Global Optimization Toolbox 3

Users Guide

How to Contact The MathWorks

Web
Newsgroup
www.mathworks.com/contact_TS.html Technical Support
www.mathworks.com

comp.soft-sys.matlab

[email protected]
[email protected]
[email protected]
[email protected]
[email protected]

Product enhancement suggestions

Bug reports
Documentation error reports
Order status, license renewals, passcodes
Sales, pricing, and general information

508-647-7000 (Phone)
508-647-7001 (Fax)
The MathWorks, Inc.
3 Apple Hill Drive
Natick, MA 01760-2098
For contact information about worldwide offices, see the MathWorks Web site.
Global Optimization Toolbox Users Guide
COPYRIGHT 20042010 by The MathWorks, Inc.
The software described in this document is furnished under a license agreement. The software may be used
or copied only under the terms of the license agreement. No part of this manual may be photocopied or
reproduced in any form without prior written consent from The MathWorks, Inc.
FEDERAL ACQUISITION: This provision applies to all acquisitions of the Program and Documentation
by, for, or through the federal government of the United States. By accepting delivery of the Program
or Documentation, the government hereby agrees that this software or documentation qualifies as
commercial computer software or commercial computer software documentation as such terms are used
or defined in FAR 12.212, DFARS Part 227.72, and DFARS 252.227-7014. Accordingly, the terms and
conditions of this Agreement and only those rights specified in this Agreement, shall pertain to and govern
the use, modification, reproduction, release, performance, display, and disclosure of the Program and
Documentation by the federal government (or other entity acquiring for or through the federal government)
and shall supersede any conflicting contractual terms or conditions. If this License fails to meet the
governments needs or is inconsistent in any respect with federal procurement law, the government agrees
to return the Program and Documentation, unused, to The MathWorks, Inc.

Trademarks

MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See
www.mathworks.com/trademarks for a list of additional trademarks. Other product or brand
names may be trademarks or registered trademarks of their respective holders.
Patents

The MathWorks products are protected by one or more U.S. patents. Please see
www.mathworks.com/patents for more information.

Revision History

January 2004
June 2004
October 2004
March 2005
September 2005
March 2006
September 2006
March 2007
September 2007
March 2008
October 2008
March 2009
September 2009
March 2010

Online only
First printing
Online only
Online only
Second printing
Online only
Online only
Online only
Third printing
Online only
Online only
Online only
Online only
Online only

New for Version 1.0 (Release 13SP1+)

Revised for Version 1.0.1 (Release 14)
Revised for Version 1.0.2 (Release 14SP1)
Revised for Version 1.0.3 (Release 14SP2)
Revised for Version 2.0 (Release 14SP3)
Revised for Version 2.0.1 (Release 2006a)
Revised for Version 2.0.2 (Release 2006b)
Revised for Version 2.1 (Release 2007a)
Revised for Version 2.2 (Release 2007b)
Revised for Version 2.3 (Release 2008a)
Revised for Version 2.4 (Release 2008b)
Revised for Version 2.4.1 (Release 2009a)
Revised for Version 2.4.2 (Release 2009b)
Revised for Version 3.0 (Release 2010a)

Contents
Introducing Global Optimization Toolbox
Functions

1
Product Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Implementation Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1-2
1-3

Example: Comparing Several Solvers . . . . . . . . . . . . . . . .

Function to Optimize . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Four Solution Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison of Syntax and Solutions . . . . . . . . . . . . . . . . . .

1-4
1-4
1-5
1-10

What Is Global Optimization? . . . . . . . . . . . . . . . . . . . . . . .

Local vs. Global Optima . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Basins of Attraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1-12
1-12
1-13

Choosing a Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Table for Choosing a Solver . . . . . . . . . . . . . . . . . . . . . . . . .
Solver Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Why Are Some Solvers Objects? . . . . . . . . . . . . . . . . . . . . . .

1-17
1-17
1-22
1-24

Writing Files for Optimization Functions

2
Computing Objective Functions . . . . . . . . . . . . . . . . . . . .
Objective (Fitness) Functions . . . . . . . . . . . . . . . . . . . . . . . .
Example: Writing a Function File . . . . . . . . . . . . . . . . . . . .
Example: Writing a Vectorized Function . . . . . . . . . . . . . .
Gradients and Hessians . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Maximizing vs. Minimizing . . . . . . . . . . . . . . . . . . . . . . . . .

2-2
2-2
2-2
2-3
2-4
2-5

Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Set Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2-6
2-6

Gradients and Hessians . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Vectorized Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2-6
2-7

Using GlobalSearch and MultiStart

vi

Contents

How to Optimize with GlobalSearch and MultiStart . .

Problems You Can Solve with GlobalSearch and
MultiStart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Outline of Steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Determining Inputs for the Problem . . . . . . . . . . . . . . . . . .
Create a Problem Structure . . . . . . . . . . . . . . . . . . . . . . . . .
Create a Solver Object . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Set Start Points for MultiStart . . . . . . . . . . . . . . . . . . . . . .
Run the Solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3-2
3-2
3-4
3-4
3-13
3-16
3-19

Examining Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Single Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Multiple Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Iterative Display . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Global Output Structures . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: Visualizing the Basins of Attraction . . . . . . . . . .

3-23
3-23
3-24
3-28
3-31
3-32

How GlobalSearch and MultiStart Work . . . . . . . . . . . . .

Multiple Runs of a Local Solver . . . . . . . . . . . . . . . . . . . . . .
Differences Between the Solver Objects . . . . . . . . . . . . . . .
GlobalSearch Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MultiStart Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3-35
3-35
3-35
3-37
3-41
3-44

Improving Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Can You Certify a Solution Is Global? . . . . . . . . . . . . . . . . .
Refining the Start Points . . . . . . . . . . . . . . . . . . . . . . . . . . .
Changing Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Reproducing Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3-45
3-45
3-48
3-55
3-59

GlobalSearch and MultiStart Examples . . . . . . . . . . . . . .

Example: Finding Global or Multiple Local Minima . . . . .
Example: Using Only Feasible Start Points . . . . . . . . . . . .

3-63
3-63
3-70

3-2

Example: Parallel MultiStart . . . . . . . . . . . . . . . . . . . . . . . .

Example: Isolated Global Minimum . . . . . . . . . . . . . . . . . .

3-74
3-77

Using Direct Search

4
What Is Direct Search? . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4-2

Performing a Pattern Search . . . . . . . . . . . . . . . . . . . . . . .

Calling patternsearch at the Command Line . . . . . . . . . . .
Using the Optimization Tool for Pattern Search . . . . . . . . .

4-3
4-3
4-3

Example: Finding the Minimum of a Function Using

the GPS Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Objective Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Finding the Minimum of the Function . . . . . . . . . . . . . . . . .
Plotting the Objective Function Values and Mesh Sizes . .

4-7
4-7
4-8
4-9

Pattern Search Terminology . . . . . . . . . . . . . . . . . . . . . . . .

Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Meshes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Polling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Expanding and Contracting . . . . . . . . . . . . . . . . . . . . . . . . .

4-11
4-11
4-12
4-13
4-13

How Pattern Search Works . . . . . . . . . . . . . . . . . . . . . . . . .

Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Successful Polls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
An Unsuccessful Poll . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Displaying the Results at Each Iteration . . . . . . . . . . . . . .
More Iterations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopping Conditions for the Pattern Search . . . . . . . . . . . .

4-14
4-14
4-15
4-18
4-19
4-20
4-21

Description of the Nonlinear Constraint Solver . . . . . .

4-24

Performing a Pattern Search Using the Optimization

Tool GUI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: A Linearly Constrained Problem . . . . . . . . . . . . .
Displaying Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4-26
4-26
4-29

vii

Example: Working with a Custom Plot Function . . . . . . . .

Performing a Pattern Search from the Command
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Calling patternsearch with the Default Options . . . . . . . . .
Setting Options for patternsearch at the Command Line . .
Using Options and Problems from the Optimization
Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Pattern Search Examples: Setting Options . . . . . . . . . . .
Poll Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Complete Poll . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: Comparing the Efficiency of Poll Options . . . . . .
Using a Search Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mesh Expansion and Contraction . . . . . . . . . . . . . . . . . . . .
Mesh Accelerator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using Cache . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Setting Tolerances for the Solver . . . . . . . . . . . . . . . . . . . . .
Constrained Minimization Using patternsearch . . . . . . . . .
Vectorizing the Objective and Constraint Functions . . . . .

4-30

4-36
4-36
4-38
4-40
4-42
4-42
4-44
4-48
4-55
4-59
4-64
4-65
4-68
4-69
4-72

Using the Genetic Algorithm

viii

Contents

What Is the Genetic Algorithm? . . . . . . . . . . . . . . . . . . . . .

5-2

Performing a Genetic Algorithm Optimization . . . . . . .

Calling the Function ga at the Command Line . . . . . . . . . .
Using the Optimization Tool . . . . . . . . . . . . . . . . . . . . . . . .

5-3
5-3
5-4

Example: Rastrigins Function . . . . . . . . . . . . . . . . . . . . . .

Rastrigins Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Finding the Minimum of Rastrigins Function . . . . . . . . . .
Finding the Minimum from the Command Line . . . . . . . . .
Displaying Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5-8
5-8
5-10
5-12
5-13

Some Genetic Algorithm Terminology . . . . . . . . . . . . . . .

Fitness Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5-17
5-17

Individuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Populations and Generations . . . . . . . . . . . . . . . . . . . . . . . .
Diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fitness Values and Best Fitness Values . . . . . . . . . . . . . . .
Parents and Children . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5-17
5-18
5-18
5-19
5-19

How the Genetic Algorithm Works . . . . . . . . . . . . . . . . . .

Outline of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Initial Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating the Next Generation . . . . . . . . . . . . . . . . . . . . . . .
Plots of Later Generations . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopping Conditions for the Algorithm . . . . . . . . . . . . . . . .

5-20
5-20
5-21
5-22
5-24
5-24

Description of the Nonlinear Constraint Solver . . . . . .

5-27

Genetic Algorithm Optimizations Using the

Optimization Tool GUI . . . . . . . . . . . . . . . . . . . . . . . . . . .
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Displaying Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: Creating a Custom Plot Function . . . . . . . . . . . .
Reproducing Your Results . . . . . . . . . . . . . . . . . . . . . . . . . .
Example: Resuming the Genetic Algorithm from the Final
Population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using the Genetic Algorithm from the Command
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Running ga with the Default Options . . . . . . . . . . . . . . . . .
Setting Options for ga at the Command Line . . . . . . . . . . .
Using Options and Problems from the Optimization
Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Reproducing Your Results . . . . . . . . . . . . . . . . . . . . . . . . . .
Resuming ga from the Final Population of a Previous
Run . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Running ga From a File . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Genetic Algorithm Examples . . . . . . . . . . . . . . . . . . . . . . .
Improving Your Results . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Population Diversity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fitness Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Reproduction Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mutation and Crossover . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5-29
5-29
5-29
5-30
5-33
5-34

5-39
5-39
5-40
5-43
5-44
5-45
5-46
5-49
5-49
5-49
5-59
5-62
5-63
5-63

ix

Setting the Amount of Mutation . . . . . . . . . . . . . . . . . . . . .

Setting the Crossover Fraction . . . . . . . . . . . . . . . . . . . . . . .
Comparing Results for Varying Crossover Fractions . . . . .
Example: Global vs. Local Minima with GA . . . . . . . . . . . .
Using a Hybrid Function . . . . . . . . . . . . . . . . . . . . . . . . . . .
Setting the Maximum Number of Generations . . . . . . . . . .
Vectorizing the Fitness Function . . . . . . . . . . . . . . . . . . . . .
Constrained Minimization Using ga . . . . . . . . . . . . . . . . . .

5-64
5-66
5-70
5-72
5-76
5-80
5-81
5-82

Using Simulated Annealing

Contents

What Is Simulated Annealing? . . . . . . . . . . . . . . . . . . . . . .

6-2

Performing a Simulated Annealing Optimization . . . . .

Calling simulannealbnd at the Command Line . . . . . . . . .
Using the Optimization Tool . . . . . . . . . . . . . . . . . . . . . . . .

6-3
6-3
6-3

Example: Minimizing De Jongs Fifth Function . . . . . . .

Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Minimizing at the Command Line . . . . . . . . . . . . . . . . . . . .
Minimizing Using the Optimization Tool . . . . . . . . . . . . . .

6-7
6-7
6-8
6-8

Some Simulated Annealing Terminology . . . . . . . . . . . . .

Objective Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Temperature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Annealing Schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Reannealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6-10
6-10
6-10
6-10
6-10

How Simulated Annealing Works . . . . . . . . . . . . . . . . . . .

Outline of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopping Conditions for the Algorithm . . . . . . . . . . . . . . . .

6-12
6-12
6-12

Using Simulated Annealing from the Command Line . .

Running simulannealbnd With the Default Options . . . . .
Setting Options for simulannealbnd at the Command
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Reproducing Your Results . . . . . . . . . . . . . . . . . . . . . . . . . .

6-14
6-14
6-15
6-17

Simulated Annealing Examples . . . . . . . . . . . . . . . . . . . . .

6-19

Multiobjective Optimization

7
What Is Multiobjective Optimization? . . . . . . . . . . . . . . .

7-2

Using gamultiobj . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using gamultiobj with Optimization Tool . . . . . . . . . . . . . .
Example: Multiobjective Optimization . . . . . . . . . . . . . . . .
Options and Syntax: Differences With ga . . . . . . . . . . . . . .

7-5
7-5
7-6
7-7
7-13

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7-14

Parallel Processing

8
Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Types of Parallel Processing in Global Optimization
Toolbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
How Toolbox Functions Distribute Processes . . . . . . . . . . .

8-2

How to Use Parallel Processing . . . . . . . . . . . . . . . . . . . . .

Multicore Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Processor Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parallel Search Functions or Hybrid Functions . . . . . . . . .

8-11
8-11
8-12
8-14

8-2
8-3

Options Reference

9
GlobalSearch and MultiStart Properties (Options) . . . .

9-2

xi

xii

Contents

How to Set Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Properties of Both Objects . . . . . . . . . . . . . . . . . . . . . . . . . .
GlobalSearch Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . .
MultiStart Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9-2
9-2
9-4
9-6

Pattern Search Options . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Optimization Tool vs. Command Line . . . . . . . . . . . . . . . . .
Plot Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Poll Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Search Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mesh Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Algorithm Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Cache Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopping Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Output Function Options . . . . . . . . . . . . . . . . . . . . . . . . . . .
Display to Command Window Options . . . . . . . . . . . . . . . .
Vectorize and Parallel Options (User Function
Evaluation) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Options Table for Pattern Search Algorithms . . . . . . . . . . .

9-7
9-7
9-8
9-10
9-13
9-17
9-19
9-19
9-20
9-21
9-23

Genetic Algorithm Options . . . . . . . . . . . . . . . . . . . . . . . . .

Optimization Tool vs. Command Line . . . . . . . . . . . . . . . . .
Plot Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Population Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fitness Scaling Options . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Selection Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Reproduction Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mutation Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Crossover Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Migration Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Algorithm Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Multiobjective Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Hybrid Function Options . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopping Criteria Options . . . . . . . . . . . . . . . . . . . . . . . . . . .
Output Function Options . . . . . . . . . . . . . . . . . . . . . . . . . . .
Display to Command Window Options . . . . . . . . . . . . . . . .
Vectorize and Parallel Options (User Function
Evaluation) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9-29
9-29
9-30
9-33
9-36
9-37
9-39
9-40
9-42
9-45
9-46
9-47
9-47
9-48
9-48
9-50

Simulated Annealing Options . . . . . . . . . . . . . . . . . . . . . . .

saoptimset At The Command Line . . . . . . . . . . . . . . . . . . . .
Plot Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Temperature Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9-53
9-53
9-53
9-55

9-24
9-25

9-51

Algorithm Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Hybrid Function Options . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopping Criteria Options . . . . . . . . . . . . . . . . . . . . . . . . . . .
Output Function Options . . . . . . . . . . . . . . . . . . . . . . . . . . .
Display Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9-56
9-57
9-58
9-59
9-61

Function Reference

10
GlobalSearch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10-2

........................................

10-2

Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10-2

.....................................

10-3

Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10-3

MultiStart

Direct Search

Class Reference

11
Functions Alphabetical List

12
Examples

A
GlobalSearch and MultiStart . . . . . . . . . . . . . . . . . . . . . . .

A-2

xiii

Pattern Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A-2

Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A-3

Simulated Annealing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A-3

Index

xiv

Contents

1
Introducing Global
Optimization Toolbox
Functions
Product Overview on page 1-2
Example: Comparing Several Solvers on page 1-4
What Is Global Optimization? on page 1-12
Choosing a Solver on page 1-17

Introducing Global Optimization Toolbox Functions

Product Overview
Global Optimization Toolbox provides methods that search for global
solutions to problems that contain multiple maxima or minima. It includes
global search, multistart, pattern search, genetic algorithm, and simulated
annealing solvers. You can use these solvers to solve optimization problems
where the objective or constraint function is continuous, discontinuous,
stochastic, does not possess derivatives, or includes simulations or black-box
functions with undefined values for some parameter settings.
Genetic algorithm and pattern search solvers support algorithmic
customization. You can create a custom genetic algorithm variant by
modifying initial population and fitness scaling options or by defining parent
selection, crossover, and mutation functions. You can customize pattern
search by defining polling, searching, and other functions.
Key features:
Interactive tools for defining and solving optimization problems and
monitoring solution progress
Global search and multistart solvers for finding single or multiple global
optima
Genetic algorithm solver that supports linear, nonlinear, and bound
constraints
Multiobjective genetic algorithm with Pareto-front identification, including
linear and bound constraints
Pattern search solver that supports linear, nonlinear, and bound
constraints
Simulated annealing tools that implement a random search method,
with options for defining annealing process, temperature schedule, and
acceptance criteria
Parallel computing support in multistart, genetic algorithm, and pattern
search solver
Custom data type support in genetic algorithm, multiobjective genetic
algorithm, and simulated annealing solvers

1-2

Product Overview

Implementation Notes
Global Optimization Toolbox solvers repeatedly attempt to locate a global
solution. However, no solver employs an algorithm that can certify a solution
as global.
You can extend the capabilities of Global Optimization Toolbox functions by
writing your own files, by using them in combination with other toolboxes, or
with the MATLAB or Simulink environments.

1-3

Introducing Global Optimization Toolbox Functions

Example: Comparing Several Solvers

In this section...
Function to Optimize on page 1-4
Four Solution Methods on page 1-5
Comparison of Syntax and Solutions on page 1-10

Function to Optimize
This example shows how to minimize Rastrigins function with four solvers.
Each solver has its own characteristics. The characteristics lead to different
solutions and run times. The results, examined in Comparison of Syntax
and Solutions on page 1-10, can help you choose an appropriate solver for
your own problems.
Rastrigins function has many local minima, with a global minimum at (0,0):

Ras( x) = 20 + x12 + x22 10 ( cos 2x1 + cos 2x2 ) .

Usually you dont know the location of the global minimum of your objective
function. To show how the solvers look for a global solution, this example
starts all the solvers around the point [20,30], which is far from the global
minimum.
The rastriginsfcn.m file implements Rastrigins function. This file comes
with Global Optimization Toolbox software. This example employs a
scaled version of Rastrigins function with larger basins of attraction. For
information, see Basins of Attraction on page 1-13.
rf2 = @(x)rastriginsfcn(x/10);

1-4

Example: Comparing Several Solvers

This example minimizes rf2 using the default settings of fminunc (an
Optimization Toolbox solver), patternsearch, and GlobalSearch. It also
uses ga with a nondefault setting, to obtain an initial population around
the point [20,30].

Four Solution Methods

fminunc on page 1-6
patternsearch on page 1-7
ga on page 1-8

1-5

Introducing Global Optimization Toolbox Functions

GlobalSearch on page 1-9

fminunc
To solve the optimization problem using the fminunc Optimization Toolbox
solver, enter:
rf2 = @(x)rastriginsfcn(x/10); % objective
x0 = [20,30]; % start point away from the minimum
[xf ff flf of] = fminunc(rf2,x0)
fminunc returns
Warning: Gradient must be provided for trust-region algorithm;
using line-search algorithm instead.
> In fminunc at 347
Local minimum found.
Optimization completed because the size of the gradient is
less than the default value of the function tolerance.
xf =
19.8991
29.8486
ff =
12.9344
flf =
1
of =
iterations: 3
funcCount: 15
stepsize: 1
firstorderopt: 5.9907e-009
algorithm: 'medium-scale: Quasi-Newton line search'
message: [1x438 char]

xf is the minimizing point.

ff is the value of the objective, rf2, at xf.
flf is the exit flag. An exit flag of 1 indicates xf is a local minimum.

1-6

Example: Comparing Several Solvers

of is the output structure, which describes the fminunc calculations

leading to the solution.

patternsearch
To solve the optimization problem using the patternsearch Global
Optimization Toolbox solver, enter:
rf2 = @(x)rastriginsfcn(x/10); % objective
x0 = [20,30]; % start point away from the minimum
[xp fp flp op] = patternsearch(rf2,x0)
patternsearch returns
Optimization terminated: mesh size less than options.TolMesh.
xp =
19.8991
-9.9496
fp =
4.9748
flp =
1
op =
function: @(x)rastriginsfcn(x/10)
problemtype: 'unconstrained'
pollmethod: 'gpspositivebasis2n'
searchmethod: []
iterations: 48
funccount: 174
meshsize: 9.5367e-007
message: 'Optimization terminated: mesh size less than
options.TolMesh.'

xp is the minimizing point.

fp is the value of the objective, rf2, at xp.
flp is the exit flag. An exit flag of 1 indicates xp is a local minimum.
op is the output structure, which describes the patternsearch calculations
leading to the solution.

1-7

Introducing Global Optimization Toolbox Functions

ga
To solve the optimization problem using the ga Global Optimization Toolbox
solver, enter:
rf2 = @(x)rastriginsfcn(x/10); % objective
x0 = [20,30]; % start point away from the minimum
initpop = 10*randn(20,2) + repmat([10 30],20,1);
opts = gaoptimset('InitialPopulation',initpop);
[xga fga flga oga] = ga(rf2,2,[],[],[],[],[],[],[],opts)
initpop is a 20-by-2 matrix. Each row of initpop has mean [10,30], and
each element is normally distributed with standard deviation 10. The rows of
initpop form an initial population matrix for the ga solver.
opts is an optimization structure setting initpop as the initial population.

The final line calls ga, using the optimization structure.

ga uses random numbers, and produces a random result. In this case ga

returns:
Optimization terminated: average change in the fitness value
less than options.TolFun.
xga =
-0.0016
19.8821
fga =
3.9804
flga =
1
oga =
problemtype: 'unconstrained'
rngstate: [1x1 struct]
generations: 54
funccount: 1100
message: [1x86 char]

xga is the minimizing point.

fga is the value of the objective, rf2, at xga.
flga is the exit flag. An exit flag of 1 indicates xga is a local minimum.

1-8

Example: Comparing Several Solvers

oga is the output structure, which describes the ga calculations leading

to the solution.

GlobalSearch
To solve the optimization problem using the GlobalSearch solver, enter:
rf2 = @(x)rastriginsfcn(x/10); % objective
x0 = [20,30]; % start point away from the minimum
problem = createOptimProblem('fmincon','objective',rf2,...
'x0',x0);
gs = GlobalSearch;
[xg fg flg og] = run(gs,problem)
problem is an optimization problem structure. problem specifies the fmincon
solver, the rf2 objective function, and x0 = [20,30]. For more information
on using createOptimProblem, see Create a Problem Structure on page 3-4.

Note You must specify fmincon as the solver for GlobalSearch, even for
unconstrained problems.
gs is a default GlobalSearch object. The object contains options for solving
the problem. Calling run(gs,problem) runs problem from multiple start
points. The start points are random, so the following result is also random.

In this case, the run returns:

GlobalSearch stopped because it analyzed all the trial points.

All 6 local solver runs converged with a positive local solver exit flag.

xg =
1.0e-007 *
-0.1405

-0.1405

fg =
0
flg =
1
og =

1-9

Introducing Global Optimization Toolbox Functions

funcCount: 2312
localSolverTotal: 6
localSolverSuccess: 6
localSolverIncomplete: 0
localSolverNoSolution: 0
message: [1x137 char]

xg is the minimizing point.

fg is the value of the objective, rf2, at xg.
flg is the exit flag. An exit flag of 1 indicates all fmincon runs converged
properly.
og is the output structure, which describes the GlobalSearch calculations
leading to the solution.

Comparison of Syntax and Solutions

One solution is better than another if its objective function value is smaller
than the other. The following table summarizes the results, accurate to one
decimal.
Results

fminunc

patternsearch

ga

GlobalSearch

solution

[19.9 29.9]

[19.9 -9.9]

[0 19.9]

[0 0]

objective

12.9

# Fevals

15

174

1040

2312

These results are typical:

fminunc quickly reaches the local solution within its starting basin, but
does not explore outside this basin at all. fminunc has simple calling
syntax.
patternsearch takes more function evaluations than fminunc, and
searches through several basins, arriving at a better solution than fminunc.
patternsearch calling syntax is the same as that of fminunc.
ga takes many more function evaluations than patternsearch. By chance
it arrived at a slightly better solution. ga is stochastic, so its results change

1-10

Example: Comparing Several Solvers

with every run. ga has simple calling syntax, but there are extra steps to
have an initial population near [20,30].
GlobalSearch run takes many more function evaluations than
patternsearch, searches many basins, and arrives at an even better
solution. In this case, GlobalSearch found the global optimum. Setting
up GlobalSearch is more involved than setting up the other solvers. As
the example shows, before calling GlobalSearch, you must create both
a GlobalSearch object (gs in the example), and a problem structure
(problem). Then, call the run method with gs and problem. For
more details on how to run GlobalSearch, see How to Optimize with
GlobalSearch and MultiStart on page 3-2.

1-11

Introducing Global Optimization Toolbox Functions

What Is Global Optimization?

In this section...
Local vs. Global Optima on page 1-12
Basins of Attraction on page 1-13

Local vs. Global Optima

Optimization is the process of finding the point that minimizes a function.
More specifically:
A local minimum of a function is a point where the function value is
smaller than or equal to the value at nearby points, but possibly greater
than at a distant point.
A global minimum is a point where the function value is smaller than or
equal to the value at all other feasible points.

Global minimum
Local minimum
Generally, Optimization Toolbox solvers find a local optimum. (This local
optimum can be a global optimum.) They find the optimum in the basin
of attraction of the starting point. For more information, see Basins of
Attraction on page 1-13.
In contrast, Global Optimization Toolbox solvers are designed to search
through more than one basin of attraction. They search in various ways:
GlobalSearch and MultiStart generate a number of starting points. They
then use a local solver to find the optima in the basins of attraction of the
starting points.

1-12

What Is Global Optimization?

ga uses a set of starting points (called the population) and iteratively

generates better points from the population. As long as the initial
population covers several basins, ga can examine several basins.
simulannealbnd performs a random search. Generally, simulannealbnd
accepts a point if it is better than the previous point. simulannealbnd
occasionally accepts a worse point, in order to reach a different basin.
patternsearch looks at a number of neighboring points before accepting
one of them. If some neighboring points belong to different basins,
patternsearch in essence looks in a number of basins at once.

Basins of Attraction
If an objective function f(x) is smooth, the vector f(x) points in the direction
where f(x) decreases most quickly. The equation of steepest descent, namely

d
x(t) = f ( x(t)),
dt
yields a path x(t) that goes to a local minimum as t gets large. Generally,
initial values x(0) that are close to each other give steepest descent paths that
tend to the same minimum point. The basin of attraction for steepest descent
is the set of initial values leading to the same local minimum.
The following figure shows two one-dimensional minima. The figure shows
different basins of attraction with different line styles, and it shows directions
of steepest descent with arrows. For this and subsequent figures, black dots
represent local minima. Every steepest descent path, starting at a point x(0),
goes to the black dot in the basin containing x(0).

f(x)

1-13

Introducing Global Optimization Toolbox Functions

The following figure shows how steepest descent paths can be more
complicated in more dimensions.

The following figure shows even more complicated paths and basins of
attraction.

1-14

What Is Global Optimization?

Constraints can break up one basin of attraction into several pieces. For
example, consider minimizing y subject to:
y |x|
y 5 4(x2)2.
The figure shows the two basins of attraction with the final points.

1-15

Introducing Global Optimization Toolbox Functions

The steepest descent paths are straight lines down to the constraint
boundaries. From the constraint boundaries, the steepest descent paths
travel down along the boundaries. The final point is either (0,0) or (11/4,11/4),
depending on whether the initial x-value is above or below 2.

1-16

Choosing a Solver

Choosing a Solver
In this section...
Table for Choosing a Solver on page 1-17
Solver Characteristics on page 1-22
Why Are Some Solvers Objects? on page 1-24

Table for Choosing a Solver

There are six Global Optimization Toolbox solvers:
ga (Genetic Algorithm)
GlobalSearch
MultiStart
patternsearch, also called direct search
simulannealbnd (Simulated Annealing)
gamultiobj, which is not a minimizer; see Chapter 7, Multiobjective
Optimization
Choose an optimizer based on problem characteristics and on the type of
solution you want. Solver Characteristics on page 1-22 contains more
information that can help you decide which solver is likely to be most suitable.
Explanation of Desired Solution on page 1-18
Choosing Between Solvers for Smooth Problems on page 1-20
Choosing Between Solvers for Nonsmooth Problems on page 1-21

1-17

Introducing Global Optimization Toolbox Functions

Desired Solution

Smooth Objective and

Constraints

Nonsmooth Objective or
Constraints

Single local solution

Optimization Toolbox functions;

see Optimization Decision
Table in the Optimization
Toolbox documentation

fminbnd, patternsearch,
fminsearch, ga, simulannealbnd

Multiple local solutions

GlobalSearch, MultiStart

Single global solution

GlobalSearch, MultiStart,
patternsearch, ga,
simulannealbnd

patternsearch, ga,
simulannealbnd

Single local solution

using parallel processing

MultiStart, Optimization

patternsearch, ga

Toolbox functions

Multiple local solutions

using parallel processing

MultiStart

Single global solution

using parallel processing

MultiStart

patternsearch, ga

Explanation of Desired Solution

To understand the meaning of the terms in Desired Solution, consider the
example
f(x) = 100x2(1x)2 x,
which has local minima x1 near 0 and x2 near 1:

1-18

Choosing a Solver

The minima are located at:

x1 = fminsearch(@(x)(100*x^2*(x - 1)^2 - x),0)
x1 =
0.0051
x2 = fminsearch(@(x)(100*x^2*(x - 1)^2 - x),1)
x2 =
1.0049

Description of the Terms

Term

Meaning

Single local solution

Find one local solution, a point x where the

objective function f(x) is a local minimum. For
more details, see Local vs. Global Optima on
page 1-12. In the example, both x1 and x2 are
local solutions.

Multiple local solutions

Find a set of local solutions. In the example, the

complete set of local solutions is {x1,x2}.

1-19

Introducing Global Optimization Toolbox Functions

Description of the Terms (Continued)

Term

Meaning

Single global solution

Find the point x where the objective function f(x)

is a global minimum. In the example, the global
solution is x2.

Choosing Between Solvers for Smooth Problems

Single Global Solution.
1 Try GlobalSearch first. It is most focused on finding a global solution, and

has an efficient local solver, fmincon.

2 Try MultiStart second. It has efficient local solvers, and can search a

wide variety of start points.

3 Try patternsearch third. It is less efficient, since it does not use gradients.

However, patternsearch is robust and is more efficient than the remaining

local solvers.
4 Try ga fourth. It can handle all types of constraints, and is usually more

efficient than simulannealbnd.

5 Try simulannealbnd last. It can handle only unconstrained or bound

constraints. simulannealbnd is usually the least efficient solver. However,

given a slow enough cooling schedule, it can find a global solution.
Multiple Local Solutions. GlobalSearch and MultiStart both provide
multiple local solutions. For the syntax to obtain multiple solutions, see
Multiple Solutions on page 3-24. GlobalSearch and MultiStart differ in
the following characteristics:
MultiStart can find more local minima. This is because GlobalSearch
rejects many generated start points (initial points for local solution).
Essentially, GlobalSearch accepts a start point only when it determines
that the point has a good chance of obtaining a global minimum. In
contrast, MultiStart passes all generated start points to a local solver. For
more information, see GlobalSearch Algorithm on page 3-37.

1-20

Choosing a Solver

MultiStart offers a choice of local solver: fmincon, fminunc, lsqcurvefit,

or lsqnonlin. The GlobalSearch solver uses only fmincon as its local
solver.
GlobalSearch uses a scatter-search algorithm for generating start points.
In contrast, MultiStart generates points uniformly at random within
bounds, or allows you to provide your own points.
MultiStart can run in parallel. See How to Use Parallel Processing
on page 8-11.

Choosing Between Solvers for Nonsmooth Problems

Choose the applicable solver with the lowest number:
1 Use fminbnd first on one-dimensional bounded problems only. fminbnd

provably converges quickly in one dimension.

2 Use patternsearch on any other type of problem. patternsearch provably

converges, and handles all types of constraints.

3 Try fminsearch next for low-dimensional unbounded problems.

fminsearch is not as general as patternsearch and can fail to converge.

For low-dimensional problems, fminsearch is simple to use, since it has

few tuning options.

4 Try ga next. ga has little supporting theory and is often less efficient than

patternsearch. It handles all types of constraints.

5 Try simulannealbnd last for unbounded problems, or for problems with

bounds. simulannealbnd provably converges only for a logarithmic cooling

schedule, which is extremely slow. simulannealbnd takes only bound
constraints, and is often less efficient than ga.

1-21

Introducing Global Optimization Toolbox Functions

Solver Characteristics
Solver

Convergence

Characteristics

GlobalSearch

Fast convergence to local

optima for smooth problems.

Deterministic iterates
Gradient-based
Automatic stochastic start points
Removes many start points
heuristically

MultiStart

Fast convergence to local

optima for smooth problems.

Deterministic iterates
Can run in parallel; see How to Use
Parallel Processing on page 8-11
Gradient-based
Stochastic or deterministic start
points, or combination of both
Automatic stochastic start points
Runs all start points
Choice of local solver: fmincon,
fminunc, lsqcurvefit, or
lsqnonlin

patternsearch

Proven convergence to
local optimum, slower than
gradient-based solvers.

Deterministic iterates
Can run in parallel; see How to Use
Parallel Processing on page 8-11
No gradients
User-supplied start point

1-22

Choosing a Solver

Solver

Convergence

Characteristics

ga

No convergence proof.

Stochastic iterates
Can run in parallel; see How to Use
Parallel Processing on page 8-11
Population-based
No gradients
Automatic start population, or
user-supplied population, or
combination of both

simulannealbnd

Proven to converge to global

optimum for bounded problems
with very slow cooling
schedule.

Stochastic iterates
No gradients
User-supplied start point
Only bound constraints

Explanation of some characteristics:

Convergence Solvers can fail to converge to any solution when
started far from a local minimum. When started near a local minimum,
gradient-based solvers converge to a local minimum quickly for smooth
problems. patternsearch provably converges for a wide range of problems,
but the convergence is slower than gradient-based solvers. Both ga and
simulannealbnd can fail to converge in a reasonable amount of time for
some problems, although they are often effective.
Iterates Solvers iterate to find solutions. The steps in the iteration are
iterates. Some solvers have deterministic iterates. Others use random
numbers and have stochastic iterates.
Gradients Some solvers use estimated or user-supplied derivatives in
calculating the iterates. Other solvers do not use or estimate derivatives,
but use only objective and constraint function values.
Start points Most solvers require you to provide a starting point for
the optimization. One reason they require a start point is to obtain the
dimension of the decision variables. ga does not require any starting points,
because it takes the dimension of the decision variables as an input. ga can
generate its start population automatically.

1-23

Introducing Global Optimization Toolbox Functions

Compare the characteristics of Global Optimization Toolbox solvers to

Optimization Toolbox solvers.
Solver

Convergence

Characteristics

fmincon, fminunc,
fseminf, lsqcurvefit,
lsqnonlin

Proven quadratic convergence

to local optima for smooth
problems

Deterministic iterates

fminsearch

No convergence proof
counterexamples exist.

Gradient-based
User-supplied starting point
Deterministic iterates
No gradients
User-supplied start point

fminbnd

Proven convergence to local

optima for smooth problems,
slower than quadratic.

Deterministic iterates
No gradients
User-supplied start point
Only one-dimensional problems

All these Optimization Toolbox solvers:

Have deterministic iterates
Start from one user-supplied point
Search just one basin of attraction

Why Are Some Solvers Objects?

GlobalSearch and MultiStart are objects. What does this mean for you?

You create a GlobalSearch or MultiStart object before running your

problem.
You can reuse the object for running multiple problems.
GlobalSearch and MultiStart objects are containers for algorithms and
global options. You use these objects to run a local solver multiple times.
The local solver has its own options.
For more information, see the Object-Oriented Programming documentation.

1-24

2
Writing Files for
Optimization Functions
Computing Objective Functions on page 2-2
Constraints on page 2-6

Writing Files for Optimization Functions

Computing Objective Functions

In this section...
Objective (Fitness) Functions on page 2-2
Example: Writing a Function File on page 2-2
Example: Writing a Vectorized Function on page 2-3
Gradients and Hessians on page 2-4
Maximizing vs. Minimizing on page 2-5

Objective (Fitness) Functions

To use Global Optimization Toolbox functions, you must first write a file (or
an anonymous function) that computes the function you want to optimize.
This function is called an objective function for most solvers or a fitness
function for ga. The function should accept a vector whose length is the
number of independent variables, and should return a scalar. For vectorized
solvers, the function should accept a matrix (where each row represents one
input vector), and return a vector of objective function values. This section
shows how to write the file.

Example: Writing a Function File

The following example shows how to write a file for the function you want to
optimize. Suppose that you want to minimize the function

f ( x) = x12 2 x1 x2 + 6 x1 + 4 x22 3 x2 .
The file that computes this function must accept a vector x of length 2,
corresponding to the variables x1 and x2, and return a scalar equal to the value
of the function at x. To write the file, do the following steps:
1 Select New > Script (Ctrl+N) from the MATLAB File menu. This opens a

new file in the editor.

2 In the file, enter the following two lines of code:

function z = my_fun(x)

2-2

Computing Objective Functions

z = x(1)^2 - 2*x(1)*x(2) + 6*x(1) + 4*x(2)^2 - 3*x(2);

3 Save the file in a folder on the MATLAB path.

To check that the file returns the correct value, enter

my_fun([2 3])
ans =
31

Example: Writing a Vectorized Function

The ga and patternsearch solvers optionally compute the objective functions
of a collection of vectors in one function call. This method can take less time
than computing the objective functions of the vectors serially. This method
is called a vectorized function call.
To compute in vectorized fashion:
Write your objective function to:

Accept a matrix with an arbitrary number of rows

Return the vector of function values of each row

If you have a nonlinear constraint, be sure to write the constraint in a

vectorized fashion. For details, see Vectorized Constraints on page 2-7.
Set the Vectorized option to 'on' with gaoptimset or psoptimset, or
set User function evaluation > Evaluate objective/fitness and
constraint functions to vectorized in the Optimization Tool. For
patternsearch, also set CompletePoll to 'on'. Be sure to pass the options
structure to the solver.
For example, to write the objective function of Example: Writing a Function
File on page 2-2 in a vectorized fashion,
function z = my_fun(x)
z = x(:,1).^2 - 2*x(:,1).*x(:,2) + 6*x(:,1) + ...
4*x(:,2).^2 - 3*x(:,2);

To use my_fun as a vectorized objective function for patternsearch:

2-3

Writing Files for Optimization Functions

options = psoptimset('CompletePoll','on','Vectorized','on');
[x fval] = patternsearch(@my_fun,[1 1],[],[],[],[],[],[],...
[],options);

To use my_fun as a vectorized objective function for ga:

options = gaoptimset('Vectorized','on');
[x fval] = ga(@my_fun,2,[],[],[],[],[],[],[],options);

For more information on writing vectorized functions for patternsearch,

see Vectorizing the Objective and Constraint Functions on page 4-72. For
more information on writing vectorized functions for ga, see Vectorizing
the Fitness Function on page 5-81.

Gradients and Hessians

If you use GlobalSearch or MultiStart, your objective function can return
derivatives (gradient, Jacobian, or Hessian). For details on how to include
this syntax in your objective function, see Writing Objective Functions in
Optimization Toolbox documentation. Use optimset to set options so that
your solver uses the derivative information:
Local Solver = fmincon, fminunc
Condition

Option Setting

Objective function contains gradient

'GradObj' = 'on'

Objective function contains Hessian

'Hessian' = 'on'

Constraint function contains

gradient

'GradConstr' = 'on'

Calculate Hessians of Lagrangian in

an extra function

'Hessian' = 'on', 'HessFcn' =

function handle

For more information about Hessians for fmincon, see Hessian.

Local Solver = lsqcurvefit, lsqnonlin

2-4

Condition

Option Setting

Objective function contains Jacobian

'Jacobian' = 'on'

Computing Objective Functions

Maximizing vs. Minimizing

Global Optimization Toolbox optimization functions minimize the objective or
fitness function. That is, they solve problems of the form

min f ( x).
x

If you want to maximize f(x), minimize f(x), because the point at which the
minimum of f(x) occurs is the same as the point at which the maximum
of f(x) occurs.
For example, suppose you want to maximize the function

f ( x) = x12 2 x1 x2 + 6 x1 + 4 x22 3 x2 .
Write your function file to compute

g( x) = f ( x) = x12 + 2 x1 x2 6 x1 4 x22 + 3 x2 ,
and minimize g(x).

2-5

Writing Files for Optimization Functions

Constraints
Many Global Optimization Toolbox functions accept bounds, linear
constraints, or nonlinear constraints. To see how to include these constraints
in your problem, see Writing Constraints in the Optimization Toolbox
documentation. Try consulting these pertinent links to sections:
Bound Constraints
Linear Inequality Constraints (linear equality constraints have the same
form)
Nonlinear Constraints

Set Bounds
It is more important to set bounds for global solvers than for local solvers.
Global solvers use bounds in a variety of ways:
GlobalSearch requires bounds for its scatter-search point generation. If
you do not provide bounds, GlobalSearch bounds each component below
by -9999 and above by 10001. However, these bounds can easily be
inappropriate.
If you do not provide bounds and do not provide custom start points,
MultiStart bounds each component below by -1000 and above by 1000.
However, these bounds can easily be inappropriate.
ga uses bounds and linear constraints for its initial population generation.
For unbounded problems, ga uses a default of 0 as the lower bound and 1
as the upper bound for each dimension for initial point generation. For
bounded problems, and problems with linear constraints, ga uses the
bounds and constraints to make the initial population.
simulannealbnd and patternsearch do not require bounds, although they
can use bounds.

Gradients and Hessians

If you use GlobalSearch or MultiStart with fmincon, your nonlinear
constraint functions can return derivatives (gradient or Hessian). For details,
see Gradients and Hessians on page 2-4.

2-6

Constraints

Vectorized Constraints
The ga and patternsearch solvers optionally compute the nonlinear
constraint functions of a collection of vectors in one function call. This method
can take less time than computing the objective functions of the vectors
serially. This method is called a vectorized function call.
For the solver to compute in a vectorized manner, you must vectorize both
your objective (fitness) function and nonlinear constraint function. For details,
see Vectorizing the Objective and Constraint Functions on page 4-72.
As an example, suppose your nonlinear constraints for a three-dimensional
problem are

x12 x22 x32

+
+
6
4
9 25
x3 cosh ( x1 + x2 )
x1 x2 x3 = 2.
The following code gives these nonlinear constraints in a vectorized fashion,
assuming that the rows of your input matrix x are your population or input
vectors:
function [c ceq] = nlinconst(x)
c(:,1) = x(:,1).^2/4 + x(:,2).^2/9 + x(:,3).^2/25 - 6;
c(:,2) = cosh(x(:,1) + x(:,2)) - x(:,3);
ceq = x(:,1).*x(:,2).*x(:,3) - 2;

For example, minimize the vectorized quadratic function

function y = vfun(x)
y = -x(:,1).^2 - x(:,2).^2 - x(:,3).^2;

over the region with constraints nlinconst using patternsearch:

options = psoptimset('CompletePoll','on','Vectorized','on');
[x fval] = patternsearch(@vfun,[1,1,2],[],[],[],[],[],[],...
@nlinconst,options)
Optimization terminated: mesh size less than options.TolMesh

2-7

Writing Files for Optimization Functions

and constraint violation is less than options.TolCon.

x =
0.2191

0.7500

12.1712

fval =
-148.7480

Using ga:
options = gaoptimset('Vectorized','on');
[x fval] = ga(@vfun,3,[],[],[],[],[],[],@nlinconst,options)
Optimization terminated: maximum number of generations exceeded.
x =
-1.4098

-0.1216

11.6664

fval =
-138.1066

For this problem patternsearch computes the solution far more quickly
and accurately.

2-8

3
Using GlobalSearch and
MultiStart
How to Optimize with GlobalSearch and MultiStart on page 3-2
Examining Results on page 3-23
How GlobalSearch and MultiStart Work on page 3-35
Improving Results on page 3-45
GlobalSearch and MultiStart Examples on page 3-63

Using GlobalSearch and MultiStart

How to Optimize with GlobalSearch and MultiStart

In this section...
Problems You Can Solve with GlobalSearch and MultiStart on page 3-2
Outline of Steps on page 3-2
Determining Inputs for the Problem on page 3-4
Create a Problem Structure on page 3-4
Create a Solver Object on page 3-13
Set Start Points for MultiStart on page 3-16
Run the Solver on page 3-19

Problems You Can Solve with GlobalSearch and

MultiStart
The GlobalSearch and MultiStart solvers apply to problems with smooth
objective and constraint functions. The solvers search for a global minimum,
or for a set of local minima. For more information on which solver to use,
see Choosing a Solver on page 1-17.
GlobalSearch and MultiStart work by starting a local solver, such as
fmincon, from a variety of start points. Generally the start points are random.
However, for MultiStart you can provide a set of start points. For more

information, see How GlobalSearch and MultiStart Work on page 3-35.

To find out how to use these solvers, see Outline of Steps on page 3-2. For
examples using these solvers, see the example index.

Outline of Steps
To find a global or multiple local solutions:
1 Create a Problem Structure on page 3-4
2 Create a Solver Object on page 3-13
3 (Optional, MultiStart only) Set Start Points for MultiStart on page 3-16

3-2

How to Optimize with GlobalSearch and MultiStart

4 Run the Solver on page 3-19

The following figure illustrates these steps.

Information
you have

Local Solver

X0

Objective Constraints

Global options

Local Options

Command
to use

Resulting
Object or
Structure

createOptimProblem

Problem Structure

Optional,
MultiStart only
Start Points

GlobalSearch
or
MultiStart

RandomStartPointSet
or
CustomStartPointSet

Solver Object

Start Points Object

run

Results

3-3

Using GlobalSearch and MultiStart

Determining Inputs for the Problem

A problem structure defines a local optimization problem using a local solver,
such as fmincon. Therefore, most of the documentation on choosing a solver
or preparing the inputs is in the Optimization Toolbox documentation.
Required Inputs
Input

Link to Documentation

Local solver

Optimization Decision Table in the Optimization

Toolbox documentation.

Objective function

Computing Objective Functions on page 2-2

Start point x0
Optional Inputs
Input

Link to Documentation

Constraint functions

Constraints on page 2-6

Local options
structure

Setting Options

Create a Problem Structure

To use the GlobalSearch or MultiStart solvers, you must first create a
problem structure. There are two ways to create a problem structure:
Using the createOptimProblem Function on page 3-4
Exporting from the Optimization Tool on page 3-8

Using the createOptimProblem Function

Follow these steps to create a problem structure using the
createOptimProblem function.
1 Define your objective function as a file or anonymous function. For

details, see Computing Objective Functions on page 2-2. If your solver

3-4

How to Optimize with GlobalSearch and MultiStart

is lsqcurvefit or lsqnonlin, ensure the objective function returns a

vector, not scalar.
2 If relevant, create your constraints, such as bounds and nonlinear

constraint functions. For details, see Constraints on page 2-6.

3 Create a start point. For example, to create a three-dimensional random

start point xstart:

xstart = randn(3,1);
4 (Optional) Create an options structure using optimset. For example,

options = optimset('Algorithm','interior-point');
5 Enter

problem = createOptimProblem(solver,

where solver is the name of your local solver:

For GlobalSearch: 'fmincon'
For MultiStart the choices are:
'fmincon'
'fminunc'
'lsqcurvefit'
'lsqnonlin'
For help choosing, see Optimization Decision Table.
6 Set an initial point using the 'x0' parameter. If your initial point is

xstart, and your solver is fmincon, your entry is now

problem = createOptimProblem('fmincon','x0',xstart,
7 Include the function handle for your objective function in objective:

problem = createOptimProblem('fmincon','x0',xstart, ...

'objective',@objfun,
8 Set bounds and other constraints as applicable.

3-5

Using GlobalSearch and MultiStart

Constraint

Name

lower bounds

'lb'

upper bounds

'ub'

matrix Aineq for linear inequalities

Aineq x bineq

'Aineq'

vector bineq for linear inequalities

Aineq x bineq

'bineq'

matrix Aeq for linear equalities Aeq x = beq

'Aeq'

vector beq for linear equalities Aeq x = beq

'beq'

nonlinear constraint function

'nonlcon'

9 If using the lsqcurvefit local solver, include vectors of input data and

response data, named 'xdata' and 'ydata' respectively.

10 Best practice: validate the problem structure by running your solver on the

structure. For example, if your local solver is fmincon:

[x fval eflag output] = fmincon(problem);

Note Specify fmincon as the solver for GlobalSearch, even if you have
no constraints. However, you cannot run fmincon on a problem without
constraints. Add an artificial constraint to the problem to validate the
structure:
problem.lb = -Inf;

Example: Creating a Problem Structure with createOptimProblem.

This example minimizes the function from Run the Solver on page 3-19,
subject to the constraint x1 + 2x2 4. The objective is
sixmin = 4x2 2.1x4 + x6/3 + xy 4y2 + 4y4.
Use the interior-point algorithm of fmincon, and set the start point to
[2;3].

3-6

How to Optimize with GlobalSearch and MultiStart

1 Write a function handle for the objective function.

sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...

+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
2 Write the linear constraint matrices. Change the constraint to less than

form:
A = [-1,-2];
b = -4;
3 Create the local options structure to use the interior-point algorithm:

opts = optimset('Algorithm','interior-point');
4 Create the problem structure with createOptimProblem:

problem = createOptimProblem('fmincon', ...

'x0',[2;3],'objective',sixmin, ...
'Aineq',A,'bineq',b,'options',opts);
5 The resulting structure:

problem =
objective: @(x)(4*x(1)^2-2.1*x(1)^4+x(1)^6/3+ ...
x(1)*x(2)-4*x(2)^2+4*x(2)^4)
x0: [2x1 double]
Aineq: [-1 -2]
bineq: -4
Aeq: []
beq: []
lb: []
ub: []
nonlcon: []
solver: 'fmincon'
options: [1x1 struct]
6 Best practice: validate the problem structure by running your solver on

the structure:
[x fval eflag output] = fmincon(problem);

3-7

Using GlobalSearch and MultiStart

Note Specify fmincon as the solver for GlobalSearch, even if you have
no constraints. However, you cannot run fmincon on a problem without
constraints. Add an artificial constraint to the problem to validate the
structure:
problem.lb = -Inf;

Exporting from the Optimization Tool

Follow these steps to create a problem structure using the Optimization Tool.
1 Define your objective function as a file or anonymous function. For

details, see Computing Objective Functions on page 2-2. If your solver

is lsqcurvefit or lsqnonlin, ensure the objective function returns a
vector, not scalar.
2 If relevant, create nonlinear constraint functions. For details, see

Nonlinear Constraints.
3 Create a start point. For example, to create a three-dimensional random

start point xstart:

xstart = randn(3,1);
4 Open the Optimization Tool by entering optimtool at the command line, or

by choosing Start > Toolboxes > Optimization > Optimization Tool.

5 Choose the local Solver.

For GlobalSearch: fmincon (default).

For MultiStart:
fmincon (default)

3-8

How to Optimize with GlobalSearch and MultiStart

fminunc
lsqcurvefit
lsqnonlin
For help choosing, see Optimization Decision Table.
6 Choose an appropriate Algorithm. For help choosing, see Choosing the

Algorithm.

7 Set an initial point (Start point).

8 Include the function handle for your objective function in Objective

function, and, if applicable, include your Nonlinear constraint

function. For example,

9 Set bounds, linear constraints, or local Options. For details on constraints,

see Writing Constraints.

3-9

Using GlobalSearch and MultiStart

10 Best practice: run the problem to verify the setup.

Note You must specify fmincon as the solver for GlobalSearch, even if
you have no constraints. However, you cannot run fmincon on a problem
without constraints. Add an artificial constraint to the problem to verify
the setup.

11 Choose File > Export to Workspace and select Export problem and

options to a MATLAB structure named

3-10

How to Optimize with GlobalSearch and MultiStart

Example: Creating a Problem Structure with the Optimization Tool.

This example minimizes the function from Run the Solver on page 3-19,
subject to the constraint x1 + 2x2 4. The objective is
sixmin = 4x2 2.1x4 + x6/3 + xy 4y2 + 4y4.
Use the interior-point algorithm of fmincon, and set the start point to
[2;3].
1 Write a function handle for the objective function.

sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...

+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
2 Write the linear constraint matrices. Change the constraint to less than

form:
A = [-1,-2];
b = -4;
3 Launch the Optimization Tool by entering optimtool at the MATLAB

command line.
4 Set the solver, algorithm, objective, start point, and constraints.

3-11

Using GlobalSearch and MultiStart

5 Best practice: run the problem to verify the setup.

The problem runs successfully.

3-12

How to Optimize with GlobalSearch and MultiStart

6 Choose File > Export to Workspace and select Export problem and

options to a MATLAB structure named

Create a Solver Object

A solver object contains your preferences for the global portion of the
optimization.
You do not need to set any preferences. Create a GlobalSearch object named
gs with default settings as follows:
gs = GlobalSearch;

3-13

Using GlobalSearch and MultiStart

Similarly, create a MultiStart object named ms with default settings as

follows:
ms = MultiStart;

Properties (Global Options) of Solver Objects

Global options are properties of a GlobalSearch or MultiStart object.
Properties for both GlobalSearch and MultiStart
Property Name

Meaning

Display

Detail level of iterative display. Set to 'off' for no

display, 'final' (default) for a report at the end of the
run, or 'iter' for reports as the solver progresses.
For more information and examples, see Iterative
Display on page 3-28.

TolFun

Solvers consider objective function values within

TolFun of each other to be identical (not distinct).
Default: 1e-6. Solvers group solutions when the
solutions satisfy both TolFun and TolX tolerances.

TolX

Solvers consider solutions within TolX distance of each

other to be identical (not distinct). Default: 1e-6.
Solvers group solutions when the solutions satisfy both
TolFun and TolX tolerances.

MaxTime

Solvers halt if the run exceeds MaxTime seconds, as

measured by a clock (not processor seconds). Default:
Inf

StartPointsToRun Choose whether to run 'all' (default) start points,

only those points that satisfy 'bounds', or only those

points that are feasible with respect to bounds and

inequality constraints with 'bounds-ineqs'. For an
example, see Example: Using Only Feasible Start
Points on page 3-70.

3-14

How to Optimize with GlobalSearch and MultiStart

Properties for GlobalSearch

Property Name

Meaning

NumTrialPoints

Number of trial points to use examine.

Default: 1000

BasinRadiusFactor

See Properties on page 12-30 for detailed

descriptions of these properties.

DistanceThresholdFactor
MaxWaitCycle
NumStageOnePoints
PenaltyThresholdFactor

Properties for MultiStart

Property Name

Meaning

UseParallel

When 'always', MultiStart attempts to distribute

start points to multiple processors for the local solver.
Disable by setting to 'never' (default). For details,
see How to Use Parallel Processing on page 8-11.
For an example, see Example: Parallel MultiStart
on page 3-74.

Example: Creating a Nondefault GlobalSearch Object

Suppose you want to solve a problem and:
Consider local solutions identical if they are within 0.01 of each other and
the function values are within the default TolFun tolerance.
Spend no more than 2000 seconds on the computation.
To solve the problem, create a GlobalSearch object gs as follows:
gs = GlobalSearch('TolX',0.01,'MaxTime',2000);

Example: Creating a Nondefault MultiStart Object

Suppose you want to solve a problem such that:

3-15

Using GlobalSearch and MultiStart

You consider local solutions identical if they are within 0.01 of each other
and the function values are within the default TolFun tolerance.
You spend no more than 2000 seconds on the computation.
To solve the problem, create a MultiStart object ms as follows:
ms = MultiStart('TolX',0.01,'MaxTime',2000);

Set Start Points for MultiStart

There are four ways you tell MultiStart which start points to use for the
local solver:
Pass a positive integer k. MultiStart generates k - 1 start points
as if using a RandomStartPointSet object and the problem structure.
MultiStart also uses the x0 start point from the problem structure, for a
total of k start points.
Pass a RandomStartPointSet object.
Pass a CustomStartPointSet object.
Pass a cell array of RandomStartPointSet and CustomStartPointSet
objects. Pass a cell array if you have some specific points you want to run,
but also want MultiStart to use other random start points.
Note You can control whether MultiStart uses all start points, or only
those points that satisfy bounds or other inequality constraints. For more
information, see Filter Start Points (Optional) on page 3-42.

Positive Integer for Start Points

The syntax for running MultiStart for k start points is
[xmin,fmin,flag,outpt,allmins] = run(ms,problem,k);

The positive integer k specifies the number of start points MultiStart uses.
MultiStart generates random start points using the dimension of the problem
and bounds from the problem structure. MultiStart generates k - 1 random
start points, and also uses the x0 start point from the problem structure.

3-16

How to Optimize with GlobalSearch and MultiStart

RandomStartPointSet Object for Start Points

Create a RandomStartPointSet object as follows:
stpoints = RandomStartPointSet;

By default a RandomStartPointSet object generates 10 start points. Control

the number of start points with the NumStartPoints property. For example,
to generate 40 start points:
stpoints = RandomStartPointSet('NumStartPoints',40);

You can set an ArtificialBound for a RandomStartPointSet.

If a component has no bounds, RandomStartPointSet uses a lower bound
of -ArtificialBound, and an upper bound of ArtificialBound.
If a component has a lower bound lb but no upper bound,
RandomStartPointSet uses an upper bound of lb + 2*ArtificialBound.
Similarly, if a component has an upper bound ub but no lower bound,
RandomStartPointSet uses a lower bound of ub - 2*ArtificialBound.
For example, to generate 100 start points with an ArtificialBound of 50:
stpoints = RandomStartPointSet('NumStartPoints',100, ...
'ArtificialBound',50);

CustomStartPointSet Object for Start Points

To use a specific set of starting points, package them in a
CustomStartPointSet as follows:
1 Place the starting points in a matrix. Each row of the matrix represents

one starting point. MultiStart runs all the rows of the matrix, subject to
filtering with the StartPointsToRun property. For more information, see
MultiStart Algorithm on page 3-41.
2 Create a CustomStartPointSet object from the matrix:

tpoints = CustomStartPointSet(ptmatrix);

For example, create a set of 40 five-dimensional points, with each component

of a point equal to 10 plus an exponentially distributed variable with mean 25:

3-17

Using GlobalSearch and MultiStart

pts = -25*log(rand(40,5)) + 10;

tpoints = CustomStartPointSet(pts);

To get the original matrix of points from a CustomStartPointSet object,

use the list method:
pts = list(tpoints); % Assumes tpoints is a CustomStartPointSet

A CustomStartPointSet has two properties: DimStartPoints

and NumStartPoints. You can use these properties to query a
CustomStartPointSet object. For example, the tpoints object in the example
has the following properties:
tpoints.DimStartPoints
ans =
5
tpoints.NumStartPoints
ans =
40

Cell Array of Objects for Start Points

To use a specific set of starting points along with some randomly generated
points, pass a cell array of RandomStartPointSet or CustomStartPointSet
objects.
For example, to use both the 40 specific five-dimensional points of
CustomStartPointSet Object for Start Points on page 3-17 and 40 additional
five-dimensional points from RandomStartPointSet:
pts = -25*log(rand(40,5)) + 10;
tpoints = CustomStartPointSet(pts);
rpts = RandomStartPointSet('NumStartPoints',40);
allpts = {tpoints,rpts};

Run MultiStart with the allpts cell array:

% Assume ms and problem exist
[xmin,fmin,flag,outpt,allmins] = run(ms,problem,allpts);

3-18

How to Optimize with GlobalSearch and MultiStart

Run the Solver

Running a solver is nearly identical for GlobalSearch and MultiStart. The
only difference in syntax is MultiStart takes an additional input describing
the start points.
For example, suppose you want to find several local minima of the sixmin
function
sixmin = 4x2 2.1x4 + x6/3 + xy 4y2 + 4y4.

This function is also called the six-hump camel back function [3]. All the local
minima lie in the region 3 x,y 3.

Example of Run with GlobalSearch

To find several local minima of the problem described in Run the Solver on
page 3-19 using GlobalSearch, enter:
gs = GlobalSearch;
opts = optimset('Algorithm','interior-point');

3-19

Using GlobalSearch and MultiStart

sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...

+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3],...
'options',opts);
[xming,fming,flagg,outptg,manyminsg] = run(gs,problem);

The output of the run (which varies, based on the random seed):
xming,fming,flagg,outptg,manyminsg
xming =
-0.0898
0.7127
fming =
-1.0316
flagg =
1
outptg =
funcCount:
localSolverTotal:
localSolverSuccess:
localSolverIncomplete:
localSolverNoSolution:
message:

2245
8
8
0
0
[1x137 char]

manyminsg =
1x4 GlobalOptimSolution
Properties:
X
Fval
Exitflag
Output
X0

Example of Run with MultiStart

To find several local minima of the problem described in Run the Solver on
page 3-19 using 50 runs of fmincon with MultiStart, enter:

3-20

How to Optimize with GlobalSearch and MultiStart

% % Set the default stream to get exactly the same output

% s = RandStream('mt19937ar','Seed',14);
% RandStream.setDefaultStream(s);
ms = MultiStart;
opts = optimset('Algorithm','interior-point');
sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...
+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3],...
'options',opts);
[xminm,fminm,flagm,outptm,manyminsm] = run(ms,problem,50);

The output of the run (which varies based on the random seed):
xminm,fminm,flagm,outptm,manyminsm
xminm =
-0.0898

0.7127

fminm =
-1.0316
flagm =
1
outptm =
funcCount:
localSolverTotal:
localSolverSuccess:
localSolverIncomplete:
localSolverNoSolution:
message:

2035
50
50
0
0
[1x128 char]

manyminsm =
1x6 GlobalOptimSolution
Properties:
X
Fval
Exitflag

3-21

Using GlobalSearch and MultiStart

Output
X0

In this case, MultiStart located all six local minima, while GlobalSearch
located four. For pictures of the MultiStart solutions, see Example:
Visualizing the Basins of Attraction on page 3-32.

3-22

Examining Results

Examining Results
In this section...
Single Solution on page 3-23
Multiple Solutions on page 3-24
Iterative Display on page 3-28
Global Output Structures on page 3-31
Example: Visualizing the Basins of Attraction on page 3-32

Single Solution
You obtain the single best solution found during the run by calling run with
the syntax
[x fval eflag output] = run(...);

x is the location of the local minimum with smallest objective function

value.
fval is the objective function value evaluated at x.
eflag is an exit flag for the global solver. Values:
Global Solver Exit Flags
2

At least one local minimum found. Some runs of the local solver
converged (had positive exit flag).

At least one local minimum found. All runs of the local solver
converged (had positive exit flag).

No local minimum found. Local solver called at least once, and

at least one local solver exceeded the MaxIter or MaxFunEvals
tolerances.

-1

One or more local solver runs stopped by the local solver output
function or plot function.

-2

No feasible local minimum found.

-5

MaxTime limit exceeded.

3-23

Using GlobalSearch and MultiStart

Global Solver Exit Flags (Continued)

-8

No solution found. All runs had local solver exit flag -2 or

smaller, not all equal -2.

-10

Failures encountered in user-provided functions.

output is a structure with details about the multiple runs of the local
solver. For more information, see Global Output Structures on page 3-31.
The list of outputs is for the case eflag > 0. If eflag <= 0, then x is the
following:
If some local solutions are feasible, x represents the location of the lowest
objective function value. Feasible means the constraint violations are
smaller than problem.options.TolCon.
If no solutions are feasible, x is the solution with lowest infeasibility.
If no solutions exist, x, fval, and output are empty ([]).

Multiple Solutions
You obtain multiple solutions in an object by calling run with the syntax
[x fval eflag output manymins] = run(...);
manymins is a vector of solution objects; see GlobalOptimSolution. The
vector is in order of objective function value, from lowest (best) to highest
(worst). Each solution object contains the following properties (fields):

X a local minimum
Fval the value of the objective function at X
Exitflag the exit flag for the global solver (described here)
Output an output structure (described here)
X0 a cell array of start points that led to the solution point X
There are several ways to examine the vector of solution objects:

3-24

Examining Results

In the MATLAB Workspace Browser. Double-click the solution object, and

then double-click the resulting display in the Variable Editor.

Using dot addressing. GlobalOptimSolution properties are capitalized.

Use proper capitalization to access the properties.
For example, to find the vector of function values, enter:
fcnvals = [manymins.Fval]
fcnvals =

3-25

Using GlobalSearch and MultiStart

-1.0316

-0.2155

To get a cell array of all the start points that led to the lowest function
value (the first element of manymins), enter:
smallX0 = manymins(1).X0

Plot some field values. For example, to see the range of resulting Fval,
enter:
hist([manymins.Fval])

This results in a histogram of the computed function values. (The figure

shows a histogram from a different example than the previous few figures.)

Example: Changing the Definition of Distinct Solutions

You might find out, after obtaining multiple local solutions, that your
tolerances were not appropriate. You can have many more local solutions
than you want, spaced too closely together. Or you can have fewer solutions

3-26

Examining Results

than you want, with GlobalSearch or MultiStart clumping together too

many solutions.
To deal with this situation, run the solver again with different tolerances. The
TolX and TolFun tolerances determine how the solvers group their outputs
into the GlobalOptimSolution vector. These tolerances are properties of the
GlobalSearch or MultiStart object.

For example, suppose you want to use the active-set algorithm in fmincon
to solve the problem in Example of Run with MultiStart on page 3-20.
Further suppose that you want to have tolerances of 0.01 for both TolX and
TolFun. The run method groups local solutions whose objective function
values are within TolFun of each other, and which are also less than TolX
apart from each other. To obtain the solution:
ms = MultiStart('TolFun',0.01,'TolX',0.01);
opts = optimset('Algorithm','active-set');
sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...
+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3],...
'options',opts);
[xminm,fminm,flagm,outptm,someminsm] = run(ms,problem,50);
MultiStart completed the runs from all start points.
All 50 local solver runs converged with a
positive local solver exit flag.
someminsm
someminsm =
1x6 GlobalOptimSolution
Properties:
X
Fval
Exitflag
Output

3-27

Using GlobalSearch and MultiStart

X0

In this case, MultiStart generated six distinct solutions. Here distinct

means that the solutions are more than 0.01 apart in either objective function
value or location.

Iterative Display
Iterative display gives you information about the progress of solvers during
their runs.
There are two types of iterative display:
Global solver display
Local solver display
Both types appear at the command line, depending on global and local options.
Obtain local solver iterative display by setting the Display option in the
problem.options structure to 'iter' or 'iter-detailed' with optimset.
For more information, see Displaying Iterative Output in the Optimization
Toolbox documentation.
Obtain global solver iterative display by setting the Display property in the
GlobalSearch or MultiStart object to 'iter'.

Global solvers set the default Display option of the local solver to 'off',
unless the problem structure has a value for this option. Global solvers do
not override any setting you make for local options.
Note Setting the local solver Display option to anything other than 'off'
can produce a great deal of output. The default Display option created by
optimset(@solver) is 'final'.

Example: Types of Iterative Display

Run the example described in Run the Solver on page 3-19 using
GlobalSearch with GlobalSearch iterative display:

3-28

Examining Results

gs = GlobalSearch('Display','iter');
opts = optimset('Algorithm','interior-point');
sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...
+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3],...
'options',opts);
[xming,fming,flagg,outptg,manyminsg] = run(gs,problem);

Num Pts

Best

Current

Threshold

f(x)

Penalty

Penalty

Analyzed

F-count

34

-1.032

200

1441

-1.032

300

1641

-1.032

400

1841

500

2041

541

Local

Local

f(x)

exitflag

Procedure

-1.032

Initial Point

Stage 1 Local

290.7

0.07006

Stage 2 Search

-1.032

101

0.8017

Stage 2 Search

-1.032

51.23

3.483

2157

-1.032

5.348

5.456

-1.032

Stage 2 Local

542

2190

-1.032

2.993

5.348

-1.032

Stage 2 Local

544

2227

-1.032

1.807

2.993

-1.032

Stage 2 Local

545

2260

-1.032

1.609

1.807

-1.032

Stage 2 Local

546

2296

-1.032

0.9061

1.609

-1.032

Stage 2 Local

552

2348

-1.032

0.6832

0.9061

-0.2155

600

2444

-1.032

1.079

0.1235

Stage 2 Search

700

2644

-1.032

0.4859

-0.4791

Stage 2 Search

800

2844

-1.032

136.9

0.8202

Stage 2 Search

900

3044

-1.032

37.77

0.4234

Stage 2 Search

1000

3244

-1.032

-0.05542

-0.598

Stage 2 Search

Stage 2 Search

Stage 2 Local

GlobalSearch stopped because it analyzed all the start points.

All 8 local solver runs converged with a positive local solver exit flag.

Run the same example without GlobalSearch iterative display, but with
fmincon iterative display:
gs.Display = 'final';
problem.options.Display = 'iter';
[xming,fming,flagg,outptg,manyminsg] = run(gs,problem);

3-29

Using GlobalSearch and MultiStart

First-order

Norm of

f(x)

Feasibility

optimality

step

4.823333e+001

0.000e+000

1.088e+002

2.020476e+000

0.000e+000

2.176e+000

2.488e+000

10

6.525252e-001

0.000e+000

1.937e+000

1.886e+000

13

-8.776121e-001

0.000e+000

9.076e-001

8.539e-001

16

-9.121907e-001

0.000e+000

9.076e-001

1.655e-001

19

-1.009367e+000

0.000e+000

7.326e-001

8.558e-002

22

-1.030423e+000

0.000e+000

2.172e-001

6.670e-002

25

-1.031578e+000

0.000e+000

4.278e-002

1.444e-002

28

-1.031628e+000

0.000e+000

8.777e-003

2.306e-003

31

-1.031628e+000

0.000e+000

8.845e-005

2.750e-004

10

34

-1.031628e+000

0.000e+000

8.744e-007

1.354e-006

Iter F-count
0

Local minimum found that satisfies the constraints.

Optimization completed because the objective function is non-decreasing in

feasible directions, to within the selected value of the function tolerance,
and constraints were satisfied to within the selected value of the constraint tolerance.

<stopping criteria details>

Iter F-count
0

First-order

Norm of

f(x)

Feasibility

optimality

step

-5.372419e-001

0.000e+000

1.913e+000

0.000e+000

2.425e-007

... MANY ITERATIONS DELETED ...

39

-2.154638e-001

6.002e-008

Local minimum found that satisfies the constraints.

Optimization completed because the objective function is non-decreasing in

feasible directions, to within the selected value of the function tolerance,
and constraints were satisfied to within the selected value of the constraint tolerance.

<stopping criteria details>

GlobalSearch stopped because it analyzed all the start points.

All 7 local solver runs converged with a positive local solver exit flag.

3-30

Examining Results

Setting GlobalSearch iterative display, as well as fmincon iterative display,

yields both displays intermingled.
For an example of iterative display in a parallel environment, see Example:
Parallel MultiStart on page 3-74.

Global Output Structures

run can produce two types of output structures:

A global output structure. This structure contains information about the

overall run from multiple starting points. Details follow.
Local solver output structures. The vector of GlobalOptimSolution objects
contains one such structure in each element of the vector. For a description
of this structure, see Output Structures in the Optimization Toolbox
documentation, or the function reference pages for the local solvers:
fmincon, fminunc, lsqcurvefit, or lsqnonlin.
Global Output Structure
Field

Meaning

funcCount

Total number of calls to user-supplied functions (objective or

nonlinear constraint)

localSolverTotal

Number of local solver runs started

localSolverSuccess

Number of local solver runs that finished with a positive exit flag

localSolverIncomplete

Number of local solver runs that finished with a 0 exit flag

localSolverNoSolution

Number of local solver runs that finished with a negative exit

flag

message

GlobalSearch or MultiStart exit message

A positive exit flag from a local solver generally indicates a successful run. A
negative exit flag indicates a failure. A 0 exit flag indicates that the solver
stopped by exceeding the iteration or function evaluation limit. For more
information, see Exit Flags and Exit Messages or Tolerances and Stopping
Criteria in the Optimization Toolbox documentation.

3-31

Using GlobalSearch and MultiStart

Example: Visualizing the Basins of Attraction

Which start points lead to which basin? For a steepest descent solver, nearby
points generally lead to the same basin; see Basins of Attraction on page
1-13. However, for Optimization Toolbox solvers, basins are more complicated.
Plot the MultiStart start points from the example, Example of Run with
MultiStart on page 3-20, color-coded with the basin where they end.
% s = RandStream('mt19937ar','Seed',14);
% RandStream.setDefaultStream(s);
% Uncomment the previous lines to get the same output
ms = MultiStart;
opts = optimset('Algorithm','interior-point');
sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...
+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3],...
'options',opts);
[xminm,fminm,flagm,outptm,manyminsm] = run(ms,problem,50);
possColors = 'kbgcrm';
hold on
for i = 1:size(manyminsm,2)
% Color of this line
cIdx = rem(i-1, length(possColors)) + 1;
color = possColors(cIdx);
% Plot start points
u = manyminsm(i).X0;
x0ThisMin = reshape([u{:}], 2, length(u));
plot(x0ThisMin(1, :), x0ThisMin(2, :), '.', ...
'Color',color,'MarkerSize',25);
% Plot the basin with color i
plot(manyminsm(i).X(1), manyminsm(i).X(2), '*', ...
'Color', color, 'MarkerSize',25);
end % basin center marked with a *, start points with dots
hold off

3-32

Examining Results

The figure shows the centers of the basins by colored * symbols. Start points
with the same color as the * symbol converge to the center of the * symbol.
Start points do not always converge to the closest basin. For example, the red
points are closer to the green basin center than to the red basin center. Also,
many black and blue start points are closer to the opposite basin centers.
The magenta and red basins are shallow, as you can see in the following
contour plot.

3-33

3-34

Using GlobalSearch and MultiStart

How GlobalSearch and MultiStart Work

How GlobalSearch and MultiStart Work

In this section...
Multiple Runs of a Local Solver on page 3-35
Differences Between the Solver Objects on page 3-35
GlobalSearch Algorithm on page 3-37
MultiStart Algorithm on page 3-41
Bibliography on page 3-44

Multiple Runs of a Local Solver

GlobalSearch and MultiStart have similar approaches to finding global or
multiple minima. Both algorithms start a local solver (such as fmincon) from
multiple start points. The algorithms use multiple start points to sample
multiple basins of attraction. For more information, see Basins of Attraction
on page 1-13.

Differences Between the Solver Objects

GlobalSearch and MultiStart Algorithm Overview on page 3-36 contains a
sketch of the GlobalSearch and MultiStart algorithms.

3-35

Using GlobalSearch and MultiStart

GlobalSearch Algorithm

MultiStart Algorithm

Run fmincon from x0

Generate start points

Generate trial points

(potential start points)

Run start points

Stage 1:
Run best start point among the first
NumStageOnePoints trial points

Create GlobalOptimSolutions vector

Stage 2:
Loop through remaining trial points,
run fmincon if point satisfies
basin, score, and constraint filters
Create GlobalOptimSolutions vector
GlobalSearch and MultiStart Algorithm Overview

The main differences between GlobalSearch and MultiStart are:

GlobalSearch uses a scatter-search mechanism for generating start points.
MultiStart uses uniformly distributed start points within bounds, or
user-supplied start points.
GlobalSearch analyzes start points and rejects those points that are
unlikely to improve the best local minimum found so far. MultiStart runs
all start points (or, optionally, all start points that are feasible with respect
to bounds or inequality constraints).
MultiStart gives a choice of local solver: fmincon, fminunc, lsqcurvefit,
or lsqnonlin. The GlobalSearch algorithm uses fmincon.
MultiStart can run in parallel, distributing start points to multiple
processors for local solution. To run MultiStart in parallel, see How to
Use Parallel Processing on page 8-11.

3-36

How GlobalSearch and MultiStart Work

Deciding Which Solver to Use

The differences between these solver objects boil down to the following
decision on which to use:
Use GlobalSearch to find a single global minimum most efficiently on a
single processor.
Use MultiStart to:

Find multiple local minima.

Run in parallel.
Use a solver other than fmincon.
Search thoroughly for a global minimum.
Explore your own start points.

GlobalSearch Algorithm
For a description of the algorithm, see Ugray et al. [1].
When you run a GlobalSearch object, the algorithm performs the following
steps:
Run fmincon from x0 on page 3-38
Generate Trial Points on page 3-38
Obtain Stage 1 Start Point, Run on page 3-38
Initialize Basins, Counters, Threshold on page 3-38
Begin Main Loop on page 3-39
Examine Stage 2 Trial Point to See if fmincon Runs on page 3-39
When fmincon Runs: on page 3-39
When fmincon Does Not Run: on page 3-40
Create GlobalOptimSolution on page 3-41

3-37

Using GlobalSearch and MultiStart

Run fmincon from x0

GlobalSearch runs fmincon from the start point you give in the problem
structure. If this run converges, GlobalSearch records the start point

and end point for an initial estimate on the radius of a basin of attraction.
Furthermore, GlobalSearch records the final objective function value for use
in the score function (see Obtain Stage 1 Start Point, Run on page 3-38).
The score function is the sum of the objective function value at a point and a
multiple of the sum of the constraint violations. So a feasible point has score
equal to its objective function value. The multiple for constraint violations is
initially 1000. GlobalSearch updates the multiple during the run.

Generate Trial Points

GlobalSearch uses the scatter search algorithm to generate a set of
NumTrialPoints trial points. Trial points are potential start points. For a

description of the scatter search algorithm, see Glover [2]. The trial points
have components in the range (1e4+1,1e4+1). This range is not symmetric
about the origin so that the origin is not in the scatter search.

Obtain Stage 1 Start Point, Run

GlobalSearch evaluates the score function of a set of NumStageOnePoints
trial points. It then takes the point with the best score and runs fmincon
from that point. GlobalSearch removes the set of NumStageOnePoints trial

points from its list of points to examine.

Initialize Basins, Counters, Threshold

The localSolverThreshold is initially the smaller of the two objective
function values at the solution points. The solution points are the fmincon
solutions starting from x0 and from the Stage 1 start point. If both of these
solution points do not exist or are infeasible, localSolverThreshold is
initially the penalty function value of the Stage 1 start point.
The GlobalSearch heuristic assumption is that basins of attraction are
spherical. The initial estimate of basins of attraction for the solution point
from x0 and the solution point from Stage 1 are spheres centered at the
solution points. The radius of each sphere is the distance from the initial
point to the solution point. These estimated basins can overlap.

3-38

How GlobalSearch and MultiStart Work

There are two sets of counters associated with the algorithm. Each counter is
the number of consecutive trial points that:
Lie within a basin of attraction. There is one counter for each basin.
Have score function greater than localSolverThreshold. For a definition
of the score, see Run fmincon from x0 on page 3-38.
All counters are initially 0.

Begin Main Loop

GlobalSearch repeatedly examines a remaining trial point from the list, and
performs the following steps. It continually monitors the time, and stops the
search if elapsed time exceeds MaxTime seconds.

Examine Stage 2 Trial Point to See if fmincon Runs

Call the trial point p. Run fmincon from p if the following conditions hold:
p is not in any existing basin. The criterion for every basin i is:
|p - center(i)| > DistanceThresholdFactor * radius(i).
DistanceThresholdFactor is an option (default value 0.75).
radius is an estimated radius that updates in Update Basin Radius and

Threshold and React to Large Counter Values.

score(p) < localSolverThreshold.
(optional) p satisfies bound and/or inequality constraints. This test occurs
if you set the StartPointsToRun property of the GlobalSearch object to
'bounds' or 'bounds-ineqs'.

When fmincon Runs:

1 Reset CountersSet the counters for basins and threshold to 0.
2 Update Solution SetIf fmincon runs starting from p, it can yield a positive

exit flag, which indicates convergence. In that case, GlobalSearch updates

the vector of GlobalOptimSolution objects. Call the solution point xp and
the objective function value fp. There are two cases:

3-39

Using GlobalSearch and MultiStart

For every other solution point xq with objective function value fq,
|xq - xp| > TolX * max(1,|xp|)

or
|fq - fp| > TolFun * max(1,|fp|).

In this case, GlobalSearch creates a new element in the vector of

GlobalOptimSolution objects. For details of the information contained
in each object, see GlobalOptimSolution.
For some other solution point xq with objective function value fq,
|xq - xp| <= TolX * max(1,|xp|)

and
|fq - fp| <= TolFun * max(1,|fp|).

In this case, GlobalSearch regards xp as equivalent to xq. The

GlobalSearch algorithm modifies the GlobalOptimSolution of xq by
adding p to the cell array of X0 points.
There is one minor tweak that can happen to this update. If the exit flag
for xq is greater than 1, and the exit flag for xp is 1, then xp replaces xq.
This replacement can lead to some points in the same basin being more
than a distance of TolX from xp.
3 Update Basin Radius and ThresholdIf the exit flag of the current fmincon

run is positive:
a Set threshold to the score value at start point p.
b Set basin radius for xp equal to the maximum of the existing radius (if

any) and the distance between p and xp.

4 Report to Iterative DisplayWhen the GlobalSearch Display property is

'iter', every point that fmincon runs creates one line in the GlobalSearch

iterative display.

When fmincon Does Not Run:

1 Update CountersIncrement the counter for every basin containing p. Reset

the counter of every other basin to 0.

3-40

How GlobalSearch and MultiStart Work

Increment the threshold counter if score(p) >= localSolverThreshold.

Otherwise, reset the counter to 0.
2 React to Large Counter ValuesFor each basin with counter equal to

MaxWaitCycle, multiply the basin radius by 1 BasinRadiusFactor. Reset

the counter to 0. (Both MaxWaitCycle and BasinRadiusFactor are settable
properties of the GlobalSearch object.)

If the threshold counter equals MaxWaitCycle, increase the threshold:

new threshold = threshold + PenaltyThresholdFactor*(1 +
abs(threshold)).
Reset the counter to 0.
3 Report to Iterative DisplayEvery 200th trial point creates one line in the

GlobalSearch iterative display.

Create GlobalOptimSolution
After reaching MaxTime seconds or running out of trial points, GlobalSearch
creates a vector of GlobalOptimSolution objects. GlobalSearch orders the
vector by objective function value, from lowest (best) to highest (worst). This
concludes the algorithm.

MultiStart Algorithm
When you run a MultiStart object, the algorithm performs the following
steps:
Generate Start Points on page 3-41
Filter Start Points (Optional) on page 3-42
Run Local Solver on page 3-42
Check Stopping Conditions on page 3-43
Create GlobalOptimSolution Object on page 3-43

Generate Start Points

If you call MultiStart with the syntax

3-41

Using GlobalSearch and MultiStart

[x fval] = run(ms,problem,k)

for an integer k, MultiStart generates k - 1 start points exactly as if you

used a RandomStartPointSet object. The algorithm also uses the x0 start
point from the problem structure, for a total of k start points.
A RandomStartPointSet object does not have any points stored inside
the object. Instead, MultiStart calls the list method, which generates
random points within the bounds given by the problem structure. If an
unbounded component exists, list uses an artificial bound given by the
ArtificialBound property of the RandomStartPointSet object.
If you provide a CustomStartPointSet object, MultiStart does not generate
start points, but uses the points in the object.

Filter Start Points (Optional)

If you set the StartPointsToRun property of the MultiStart object to
'bounds' or 'bounds-ineqs', MultiStart does not run the local solver from
infeasible start points. In this context, infeasible means start points that
do not satisfy bounds, or start points that do not satisfy both bounds and
inequality constraints.
The default setting of StartPointsToRun is 'all'. In this case, MultiStart
does not discard infeasible start points.

Run Local Solver

MultiStart runs the local solver specified in problem.solver, starting at the
points that pass the StartPointsToRun filter. If MultiStart is running in
parallel, it sends start points to worker processors one at a time, and the
worker processors run the local solver.

The local solver checks whether MaxTime seconds have elapsed at each of its
iterations. If so, it exits that iteration without reporting a solution.
When the local solver stops, MultiStart stores the results and continues
to the next step.

3-42

How GlobalSearch and MultiStart Work

Report to Iterative Display. When the MultiStart Display property

is 'iter', every point that the local solver runs creates one line in the
MultiStart iterative display.

Check Stopping Conditions

MultiStart stops when it runs out of start points. It also stops when it
exceeds a total run time of MaxTime seconds.

Create GlobalOptimSolution Object

After MultiStart reaches a stopping condition, the algorithm creates a vector
of GlobalOptimSolution objects as follows:
1 Sort the local solutions by objective function value (Fval) from lowest to

highest. For the lsqnonlin and lsqcurvefit local solvers, the objective
function is the norm of the residual.
2 Loop over the local solutions j beginning with the lowest (best) Fval.
3 Find all the solutions k satisfying both:

|Fval(k) - Fval(j)| <= TolFun*max(1,|Fval(j)|)

|x(k) - x(j)| <= TolX*max(1,|x(j)|)
4 Record j, Fval(j), the local solver output structure for j, and a cell

array of the start points for j and all the k. Remove those points k
from the list of local solutions. This point is one entry in the vector of
GlobalOptimSolution objects.
The resulting vector of GlobalOptimSolution objects is in order by Fval,
from lowest (best) to highest (worst).
Report to Iterative Display. After examining all the local solutions,

MultiStart gives a summary to the iterative display. This summary includes

the number of local solver runs that converged, the number that failed to
converge, and the number that had errors.

3-43

Using GlobalSearch and MultiStart

Bibliography
[1] Ugray, Zsolt, Leon Lasdon, John C. Plummer, Fred Glover, James Kelly,
and Rafael Mart. Scatter Search and Local NLP Solvers: A Multistart
Framework for Global Optimization. INFORMS Journal on Computing, Vol.
19, No. 3, 2007, pp. 328340.
[2] Glover, F. A template for scatter search and path relinking. Artificial
Evolution (J.-K. Hao, E.Lutton, E.Ronald, M.Schoenauer, D.Snyers, eds.).
Lecture Notes in Computer Science, 1363, Springer, Berlin/Heidelberg, 1998,
pp. 1354.
[3] Dixon, L. and G. P. Szeg. The Global Optimization Problem: an
Introduction. Towards Global Optimisation 2 (Dixon, L. C. W. and G. P.
Szeg, eds.). Amsterdam, The Netherlands: North Holland, 1978.

3-44

Improving Results

Improving Results
In this section...
Can You Certify a Solution Is Global? on page 3-45
Refining the Start Points on page 3-48
Changing Options on page 3-55
Reproducing Results on page 3-59

Can You Certify a Solution Is Global?

How can you tell if you have located the global minimum of your objective
function? The short answer is that you cannot; you have no guarantee that
the result of a Global Optimization Toolbox solver is a global optimum.
However, you can use the following strategies for investigating solutions:
Check if a Solution Is a Local Solution with patternsearch on page 3-45
Identify a Bounded Region That Contains a Global Solution on page 3-46
Use MultiStart with More Start Points on page 3-47

Check if a Solution Is a Local Solution with patternsearch

Before you can determine if a purported solution is a global minimum, first
check that it is a local minimum. To do so, run patternsearch on the problem.
To convert the problem to use patternsearch instead of fmincon or fminunc,
enter
problem.solver = 'patternsearch';

Also, change the start point to the solution you just found:
problem.x0 = x;

For example, Check Nearby Points (in the Optimization Toolbox

documentation) shows the following:
options = optimset('Algorithm','active-set');

3-45

Using GlobalSearch and MultiStart

ffun = @(x)(x(1)-(x(1)-x(2))^2);
problem = createOptimProblem('fmincon', ...
'objective',ffun,'x0',[1/2 1/3], ...
'lb',[0 -1],'ub',[1 1],'options',options);
[x fval exitflag] = fmincon(problem)
x =
1.0e-007 *
0

0.1614

fval =
-2.6059e-016
exitflag =
1

However, checking this purported solution with patternsearch shows that

there is a better solution. Start patternsearch from the reported solution x:
% set the candidate solution x as the start point
problem.x0 = x;
problem.solver = 'patternsearch';
[xp fvalp exitflagp] = patternsearch(problem)
xp =
1.0000

-1.0000

fvalp =
-3.0000
exitflagp =
1

Identify a Bounded Region That Contains a Global Solution

Suppose you have a smooth objective function in a bounded region. Given
enough time and start points, MultiStart eventually locates a global solution.
Therefore, if you can bound the region where a global solution can exist, you
can obtain some degree of assurance that MultiStart locates the global
solution.

3-46

Improving Results

For example, consider the function

x2
f = x6 + y6 + sin( x + y) ( x2 + y2 ) cos
( 2 + x 4 + x 2 y2 + y4 ) .
1 + y2

The initial summands x6 + y6 force the function to become large and positive
for large values of |x| or |y|. The components of the global minimum of the
function must be within the bounds
10 x,y 10,
since 106 is much larger than all the multiples of 104 that occur in the other
summands of the function.
You can identify smaller bounds for this problem; for example, the global
minimum is between 2 and 2. It is more important to identify reasonable
bounds than it is to identify the best bounds.

Use MultiStart with More Start Points

To check whether there is a better solution to your problem, run MultiStart
with additional start points. Use MultiStart instead of GlobalSearch for this
task because GlobalSearch does not run the local solver from all start points.
For example, see Example: Searching for a Better Solution on page 3-52.
Updating Unconstrained Problem from GlobalSearch. If you use

GlobalSearch on an unconstrained problem, change your problem structure

before using MultiStart. You have two choices in updating a problem
structure for an unconstrained problem using MultiStart:

Change the solver field to 'fminunc':

problem.solver = 'fminunc';

To avoid a warning if your objective function does not compute a gradient,

change the local options structure to have LargeScale set to 'off':
problem.options.LargeScale = 'off';

3-47

Using GlobalSearch and MultiStart

Add an artificial constraint, retaining fmincon as the local solver:

problem.lb = -Inf;

To search a larger region than the default, see Refining the Start Points
on page 3-48.

Refining the Start Points

If some components of your problem are unconstrained, GlobalSearch and
MultiStart use artificial bounds to generate random start points uniformly
in each component. However, if your problem has far-flung minima, you need
widely dispersed start points to find these minima.
Use these methods to obtain widely dispersed start points:
Give widely separated bounds in your problem structure.
Use a RandomStartPointSet object with the MultiStart algorithm. Set a
large value of the ArtificialBound property in the RandomStartPointSet
object.
Use a CustomStartPointSet object with the MultiStart algorithm. Use
widely dispersed start points.
There are advantages and disadvantages of each method.
Method

Advantages

Disadvantages

Give bounds in problem

Automatic point generation

Makes a more complex Hessian

Can use with GlobalSearch

Unclear how large to set the

bounds

Easy to do

Changes problem

Bounds can be asymmetric

Only uniform points

Large ArtificialBound in

Automatic point generation

MultiStart only

RandomStartPointSet

Does not change problem

Only symmetric, uniform points

Easy to do

Unclear how large to set

ArtificialBound

3-48

Improving Results

Method

Advantages

Disadvantages

CustomStartPointSet

Customizable

MultiStart only

Does not change problem

Requires programming for

generating points

Methods of Generating Start Points

Uniform Grid on page 3-49
Perturbed Grid on page 3-50
Widely Dispersed Points for Unconstrained Components on page 3-50
Uniform Grid. To generate a uniform grid of start points:
1 Generate multidimensional arrays with ndgrid. Give the lower bound,

spacing, and upper bound for each component.

For example, to generate a set of three-dimensional arrays with
First component from 2 through 0, spacing 0.5
Second component from 0 through 2, spacing 0.25
Third component from 10 through 5, spacing 1
[X,Y,Z] = ndgrid(-2:.5:0,0:.25:2,-10:5);
2 Place the arrays into a single matrix, with each row representing one start

point. For example:

W = [X(:),Y(:),Z(:)];

In this example, W is a 720-by-3 matrix.

3 Put the matrix into a CustomStartPointSet object. For example:

custpts = CustomStartPointSet(W);

Call MultiStart run with the CustomStartPointSet object as the third

input. For example,

3-49

Using GlobalSearch and MultiStart

% Assume problem structure and ms MultiStart object exist

[x fval flag outpt manymins] = run(ms,problem,custpts);

Perturbed Grid. Integer start points can yield less robust solutions than
slightly perturbed start points.
To obtain a perturbed set of start points:
1 Generate a matrix of start points as in steps 12 of Uniform Grid on

page 3-49.
2 Perturb the start points by adding a random normal matrix with 0 mean

and relatively small variance.

For the example in Uniform Grid on page 3-49, after making the W matrix,
add a perturbation:
[X,Y,Z] = ndgrid(-2:.5:0,0:.25:2,-10:5);
W = [X(:),Y(:),Z(:)];
W = W + 0.01*randn(size(W));
3 Put the matrix into a CustomStartPointSet object. For example:

custpts = CustomStartPointSet(W);

Call MultiStart run with the CustomStartPointSet object as the third

input. For example,
% Assume problem structure and ms MultiStart object exist
[x fval flag outpt manymins] = run(ms,problem,custpts);

Widely Dispersed Points for Unconstrained Components. Some

components of your problem can lack upper or lower bounds. For example:
Although no explicit bounds exist, there are levels that the components
cannot attain. For example, if one component represents the weight of a
single diamond, there is an implicit upper bound of 1 kg (the Hope Diamond
is under 10 g). In such a case, give the implicit bound as an upper bound.
There truly is no upper bound. For example, the size of a computer file in
bytes has no effective upper bound. The largest size can be in gigabytes
or terabytes today, but in 10 years, who knows?

3-50

Improving Results

For truly unbounded components, you can use the following methods
of sampling. To generate approximately 1/n points in each region
(exp(n),exp(n+1)), use the following formula. If u is random and uniformly
distributed from 0 through 1, then r = 2u 1 is uniformly distributed between
1 and 1. Take

y = sgn(r) ( exp (1 / r ) e ) .
y is symmetric and random. For a variable bounded below by lb, take

y = lb + ( exp (1 / u ) e ) .
Similarly, for variable bounded above by ub, take

y = ub ( exp (1 / u ) e ) .
For example, suppose you have a three-dimensional problem with
x(1) > 0
x(2) < 100
x(3) unconstrained
To make 150 start points satisfying these constraints:
u = rand(150,3);
r1 = 1./u(:,1);
r1 = exp(r1) - exp(1);
r2 = 1./u(:,2);
r2 = -exp(r2) + exp(1) + 100;
r3 = 1./(2*u(:,3)-1);
r3 = sign(r3).*(exp(abs(r3)) - exp(1));
custpts = CustomStartPointSet([r1,r2,r3]);

The following is a variant of this algorithm. Generate a number between 0

and infinity by the method for lower bounds. Use this number as the radius
of a point. Generate the other components of the point by taking random
numbers for each component and multiply by the radius. You can normalize
the random numbers, before multiplying by the radius, so their norm is 1. For

3-51

Using GlobalSearch and MultiStart

a worked example of this method, see MultiStart Without Bounds, Widely

Dispersed Start Points on page 3-81.

Example: Searching for a Better Solution

MultiStart fails to find the global minimum in Multiple Local Minima Via

MultiStart on page 3-67. There are two simple ways to search for a better
solution:
Use more start points
Give tighter bounds on the search space
Set up the problem structure and MultiStart object:
problem = createOptimProblem('fminunc',...
'objective',@(x)sawtoothxy(x(1),x(2)),...
'x0',[100,-50],'options',...
optimset('LargeScale','off'));
ms = MultiStart;

Use More Start Points. Run MultiStart on the problem for 200 start
points instead of 50:
[x fval eflag output manymins] = run(ms,problem,200)
MultiStart completed some of the runs from the start points.
53 out of 200 local solver runs converged with a positive
local solver exit flag.
x =
1.0e-005 *
0.7460
-0.2550
fval =
7.8236e-010
eflag =
2
output =

3-52

Improving Results

funcCount:
localSolverTotal:
localSolverSuccess:
localSolverIncomplete:
localSolverNoSolution:
message:
manymins =
1x53 GlobalOptimSolution

32841
200
53
147
0
[1x143 char]

Properties:
X
Fval
Exitflag
Output
X0

This time MultiStart found the global minimum, and found 53 total local
minima.
To see the range of local solutions, enter hist([manymins.Fval]).

3-53

Using GlobalSearch and MultiStart

Tighter Bound on the Start Points. Suppose you believe that the
interesting local solutions have absolute values of all components less
than 100. The default value of the bound on start points is 1000. To use
a different value of the bound, generate a RandomStartPointSet with the
ArtificialBound property set to 100:
startpts = RandomStartPointSet('ArtificialBound',100,...
'NumStartPoints',50);
[x fval eflag output manymins] = run(ms,problem,startpts)
MultiStart completed some of the runs from the start points.
29 out of 50 local solver runs converged with a
positive local solver exit flag.
x =
1.0e-007 *
0.1771

0.3198

fval =
1.3125e-014
eflag =
2
output =
funcCount:
localSolverTotal:
localSolverSuccess:
localSolverIncomplete:
localSolverNoSolution:
message:
manymins =
1x25 GlobalOptimSolution
Properties:
X
Fval
Exitflag

3-54

7557
50
29
21
0
[1x142 char]

Improving Results

Output
X0
MultiStart found the global minimum, and found 25 distinct local solutions.
To see the range of local solutions, enter hist([manymins.Fval]).

Compared to the minima found in Use More Start Points on page 3-52, this
run found better (smaller) minima, and had a higher percentage of successful
runs.

Changing Options
How to Determine Which Options to Change
After you run a global solver, you might want to change some global or local
options. To determine which options to change, the guiding principle is:
To affect the local solver, set local solver options.
To affect the start points or solution set, change the problem structure, or
set the global solver object properties.
For example, to obtain:
More local minima Set global solver object properties.
Faster local solver iterations Set local solver options.
Different tolerances for considering local solutions identical (to obtain more
or fewer local solutions) Set global solver object properties.

3-55

Using GlobalSearch and MultiStart

Different information displayed at the command line Decide if you want

iterative display from the local solver (set local solver options) or global
information (set global solver object properties).
Different bounds, to examine different regions Set the bounds in the
problem structure.
Examples of Choosing Problem Options.
To start your local solver at points only satisfying inequality constraints,
set the StartPointsToRun property in the global solver object to
'bounds-ineqs'. This setting can speed your solution, since local solvers
do not have to attempt to find points satisfying these constraints. However,
the setting can result in many fewer local solver runs, since the global
solver can reject many start points. For an example, see Example: Using
Only Feasible Start Points on page 3-70.
To use the fmincon interior-point algorithm, set the local solver
Algorithm option to 'interior-point'. For an example showing how to
do this, see Examples of Updating Problem Options on page 3-57.
For your local solver to have different bounds, set the bounds in the
problem structure. Examine different regions by setting bounds.
To see every solution that has positive local exit flag, set the TolX property
in the global solver object to 0. For an example showing how to do this, see
Changing Global Options on page 3-57.

Changing Local Solver Options

There are several ways to change values in a local options structure:
Update the values using dot addressing and optimset. The syntax is
problem.options =
optimset(problem.options,'Parameter',value,...);

You can also replace the local options entirely:

problem.options = optimset('Parameter',value,...);

Use dot addressing on one local option. The syntax is

3-56

Improving Results

problem.options.Parameter = newvalue;

Recreate the entire problem structure. For details, see Create a Problem
Structure on page 3-4.
Examples of Updating Problem Options.
1 Create a problem structure:

problem = createOptimProblem('fmincon','x0',[-1 2], ...

'objective',@rosenboth);
2 Set the problem to use the sqp algorithm in fmincon:

problem.options.Algorithm = 'sqp';
3 Update the problem to use the gradient in the objective function, have a

TolFun value of 1e-8, and a TolX value of 1e-7:

problem.options = optimset(problem.options,'GradObj','on', ...
'TolFun',1e-8,'TolX',1e-7);

Changing Global Options

There are several ways to change characteristics of a GlobalSearch or
MultiStart object:
Use dot addressing. For example, suppose you have a default MultiStart
object:
ms = MultiStart
Properties:
UseParallel:
StartPointsToRun:
Display:
TolFun:
TolX:
MaxTime:

'never'
'all'
'final'
1.0000e-006
1.0000e-006
Inf

To change ms to have its TolX value equal to 1e-3, update the TolX field:

3-57

Using GlobalSearch and MultiStart

ms.TolX = 1e-3
Properties:
UseParallel:
StartPointsToRun:
Display:
TolFun:
TolX:
MaxTime:

'never'
'all'
'final'
1.0000e-006
1.0000e-003
Inf

Reconstruct the object starting from the current settings. For example, to
set the TolFun field in ms to 1e-3, retaining the nondefault value for TolX:
ms = MultiStart(ms,'TolFun',1e-3)
Properties:
UseParallel:
StartPointsToRun:
Display:
TolFun:
TolX:
MaxTime:

'never'
'all'
'final'
1.0000e-003
1.0000e-003
Inf

Convert a GlobalSearch object to a MultiStart object, or vice-versa.

For example, with the ms object from the previous example, create a
GlobalSearch object with the same values of TolX and TolFun:
gs = GlobalSearch(ms)
Properties:
NumTrialPoints:
BasinRadiusFactor:
DistanceThresholdFactor:
MaxWaitCycle:
NumStageOnePoints:
PenaltyThresholdFactor:
StartPointsToRun:
Display:
TolFun:
TolX:
MaxTime:

3-58

1000
0.2000
0.7500
20
200
0.2000
'all'
'final'
1.0000e-003 % from ms
1.0000e-003 % from ms
Inf

Improving Results

Reproducing Results
GlobalSearch and MultiStart use pseudorandom numbers in choosing start
points. Use the same pseudorandom number stream again to:

Compare various algorithm settings.

Have an example run repeatably.
Extend a run, with known initial segment of a previous run.
Both GlobalSearch and MultiStart use the default random number stream.

Steps to Take in Reproducing Results

1 Before running your problem, store the current state of the default random

number stream:
stream = RandStream.getDefaultStream;
strmstate = stream.State;
2 Run your GlobalSearch or MultiStart problem.
3 Restore the state of the random number stream:

stream.State = strmstate;
4 If you run your problem again, you get the same result.

Example: Reproducing a GlobalSearch or MultiStart Result

This example shows how to obtain reproducible results for Example: Finding
Global or Multiple Local Minima on page 3-63. The example follows the
procedure in Steps to Take in Reproducing Results on page 3-59.
1 Store the current state of the default random number stream:

stream = RandStream.getDefaultStream;
strmstate = stream.State;
2 Create the sawtoothxy function file:

function f = sawtoothxy(x,y)

3-59

Using GlobalSearch and MultiStart

[t r] = cart2pol(x,y); % change to polar coordinates

h = cos(2*t - 1/2)/2 + cos(t) + 2;
g = (sin(r) - sin(2*r)/2 + sin(3*r)/3 - sin(4*r)/4 + 4) ...
.*r.^2./(r+1);
f = g.*h;
end
3 Create the problem structure and GlobalSearch object:

problem = createOptimProblem('fmincon',...
'objective',@(x)sawtoothxy(x(1),x(2)),...
'x0',[100,-50],'options',...
optimset('Algorithm','sqp'));
gs = GlobalSearch('Display','iter');
4 Run the problem:

[x fval] = run(gs,problem)

Best

Current

Threshold

Local

Local

Analyzed

Num Pts
F-count

f(x)

Penalty

Penalty

f(x)

exitflag

Procedure

465

422.9

422.9

Initial Point

200

1730

1.547e-015

1.547e-015

300

1830

1.547e-015

6.01e+004

400

1930

1.547e-015

1.47e+005

4.16

Stage 2 Search

500

2030

1.547e-015

2.63e+004

11.84

Stage 2 Search

600

2130

1.547e-015

1.341e+004

30.95

Stage 2 Search

700

2230

1.547e-015

2.562e+004

65.25

Stage 2 Search

800

2330

1.547e-015

5.217e+004

163.8

Stage 2 Search

900

2430

1.547e-015

7.704e+004

409.2

981

2587

1.547e-015

42.24

516.6

1000

2606

1.547e-015

3.299e+004

42.24

1.074

Stage 2 Search
7.573

GlobalSearch stopped because it analyzed all the trial points.

All 3 local solver runs converged with a positive local solver exit flag.

x =
1.0e-007 *
0.0414

3-60

0.1298

Stage 1 Local
Stage 2 Search

Stage 2 Local
Stage 2 Search

Improving Results

fval =
1.5467e-015

You might obtain a different result when running this problem, since the
random stream was in an unknown state at the beginning of the run.
5 Restore the state of the random number stream:

stream.State = strmstate;
6 Run the problem again. You get the same output.

[x fval] = run(gs,problem)

Best

Current

Threshold

Local

Local

Analyzed

Num Pts
F-count

f(x)

Penalty

Penalty

f(x)

exitflag

Procedure

465

422.9

422.9

Initial Point

200

1730

1.547e-015

1.547e-015

Stage 1 Local

... Output deleted to save space ...

x =
1.0e-007 *
0.0414

0.1298

fval =
1.5467e-015

Parallel Processing and Random Number Streams

You obtain reproducible results from MultiStart when you run the
algorithm in parallel the same way as you do for serial computation. Runs
are reproducible because MultiStart generates pseudorandom start points
locally, and then distributes the start points to parallel processors. Therefore,
the parallel processors do not use random numbers.

3-61

Using GlobalSearch and MultiStart

To reproduce a parallel MultiStart run, use the procedure described in Steps

to Take in Reproducing Results on page 3-59. For a description of how to run
MultiStart in parallel, see How to Use Parallel Processing on page 8-11.

3-62

GlobalSearch and MultiStart Examples

GlobalSearch and MultiStart Examples

In this section...
Example: Finding Global or Multiple Local Minima on page 3-63
Example: Using Only Feasible Start Points on page 3-70
Example: Parallel MultiStart on page 3-74
Example: Isolated Global Minimum on page 3-77

Example: Finding Global or Multiple Local Minima

This example illustrates how GlobalSearch finds a global minimum
efficiently, and how MultiStart finds many more local minima.
The objective function for this example has many local minima and a unique
global minimum. In polar coordinates, the function is
f(r,t) = g(r)h(t),
where
2
sin(2r) sin(3r) sin(4 r)

r
g(r) = sin(r)
+

+ 4
2
3
4

r +1
1

cos 2t

2.
h(t) = 2 + cos(t) +
2

3-63

Using GlobalSearch and MultiStart

The global minimum is at r = 0, with objective function 0. The function g(r)

grows approximately linearly in r, with a repeating sawtooth shape. The
function h(t) has two local minima, one of which is global.

3-64

GlobalSearch and MultiStart Examples

Single Global Minimum Via GlobalSearch

1 Write a function file to compute the objective:

function f = sawtoothxy(x,y)
[t r] = cart2pol(x,y); % change to polar coordinates
h = cos(2*t - 1/2)/2 + cos(t) + 2;
g = (sin(r) - sin(2*r)/2 + sin(3*r)/3 - sin(4*r)/4 + 4) ...
.*r.^2./(r+1);
f = g.*h;
end

3-65

Using GlobalSearch and MultiStart

2 Create the problem structure. Use the 'sqp' algorithm for fmincon:

problem = createOptimProblem('fmincon',...
'objective',@(x)sawtoothxy(x(1),x(2)),...
'x0',[100,-50],'options',...
optimset('Algorithm','sqp'));

The start point is [100,-50] instead of [0,0], so GlobalSearch does not

start at the global solution.
3 Add an artificial bound, and validate the problem structure by running

fmincon:
problem.lb = -Inf;
[x fval] = fmincon(problem)
x =
45.6965 -107.6645
fval =
555.6941
4 Create the GlobalSearch object, and set iterative display:

gs = GlobalSearch('Display','iter');
5 Run the solver:

[x fval] = run(gs,problem)

Num Pts

3-66

Best

Current

Threshold

f(x)

Penalty

Penalty

Local

F-count

465

422.9

200

1730

1.547e-015

300

1830

1.547e-015

5.709e+004

400

1930

1.547e-015

1.646e+005

4.16

Stage 2 Search

500

2030

1.547e-015

2.262e+004

11.84

Stage 2 Search

600

2130

1.547e-015

1680

30.95

Stage 2 Search

700

2230

1.547e-015

1.138e+004

65.25

Stage 2 Search

800

2330

1.547e-015

1.573e+005

163.8

Stage 2 Search

1.074

f(x)

Local

Analyzed

exitflag

Procedure

422.9

Initial Point

1.547e-015

Stage 1 Local
Stage 2 Search

GlobalSearch and MultiStart Examples

900

2430

1.547e-015

3.676e+004

409.2

977

3147

1.547e-015

669

707.8

1000

3170

1.547e-015

1634

505.5

Stage 2 Search
445.5

Stage 2 Local
Stage 2 Search

GlobalSearch stopped because it analyzed all the trial points.

All 3 local solver runs converged with a positive local solver exit flag.

x =
1.0e-007 *
0.0414

0.1298

fval =
1.5467e-015

You can get different results, since GlobalSearch is stochastic.

The solver found three local minima, and it found the global minimum near
[0,0].

Multiple Local Minima Via MultiStart

1 Write a function file to compute the objective:

function f = sawtoothxy(x,y)
[t r] = cart2pol(x,y); % change to polar coordinates
h = cos(2*t - 1/2)/2 + cos(t) + 2;
g = (sin(r) - sin(2*r)/2 + sin(3*r)/3 - sin(4*r)/4 + 4) ...
.*r.^2./(r+1);
f = g.*h;
end
2 Create the problem structure. Use the fminunc algorithm with the

LargeScale option set to 'off'. The reasons for these choices are:

The problem is unconstrained. Therefore, fminunc is the appropriate

solver; see Optimization Decision Table in the Optimization Toolbox
documentation.

3-67

Using GlobalSearch and MultiStart

The fminunc LargeScale algorithm requires a gradient; see Choosing

the Algorithm in the Optimization Toolbox documentation. Therefore,
set LargeScale to 'off'.
problem = createOptimProblem('fminunc',...
'objective',@(x)sawtoothxy(x(1),x(2)),...
'x0',[100,-50],'options',...
optimset('LargeScale','off'));
3 Validate the problem structure by running it:

[x fval] = fminunc(problem)
x =
8.4420 -110.2602
fval =
435.2573
4 Create a default MultiStart object:

ms = MultiStart;
5 Run the solver for 50 iterations, recording the local minima:

[x fval eflag output manymins] = run(ms,problem,50)

MultiStart completed some of the runs from the start points.
16 out of 50 local solver runs converged with a positive
local solver exit flag.
x =
-379.3434

559.6154

fval =
1.7590e+003
eflag =
2

3-68

GlobalSearch and MultiStart Examples

output =
funcCount:
localSolverTotal:
localSolverSuccess:
localSolverIncomplete:
localSolverNoSolution:
message:
manymins =
1x16 GlobalOptimSolution

7803
50
16
34
0
[1x142 char]

Properties:
X
Fval
Exitflag
Output
X0

You can get different results, since MultiStart is stochastic.

The solver did not find the global minimum near [0,0]. It found 16 distinct
local minima.
6 Plot the function values at the local minima:

hist([manymins.Fval])

3-69

Using GlobalSearch and MultiStart

Plot the function values at the three best points:

bestf = [manymins.Fval];
hist(bestf(1:3))

MultiStart started fminunc from start points with components uniformly

distributed between 1000 and 1000. fminunc often got stuck in one of
the many local minima. fminunc exceeded its iteration limit or function
evaluation limit 34 times.

Example: Using Only Feasible Start Points

You can set the StartPointsToRun option so that MultiStart and
GlobalSearch use only start points that satisfy inequality constraints. This
option can speed your optimization, since the local solver does not have to
search for a feasible region. However, the option can cause the solvers to
miss some basins of attraction.
There are three settings for the StartPointsToRun option:
all Accepts all start points
bounds Rejects start points that do not satisfy bounds
bounds-ineqs Rejects start points that do not satisfy bounds or
inequality constraints

3-70

GlobalSearch and MultiStart Examples

For example, suppose your objective function is

function y = tiltcircle(x)
vx = x(:)-[4;4]; % ensure vx is in column form
y = vx'*[1;1] + sqrt(16 - vx'*vx); % complex if norm(x-[4;4])>4
tiltcircle returns complex values for norm(x - [4 4]) > 4.

Write a constraint function that is positive on the set where norm(x - [4

4]) > 4
function [c ceq] = myconstraint(x)
ceq = [];

3-71

Using GlobalSearch and MultiStart

cx = x(:) - [4;4]; % ensure x is a column vector

c = cx'*cx - 16; % negative where tiltcircle(x) is real

Set GlobalSearch to use only start points satisfying inequality constraints:

gs = GlobalSearch('StartPointsToRun','bounds-ineqs');

To complete the example, create a problem structure and run the solver:
opts = optimset('Algorithm','interior-point');
problem = createOptimProblem('fmincon',...
'x0',[4 4],'objective',@tiltcircle,...
'nonlcon',@myconstraint,'lb',[-10 -10],...
'ub',[10 10],'options',opts);
[x,fval,exitflag,output,solutionset] = run(gs,problem)
GlobalSearch stopped because it analyzed all the trial points.
All 5 local solver runs converged with a positive
local solver exit flag.
x =
1.1716

1.1716

fval =
-5.6530
exitflag =
1
output =
funcCount:
localSolverTotal:
localSolverSuccess:
localSolverIncomplete:
localSolverNoSolution:
message:
solutionset =
1x4 GlobalOptimSolution
Properties:

3-72

3302
5
5
0
0
[1x137 char]

GlobalSearch and MultiStart Examples

X
Fval
Exitflag
Output
X0

tiltcircle With Local Minima

Why Does GlobalSearch Find Several Local Minima?

The tiltcircle function has just one local minimum. Yet GlobalSearch
(fmincon) stops at several points. Does this mean fmincon makes an error?

3-73

Using GlobalSearch and MultiStart

The reason that fmincon stops at several boundary points is subtle. The
tiltcircle function has an infinite gradient on the boundary, as you can see
from a one-dimensional calculation:

d
x
16 x2 =
= at x = 4.
dx
16 x2
So there is a huge gradient normal to the boundary. This gradient overwhelms
the small additional tilt from the linear term. As far as fmincon can tell,
boundary points are stationary points for the constrained problem.
This behavior can arise whenever you have a function that has a square root.

Example: Parallel MultiStart

If you have a multicore processor or access to a processor network, you can
use Parallel Computing Toolbox functions with MultiStart. This example
shows how to find multiple minima in parallel for a problem, using a processor
with two cores. The problem is the same as in Multiple Local Minima Via
MultiStart on page 3-67.
1 Write a function file to compute the objective:

function f = sawtoothxy(x,y)
[t r] = cart2pol(x,y); % change to polar coordinates
h = cos(2*t - 1/2)/2 + cos(t) + 2;
g = (sin(r) - sin(2*r)/2 + sin(3*r)/3 - sin(4*r)/4 + 4) ...
.*r.^2./(r+1);
f = g.*h;
end
2 Create the problem structure:

problem = createOptimProblem('fminunc',...
'objective',@(x)sawtoothxy(x(1),x(2)),...
'x0',[100,-50],'options',...
optimset('LargeScale','off'));
3 Validate the problem structure by running it:

3-74

GlobalSearch and MultiStart Examples

[x fval] = fminunc(problem)
x =
8.4420 -110.2602
fval =
435.2573
4 Create a MultiStart object, and set the object to use parallel processing

and iterative display:

ms = MultiStart('UseParallel','always','Display','iter');
5 Set up parallel processing:

matlabpool open 2
Starting matlabpool using the 'local' configuration
... connected to 2 labs.
6 Run the problem on 50 start points:

[x fval eflag output manymins] = run(ms,problem,50);

Running the local solvers in parallel.
Run
Index
17
16
34
33

Local
Local
Local
exitflag
f(x)
# iter
2
3953
4
0
1331
45
0
7271
54
2
8249
4
... Many iterations omitted ...
47
2
2740
5
35
0
8501
48
50
0
1225
40

Local
F-count
21
201
201
18

First-order
optimality
0.1626
65.02
520.9
2.968

21
201
201

0.0422
424.8
21.89

MultiStart completed some of the runs from the start points.

17 out of 50 local solver runs converged with a positive
local solver exit flag.

3-75

Using GlobalSearch and MultiStart

Notice that the run indexes look random. Parallel MultiStart runs its
start points in an unpredictable order.
Notice that MultiStart confirms parallel processing in the first line of
output, which states: Running the local solvers in parallel.
7 When finished, shut down the parallel environment:

matlabpool close
Sending a stop signal to all the labs ... stopped.

For an example of how to obtain better solutions to this problem, see

Example: Searching for a Better Solution on page 3-52. You can use parallel
processing along with the techniques described in that example.

Speedup with Parallel Computing

The results of MultiStart runs are stochastic. The timing of runs is
stochastic, too. Nevertheless, some clear trends are apparent in the following
table. The data for the table came from one run at each number of start
points, on a machine with two cores.
Start Points

Parallel Seconds

Serial Seconds

50

3.6

3.4

100

4.9

5.7

200

8.3

10

500

16

23

1000

31

46

Parallel computing can be slower than serial when you use only a few start
points. As the number of start points increases, parallel computing becomes
increasingly more efficient than serial.
There are many factors that affect speedup (or slowdown) with parallel
processing. For more information, see Improving Performance with Parallel
Computing in the Optimization Toolbox documentation.

3-76

GlobalSearch and MultiStart Examples

Example: Isolated Global Minimum

Finding a start point in the basin of attraction of the global minimum can be
difficult when the basin is small or when you are unsure of the location of the
minimum. To solve this type of problem you can:
Add sensible bounds
Take a huge number of random start points
Make a methodical grid of start points
For an unconstrained problem, take widely dispersed random start points
This example shows these methods and some variants.
The function sech(x) is nearly 0 for all |x| > 5, and sech(0) = 1. The
example is a two-dimensional version of the sech function, with one minimum
at [1,1], the other at [1e5,-1e5]:
f(x,y) = 10sech(|x (1,1)|) 20sech(.0003(|x (1e5,1e5)|) 1.
f has a global minimum of 21 at (1e5,1e5), and a local minimum of 11
at (1,1).

3-77

Using GlobalSearch and MultiStart

The minimum at (1e5,1e5) shows as a narrow spike. The minimum at (1,1)

does not show since it is too narrow.
The following sections show various methods of searching for the global
minimum. Some of the methods are not successful on this problem.
Nevertheless, you might find each method useful for different problems.
Default Settings Cannot Find the Global Minimum Add Bounds on
page 3-79
GlobalSearch with Bounds and More Start Points on page 3-79
MultiStart with Bounds and Many Start Points on page 3-80

3-78

GlobalSearch and MultiStart Examples

MultiStart Without Bounds, Widely Dispersed Start Points on page 3-81

MultiStart with a Regular Grid of Start Points on page 3-81
MultiStart with Regular Grid and Promising Start Points on page 3-82

Default Settings Cannot Find the Global Minimum Add

Bounds
GlobalSearch and MultiStart cannot find the global minimum using default

global options, since the default start point components are in the range
(9999,10001) for GlobalSearch and (1000,1000) for MultiStart.
With additional bounds of 1e6 and 1e6 in problem, GlobalSearch usually
does not find the global minimum:
x1 = [1;1];x2 = [1e5;-1e5];
f = @(x)-10*sech(norm(x(:)-x1)) -20*sech((norm(x(:)-x2))*3e-4) -1;
problem = createOptimProblem('fmincon','x0',[0,0],'objective',f,...
'lb',[-1e6;-1e6],'ub',[1e6;1e6]);
gs = GlobalSearch;
[xfinal fval] = run(gs,problem)
GlobalSearch stopped because it analyzed all the trial points.
All 57 local solver runs converged with a positive
local solver exit flag.
xfinal =
1.0000
1.0000
fval =
-11.0000

GlobalSearch with Bounds and More Start Points

To find the global minimum, you can search more points. This example uses
1e5 start points, and a MaxTime of 300 s:
gs.NumTrialPoints = 1e5;
gs.MaxTime = 300;

3-79

Using GlobalSearch and MultiStart

[xg fvalg] = run(gs,problem)

GlobalSearch stopped because maximum time is exceeded.
GlobalSearch called the local solver 2186 times before exceeding
the clock time limit (MaxTime = 300 seconds).
1943 local solver runs converged with a positive
local solver exit flag.
xg =
1.0e+005 *
1.0000
-1.0000
fvalg =
-21.0000

In this case, GlobalSearch found the global minimum.

MultiStart with Bounds and Many Start Points

Alternatively, you can search using MultiStart with many start points. This
example uses 1e5 start points, and a MaxTime of 300 s:
ms = MultiStart(gs);
[xm fvalm] = run(ms,problem,1e5)
MultiStart stopped because maximum time was exceeded.
MultiStart called the local solver 17266 times before exceeding
the clock time limit (MaxTime = 300 seconds).
17266 local solver runs converged with a positive
local solver exit flag.
xm =
1.0000
1.0000
fvalm =

3-80

GlobalSearch and MultiStart Examples

-11.0000

In this case, MultiStart failed to find the global minimum.

MultiStart Without Bounds, Widely Dispersed Start Points

You can also use MultiStart to search an unbounded region to find the global
minimum. Again, you need many start points to have a good chance of finding
the global minimum.
The first five lines of code generate 10,000 widely dispersed random
start points using the method described in Widely Dispersed Points for
Unconstrained Components on page 3-50. newprob is a problem structure
using the fminunc local solver and no bounds:
u = rand(1e4,1);
u = 1./u;
u = exp(u) - exp(1);
s = rand(1e4,1)*2*pi;
stpts = [u.*cos(s),u.*sin(s)];
startpts = CustomStartPointSet(stpts);
newprob = createOptimProblem('fminunc','x0',[0;0],'objective',f);
[xcust fcust] = run(ms,newprob,startpts)
MultiStart completed the runs from all start points.
All 10000 local solver runs converged with a positive
local solver exit flag.
xm =
1.0000e+005
fvalm =
-21.0000

In this case, MultiStart found the global minimum.

MultiStart with a Regular Grid of Start Points

You can also use a grid of start points instead of random start points. To
learn how to construct a regular grid for more dimensions, or one that has

3-81

Using GlobalSearch and MultiStart

small perturbations, see Uniform Grid on page 3-49 or Perturbed Grid

on page 3-50.
xx = -1e6:1e4:1e6;
[xxx yyy] = meshgrid(xx,xx);
z = [xxx(:),yyy(:)];
bigstart = CustomStartPointSet(z);
[xgrid fgrid] = run(ms,newprob,bigstart)
MultiStart completed the runs from all start points.
All 10000 local solver runs converged with a positive
local solver exit flag.
xcust =
1.0e+004 *
10.0000
-10.0000
fcust =
-21.0000

In this case, MultiStart found the global minimum.

MultiStart with Regular Grid and Promising Start Points

Making a regular grid of start points, especially in high dimensions, can use
an inordinate amount of memory or time. You can filter the start points to
run only those with small objective function value.
To perform this filtering most efficiently, write your objective function in
a vectorized fashion. For information, see Example: Writing a Vectorized
Function on page 2-3 or Vectorizing the Objective and Constraint Functions
on page 4-72. The following function handle computes a vector of objectives
based on an input matrix whose rows represent start points:
x1 = [1;1];x2 = [1e5;-1e5];
g = @(x) -10*sech(sqrt((x(:,1)-x1(1)).^2 + (x(:,2)-x1(2)).^2)) ...
-20*sech(sqrt((x(:,1)-x2(1)).^2 + (x(:,2)-x2(2)).^2))-1;

3-82

GlobalSearch and MultiStart Examples

Suppose you want to run the local solver only for points where the value is
less than 2. Start with a denser grid than in MultiStart with a Regular
Grid of Start Points on page 3-81, then filter out all the points with high
function value:
xx = -1e6:1e3:1e6;
[xxx yyy] = meshgrid(xx,xx);
z = [xxx(:),yyy(:)];
idx = g(z) < -2; % index of promising start points
zz = z(idx,:);
smallstartset = CustomStartPointSet(zz);
newprobg = createOptimProblem('fminunc','x0',[0,0],...
'objective',g); % row vector x0 since g expects rows
[xfew ffew] = run(ms,newprobg,smallstartset)
MultiStart completed the runs from all start points.
All 2 local solver runs converged with a positive
local solver exit flag.
xfew =
100000

-100000

ffew =
-21

In this case, MultiStart found the global minimum. There are only two start
points in smallstartset, one of which is the global minimum.

3-83

3-84

Using GlobalSearch and MultiStart

4
Using Direct Search
What Is Direct Search? on page 4-2
Performing a Pattern Search on page 4-3
Example: Finding the Minimum of a Function Using the GPS Algorithm
on page 4-7
Pattern Search Terminology on page 4-11
How Pattern Search Works on page 4-14
Description of the Nonlinear Constraint Solver on page 4-24
Performing a Pattern Search Using the Optimization Tool GUI on page
4-26
Performing a Pattern Search from the Command Line on page 4-36
Pattern Search Examples: Setting Options on page 4-42

Using Direct Search

What Is Direct Search?

Direct search is a method for solving optimization problems that does not
require any information about the gradient of the objective function. Unlike
more traditional optimization methods that use information about the
gradient or higher derivatives to search for an optimal point, a direct search
algorithm searches a set of points around the current point, looking for one
where the value of the objective function is lower than the value at the current
point. You can use direct search to solve problems for which the objective
function is not differentiable, or is not even continuous.
Global Optimization Toolbox functions include three direct search algorithms
called the generalized pattern search (GPS) algorithm, the generating set
search (GSS) algorithm, and the mesh adaptive search (MADS) algorithm.
All are pattern search algorithms that compute a sequence of points that
approach an optimal point. At each step, the algorithm searches a set of
points, called a mesh, around the current pointthe point computed at the
previous step of the algorithm. The mesh is formed by adding the current
point to a scalar multiple of a set of vectors called a pattern. If the pattern
search algorithm finds a point in the mesh that improves the objective
function at the current point, the new point becomes the current point at the
next step of the algorithm.
The GPS algorithm uses fixed direction vectors. The GSS algorithm is
identical to the GPS algorithm, except when there are linear constraints, and
when the current point is near a linear constraint boundary. The MADS
algorithm uses a random selection of vectors to define the mesh. For details,
see Patterns on page 4-11.

4-2

Performing a Pattern Search

Performing a Pattern Search

In this section...
Calling patternsearch at the Command Line on page 4-3
Using the Optimization Tool for Pattern Search on page 4-3

Calling patternsearch at the Command Line

To perform a pattern search on an unconstrained problem at the command
line, call the function patternsearch with the syntax
[x fval] = patternsearch(@objfun, x0)

where
@objfun is a handle to the objective function.
x0 is the starting point for the pattern search.
The results are:
x Point at which the final value is attained
fval Final value of the objective function
Performing a Pattern Search from the Command Line on page 4-36 explains
in detail how to use the patternsearch function.

Using the Optimization Tool for Pattern Search

To open the Optimization Tool, enter
optimtool('patternsearch')

at the command line, or enter optimtool and then choose patternsearch

from the Solver menu.

4-3

Using Direct Search

Set options

Expand or contract help

Choose solver

Enter problem
and constraints

Run solver

View results

See final point

You can also start the tool from the MATLAB Start menu as pictured:

4-4

Performing a Pattern Search

To use the Optimization Tool, first enter the following information:

Objective function The objective function you want to minimize. Enter
the objective function in the form @objfun, where objfun.m is a file that
computes the objective function. The @ sign creates a function handle to
objfun.
Start point The initial point at which the algorithm starts the
optimization.

4-5

Using Direct Search

In the Constraints pane, enter linear constraints, bounds, or a nonlinear

constraint function as a function handle for the problem. If the problem is
unconstrained, leave these fields blank.
Then, click Start. The tool displays the results of the optimization in the
Run solver and view results pane.
In the Options pane, set the options for the pattern search. To view the
options in a category, click the + sign next to it.
Finding the Minimum of the Function on page 4-8 gives an example of using
the Optimization Tool.
The Optimization Tool chapter in the Optimization Toolbox documentation
provides a detailed description of the Optimization Tool.

4-6

Example: Finding the Minimum of a Function Using the GPS Algorithm

Example: Finding the Minimum of a Function Using the

GPS Algorithm
In this section...
Objective Function on page 4-7
Finding the Minimum of the Function on page 4-8
Plotting the Objective Function Values and Mesh Sizes on page 4-9

Objective Function
This example uses the objective function, ps_example, which is included
with Global Optimization Toolbox software. View the code for the function
by entering
type ps_example

The following figure shows a plot of the function.

4-7

Using Direct Search

Finding the Minimum of the Function

To find the minimum of ps_example, perform the following steps:
1 Enter
optimtool

and then choose the patternsearch solver.

2 In the Objective function field of the Optimization Tool, enter

@ps_example.
3 In the Start point field, type [2.1 1.7].

Leave the fields in the Constraints pane blank because the problem is
unconstrained.
4 Click Start to run the pattern search.

The Run solver and view results pane displays the results of the pattern
search.

4-8

Example: Finding the Minimum of a Function Using the GPS Algorithm

The reason the optimization terminated is that the mesh size became smaller
than the acceptable tolerance value for the mesh size, defined by the Mesh
tolerance parameter in the Stopping criteria pane. The minimum function
value is approximately 2. The Final point pane displays the point at which
the minimum occurs.

Plotting the Objective Function Values and Mesh Sizes

To see the performance of the pattern search, display plots of the best function
value and mesh size at each iteration. First, select the following check boxes
in the Plot functions pane:
Best function value
Mesh size

4-9

Using Direct Search

Then click Start to run the pattern search. This displays the following plots.

The upper plot shows the objective function value of the best point at each
iteration. Typically, the objective function values improve rapidly at the early
iterations and then level off as they approach the optimal value.
The lower plot shows the mesh size at each iteration. The mesh size increases
after each successful iteration and decreases after each unsuccessful one,
explained in How Pattern Search Works on page 4-14.

4-10

Pattern Search Terminology

Pattern Search Terminology

In this section...
Patterns on page 4-11
Meshes on page 4-12
Polling on page 4-13
Expanding and Contracting on page 4-13

Patterns
A pattern is a set of vectors {vi} that the pattern search algorithm uses to
determine which points to search at each iteration. The set {vi} is defined by
the number of independent variables in the objective function, N, and the
positive basis set. Two commonly used positive basis sets in pattern search
algorithms are the maximal basis, with 2N vectors, and the minimal basis,
with N+1 vectors.
With GPS, the collection of vectors that form the pattern are fixed-direction
vectors. For example, if there are three independent variables in the
optimization problem, the default for a 2N positive basis consists of the
following pattern vectors:

v1 = [1 0 0] v2 = [0 1 0] v3 = [0 0 1]
v4 = [ 1 0 0] v5 = [0 1 0] v6 = [0 0 1]
An N+1 positive basis consists of the following default pattern vectors.

v1 = [1 0 0]

v2 = [0 1 0]

v3 = [0 0 1]

v4 = [ 1 1 1]
With GSS, the pattern is identical to the GPS pattern, except when there are
linear constraints and the current point is near a constraint boundary. For a
description of the way in which GSS forms a pattern with linear constraints,
see Kolda, Lewis, and Torczon [1]. The GSS algorithm is more efficient
than the GPS algorithm when you have linear constraints. For an example
showing the efficiency gain, see Example: Comparing the Efficiency of Poll
Options on page 4-48.

4-11

Using Direct Search

With MADS, the collection of vectors that form the pattern are randomly
selected by the algorithm. Depending on the poll method choice, the number
of vectors selected will be 2N or N+1. As in GPS, 2N vectors consist of N
vectors and their N negatives, while N+1 vectors consist of N vectors and one
that is the negative of the sum of the others.
[1] Kolda, Tamara G., Robert Michael Lewis, and Virginia Torczon.
A generating set direct search augmented Lagrangian algorithm for
optimization with a combination of general and linear constraints. Technical
Report SAND2006-5315, Sandia National Laboratories, August 2006.

Meshes
At each step, patternsearch searches a set of points, called a mesh, for a
point that improves the objective function. patternsearch forms the mesh by
1 Generating a set of vectors {di} by multiplying each pattern vector vi by a

scalar m. m is called the mesh size.

2 Adding the { di } to the current pointthe point with the best objective

function value found at the previous step.

For example, using the GPS algorithm. suppose that:

The current point is [1.6 3.4].
The pattern consists of the vectors

v1 = [1 0]

v2 = [0 1]

v3 = [ 1 0]

v4 = [0 1]
The current mesh size m is 4.
The algorithm multiplies the pattern vectors by 4 and adds them to the
current point to obtain the following mesh.
[1.6 3.4] + 4*[1 0] = [5.6 3.4]

4-12

Pattern Search Terminology

[1.6 3.4] + 4*[0 1] = [1.6 7.4]

[1.6 3.4] + 4*[-1 0] = [-2.4 3.4]
[1.6 3.4] + 4*[0 -1] = [1.6 -0.6]

The pattern vector that produces a mesh point is called its direction.

Polling
At each step, the algorithm polls the points in the current mesh by computing
their objective function values. When the Complete poll option has the
(default) setting Off, the algorithm stops polling the mesh points as soon as it
finds a point whose objective function value is less than that of the current
point. If this occurs, the poll is called successful and the point it finds becomes
the current point at the next iteration.
The algorithm only computes the mesh points and their objective function
values up to the point at which it stops the poll. If the algorithm fails to find a
point that improves the objective function, the poll is called unsuccessful and
the current point stays the same at the next iteration.
When the Complete poll option has the setting On, the algorithm computes
the objective function values at all mesh points. The algorithm then compares
the mesh point with the smallest objective function value to the current
point. If that mesh point has a smaller value than the current point, the
poll is successful.

Expanding and Contracting

After polling, the algorithm changes the value of the mesh size m. The
default is to multiply m by 2 after a successful poll, and by 0.5 after an
unsuccessful poll.

4-13

Using Direct Search

How Pattern Search Works

In this section...
Context on page 4-14
Successful Polls on page 4-15
An Unsuccessful Poll on page 4-18
Displaying the Results at Each Iteration on page 4-19
More Iterations on page 4-20
Stopping Conditions for the Pattern Search on page 4-21

Context
patternsearch finds a sequence of points, x0, x1, x2, ... , that approach an

optimal point. The value of the objective function either decreases or remains
the same from each point in the sequence to the next. This section explains
how pattern search works for the function described in Example: Finding the
Minimum of a Function Using the GPS Algorithm on page 4-7.
To simplify the explanation, this section describes how the generalized
pattern search (GPS) works using a maximal positive basis of 2N, with Scale
set to Off in Mesh options.

4-14

How Pattern Search Works

The problem setup:

Successful Polls
The pattern search begins at the initial point x0 that you provide. In this
example, x0 = [2.1 1.7].

Iteration 1
At the first iteration, the mesh size is 1 and the GPS algorithm adds the
pattern vectors to the initial point x0 = [2.1 1.7] to compute the following
mesh points:
[1 0] + x0 = [3.1 1.7]
[0 1] + x0 = [2.1 2.7]
[-1 0] + x0 = [1.1 1.7]
[0 -1] + x0 = [2.1 0.7]

4-15

Using Direct Search

The algorithm computes the objective function at the mesh points in the order
shown above. The following figure shows the value of ps_example at the
initial point and mesh points.
Objective Function Values at Initial Point and Mesh Points
3
Initial point x0
Mesh points

5.6347
2.5

2
4.5146

4.6347

4.782

1.5

1
3.6347
0.5

1.5

2.5

3.5

First polled point that improves the objective function

The algorithm polls the mesh points by computing their objective function
values until it finds one whose value is smaller than 4.6347, the value at x0.
In this case, the first such point it finds is [1.1 1.7], at which the value of
the objective function is 4.5146, so the poll at iteration 1 is successful. The
algorithm sets the next point in the sequence equal to
x1 = [1.1 1.7]

4-16

How Pattern Search Works

Note By default, the GPS pattern search algorithm stops the current
iteration as soon as it finds a mesh point whose fitness value is smaller than
that of the current point. Consequently, the algorithm might not poll all the
mesh points. You can make the algorithm poll all the mesh points by setting
Complete poll to On.

Iteration 2
After a successful poll, the algorithm multiplies the current mesh size by 2,
the default value of Expansion factor in the Mesh options pane. Because
the initial mesh size is 1, at the second iteration the mesh size is 2. The mesh
at iteration 2 contains the following points:
2*[1 0] + x1 = [3.1 1.7]
2*[0 1] + x1 = [1.1 3.7]
2*[-1 0] + x1 = [-0.9 1.7]
2*[0 -1] + x1 = [1.1 -0.3]

The following figure shows the point x1 and the mesh points, together with
the corresponding values of ps_example.

4-17

Using Direct Search

Objective Function Values at x1 and Mesh Points

4
x1
Mesh points

6.5416
3.5
3
2.5
2
3.25

4.5146

4.7282

1.5
1
0.5
0
3.1146
0.5
1

0.5

0.5

1.5

2.5

3.5

The algorithm polls the mesh points until it finds one whose value is smaller
than 4.5146, the value at x1. The first such point it finds is [-0.9 1.7], at
which the value of the objective function is 3.25, so the poll at iteration 2 is
again successful. The algorithm sets the second point in the sequence equal to
x2 = [-0.9 1.7]

Because the poll is successful, the algorithm multiplies the current mesh size
by 2 to get a mesh size of 4 at the third iteration.

An Unsuccessful Poll
By the fourth iteration, the current point is
x3 = [-4.9 1.7]

and the mesh size is 8, so the mesh consists of the points

8*[1 0] + x3 = [3.1 1.7]
8*[0 1] + x3 = [-4.9 9.7]
8*[-1 0] + x3 = [-12.9 1.7]

4-18

How Pattern Search Works

8*[0 -1] + x3 = [-4.9 -1.3]

The following figure shows the mesh points and their objective function values.
Objective Function Values at x3 and Mesh Points
10

x3
Mesh points

7.7351

8
6
4
2

64.11

0.2649

4.7282

0
2
4
6
8

4.3351
10

At this iteration, none of the mesh points has a smaller objective function
value than the value at x3, so the poll is unsuccessful. In this case, the
algorithm does not change the current point at the next iteration. That is,
x4 = x3;

At the next iteration, the algorithm multiplies the current mesh size by
0.5, the default value of Contraction factor in the Mesh options pane, so
that the mesh size at the next iteration is 4. The algorithm then polls with
a smaller mesh size.

Displaying the Results at Each Iteration

You can display the results of the pattern search at each iteration by setting
Level of display to Iterative in the Display to command window
options. This enables you to evaluate the progress of the pattern search and
to make changes to options if necessary.

4-19

Using Direct Search

With this setting, the pattern search displays information about each iteration
at the command line. The first four iterations are
Iter
0
1
2
3
4

f-count
1
4
7
10
14

f(x)
4.63474
4.51464
3.25
-0.264905
-0.264905

MeshSize
1
2
4
8
4

Method
Successful Poll
Successful Poll
Successful Poll
Refine Mesh

The entry Successful Poll below Method indicates that the current iteration
was successful. For example, the poll at iteration 2 is successful. As a result,
the objective function value of the point computed at iteration 2, displayed
below f(x), is less than the value at iteration 1.
At iteration 4, the entry Refine Mesh tells you that the poll is unsuccessful.
As a result, the function value at iteration 4 remains unchanged from
iteration 3.
By default, the pattern search doubles the mesh size after each successful poll
and halves it after each unsuccessful poll.

More Iterations
The pattern search performs 60 iterations before stopping. The following
plot shows the points in the sequence computed in the first 13 iterations of
the pattern search.

4-20

How Pattern Search Works

Points at First 13 Iterations of Pattern Search

2

1.5

0.5

0.5

1
6

10

13

The numbers below the points indicate the first iteration at which
the algorithm finds the point. The plot only shows iteration numbers
corresponding to successful polls, because the best point doesnt change after
an unsuccessful poll. For example, the best point at iterations 4 and 5 is
the same as at iteration 3.

Stopping Conditions for the Pattern Search

The criteria for stopping the pattern search algorithm are listed in the
Stopping criteria section of the Optimization Tool:

4-21

Using Direct Search

The algorithm stops when any of the following conditions occurs:

The mesh size is less than Mesh tolerance.
The number of iterations performed by the algorithm reaches the value of
Max iteration.
The total number of objective function evaluations performed by the
algorithm reaches the value of Max function evaluations.
The time, in seconds, the algorithm runs until it reaches the value of
Time limit.
The distance between the point found in two consecutive iterations and the
mesh size are both less than X tolerance.

4-22

How Pattern Search Works

The change in the objective function in two consecutive iterations and the
mesh size are both less than Function tolerance.
Nonlinear constraint tolerance is not used as stopping criterion. It
determines the feasibility with respect to nonlinear constraints.
The MADS algorithm uses an additional parameter called the poll parameter,
p, in the mesh size stopping criterion:

N m
p =
m

for positive basis N + 1 poll

for positive basis 2N poll,

where m is the mesh size. The MADS stopping criterion is:

p Mesh tolerance.

4-23

Using Direct Search

Description of the Nonlinear Constraint Solver

The pattern search algorithm uses the Augmented Lagrangian Pattern Search
(ALPS) algorithm to solve nonlinear constraint problems. The optimization
problem solved by the ALPS algorithm is

min f ( x)
x

such that

ci ( x) 0, i = 1 m
ceqi ( x) = 0, i = m + 1 mt
Ax b
Aeq x = beq
lb x ub,
where c(x) represents the nonlinear inequality constraints, ceq(x) represents
the equality constraints, m is the number of nonlinear inequality constraints,
and mt is the total number of nonlinear constraints.
The ALPS algorithm attempts to solve a nonlinear optimization problem
with nonlinear constraints, linear constraints, and bounds. In this approach,
bounds and linear constraints are handled separately from nonlinear
constraints. A subproblem is formulated by combining the objective function
and nonlinear constraint function using the Lagrangian and the penalty
parameters. A sequence of such optimization problems are approximately
minimized using a pattern search algorithm such that the linear constraints
and bounds are satisfied.
A subproblem formulation is defined as
m

( x, , s, ) = f ( x) i si log(si ci ( x)) +
i=1

where

4-24

mt

i= m +1

i ci ( x) +

mt
ci ( x)2 ,
2 i= m+1

Description of the Nonlinear Constraint Solver

the components i of the vector are nonnegative and are known as

Lagrange multiplier estimates
the elements si of the vector s are nonnegative shifts
is the positive penalty parameter.
The algorithm begins by using an initial value for the penalty parameter
(InitialPenalty).
The pattern search algorithm minimizes a sequence of the subproblem,
which is an approximation of the original problem. When the subproblem
is minimized to a required accuracy and satisfies feasibility conditions,
the Lagrangian estimates are updated. Otherwise, the penalty parameter
is increased by a penalty factor (PenaltyFactor). This results in a new
subproblem formulation and minimization problem. These steps are repeated
until the stopping criteria are met.
For a complete description of the algorithm, see the following references:
[1] Kolda, Tamara G., Robert Michael Lewis, and Virginia Torczon.
A generating set direct search augmented Lagrangian algorithm for
optimization with a combination of general and linear constraints. Technical
Report SAND2006-5315, Sandia National Laboratories, August 2006.
[2] Conn, A. R., N. I. M. Gould, and Ph. L. Toint. A Globally Convergent
Augmented Lagrangian Algorithm for Optimization with General Constraints
and Simple Bounds, SIAM Journal on Numerical Analysis, Volume 28,
Number 2, pages 545572, 1991.
[3] Conn, A. R., N. I. M. Gould, and Ph. L. Toint. A Globally Convergent
Augmented Lagrangian Barrier Algorithm for Optimization with General
Inequality Constraints and Simple Bounds, Mathematics of Computation,
Volume 66, Number 217, pages 261288, 1997.

4-25

Using Direct Search

Performing a Pattern Search Using the Optimization Tool

GUI
In this section...
Example: A Linearly Constrained Problem on page 4-26
Displaying Plots on page 4-29
Example: Working with a Custom Plot Function on page 4-30

Example: A Linearly Constrained Problem

This section presents an example of performing a pattern search on a
constrained minimization problem. The example minimizes the function

F ( x) =

1 T
x Hx + f T x,
2

where

36
17

19
H=
12
8

15

17 19 12 8 15
33 18 11 7 14
18 43 13 8 16
,
11 13 18 6 11
7 8 6 9 8

14 16 11 8 29

f = [20 15 21 18 29 24 ],
subject to the constraints

A x b,
Aeq x = beq,
where

4-26

Performing a Pattern Search Using the Optimization Tool GUI

A = [ 8 7 3 4 9 0],
b = 7,
7 1 8 3 3 3
5 0 5 1 5 8
,
Aeq =
2 6 7 1 1 9

1 1 2 2 3 3
beq = [84 62 65 1].

Performing a Pattern Search on the Example

To perform a pattern search on the example, first enter
optimtool('patternsearch')

to open the Optimization Tool, or enter optimtool and then choose

patternsearch from the Solver menu. Then type the following function in
the Objective function field:
@lincontest7
lincontest7 is a file included in Global Optimization Toolbox software that

computes the objective function for the example. Because the matrices and
vectors defining the starting point and constraints are large, it is more
convenient to set their values as variables in the MATLAB workspace first
and then enter the variable names in the Optimization Tool. To do so, enter
x0 = [2 1 0 9 1 0];
Aineq = [-8 7 3 -4 9 0];
bineq = 7;
Aeq = [7 1 8 3 3 3; 5 0 -5 1 -5 8; -2 -6 7 1 1 9; 1 -1 2 -2 3 -3];
beq = [84 62 65 1];

Then, enter the following in the Optimization Tool:

Set Start point to x0.
Set the following Linear inequalities:

Set A to Aineq.
Set b to bineq.

4-27

Using Direct Search

Set Aeq to Aeq.

Set beq to beq.

The following figure shows these settings in the Optimization Tool.

Since this is a linearly constrained problem, set the Poll method to GSS
Positive basis 2N. For more information about the efficiency of the GSS
search methods for linearly constrained problems, see Example: Comparing
the Efficiency of Poll Options on page 4-48.

Then click Start to run the pattern search. When the search is finished, the
results are displayed in Run solver and view results pane, as shown in
the following figure.

4-28

Performing a Pattern Search Using the Optimization Tool GUI

Displaying Plots
The Plot functions pane, shown in the following figure, enables you to
display various plots of the results of a pattern search.

Select the check boxes next to the plots you want to display. For example,
if you select Best function value and Mesh size, and run the example
described in Example: Finding the Minimum of a Function Using the GPS
Algorithm on page 4-7, the tool displays the plots shown in the following
figure.

4-29

Using Direct Search

The upper plot displays the objective function value at each iteration. The
lower plot displays the mesh size at each iteration.
Note When you display more than one plot, you can open a larger version
of a plot in a separate window. Right-click (Ctrl-click for Mac) on a blank
area in a plot while patternsearch is running, or after it has stopped, and
choose the sole menu item.
Plot Options on page 9-30 describes the types of plots you can create.

Example: Working with a Custom Plot Function

To use a plot function other than those included with the software, you can
write your own custom plot function that is called at each iteration of the
pattern search to create the plot. This example shows how to create a plot

4-30

Performing a Pattern Search Using the Optimization Tool GUI

function that displays the logarithmic change in the best objective function
value from the previous iteration to the current iteration.
This section covers the following topics:
Creating the Custom Plot Function on page 4-31
Setting Up the Problem on page 4-32
Using the Custom Plot Function on page 4-33
How the Plot Function Works on page 4-34

Creating the Custom Plot Function

To create the plot function for this example, copy and paste the following code
into a new function file in the MATLAB Editor:
function stop = psplotchange(optimvalues, flag)
% PSPLOTCHANGE Plots the change in the best objective function
% value from the previous iteration.
% Best objective function value in the previous iteration
persistent last_best
stop = false;
if(strcmp(flag,'init'))
set(gca,'Yscale','log'); %Set up the plot
hold on;
xlabel('Iteration');
ylabel('Log Change in Values');
title(['Change in Best Function Value']);
end
% Best objective function value in the current iteration
best = min(optimvalues.fval);
% Set last_best to best
if optimvalues.iteration == 0
last_best = best;
else

4-31

Using Direct Search

%Change in objective function value

change = last_best - best;
plot(optimvalues.iteration, change, '.r');
end

Then save the file as psplotchange.m in a folder on the MATLAB path.

Setting Up the Problem

The problem is the same as Example: A Linearly Constrained Problem
on page 4-26. To set up the problem:
1 Enter the following at the MATLAB command line:

x0 = [2 1 0 9 1 0];
Aineq = [-8 7 3 -4 9 0];
bineq = 7;
Aeq = [7 1 8 3 3 3; 5 0 -5 1 -5 8; -2 -6 7 1 1 9; 1 -1 2 -2 3 -3];
beq = [84 62 65 1];
2 Enter optimtool to open the Optimization Tool.
3 Choose the patternsearch solver.
4 Set up the problem to match the following figure.

4-32

Performing a Pattern Search Using the Optimization Tool GUI

5 Since this is a linearly constrained problem, set the Poll method to GSS

Positive basis 2N.

Using the Custom Plot Function

To use the custom plot function, select Custom function in the Plot
functions pane and enter @psplotchange in the field to the right. To
compare the custom plot with the best function value plot, also select Best
function value.

Now, when you run the example, the pattern search tool displays the plots
shown in the following figure.

4-33

Using Direct Search

Note that because the scale of the y-axis in the lower custom plot is
logarithmic, the plot will only show changes that are greater than 0.

How the Plot Function Works

The plot function uses information contained in the following structures,
which the Optimization Tool passes to the function as input arguments:
optimvalues Structure containing the current state of the solver
flag String indicating the current status of the algorithm
The most important statements of the custom plot function, psplotchange.m,
are summarized in the following table.

4-34

Performing a Pattern Search Using the Optimization Tool GUI

Custom Plot Function Statements

Statement

Description

persistent last_best

Creates the persistent variable

last_best, the best objective
function value in the previous
generation. Persistent variables are
preserved over multiple calls to the
plot function.

set(gca,'Yscale','log')

Sets up the plot before the algorithm

starts.

best = min(optimvalues.fval)

Sets best equal to the minimum

objective function value. The field
optimvalues.fval contains the
objective function value in the
current iteration. The variable best
is the minimum objective function
value. For a complete description
of the fields of the structure
optimvalues, see Structure of the
Plot Functions on page 9-9.

change = last_best - best

Sets the variable change to the

best objective function value at
the previous iteration minus the
best objective function value in the
current iteration.

plot(optimvalues.iteration,
change, '.r')

Plots the variable change at the

current objective function value,
for the current iteration contained
inoptimvalues.iteration.

4-35

Using Direct Search

Performing a Pattern Search from the Command Line

In this section...
Calling patternsearch with the Default Options on page 4-36
Setting Options for patternsearch at the Command Line on page 4-38
Using Options and Problems from the Optimization Tool on page 4-40

Calling patternsearch with the Default Options

This section describes how to perform a pattern search with the default
options.

Pattern Search on Unconstrained Problems

For an unconstrained problem, call patternsearch with the syntax
[x fval] = patternsearch(@objectfun, x0)

The output arguments are

x The final point
fval The value of the objective function at x
The required input arguments are
@objectfun A function handle to the objective function objectfun,
which you can write as a function file. See Computing Objective Functions
on page 2-2 to learn how to do this.
x0 The initial point for the pattern search algorithm.
As an example, you can run the example described in Example: Finding
the Minimum of a Function Using the GPS Algorithm on page 4-7 from the
command line by entering
[x fval] = patternsearch(@ps_example, [2.1 1.7])

This returns

4-36

Performing a Pattern Search from the Command Line

Optimization terminated: mesh size less than options.TolMesh.

x =
-4.7124

-0.0000

fval =
-2.0000

Pattern Search on Constrained Problems

If your problem has constraints, use the syntax
[x fval] = patternsearch(@objfun,x0,A,b,Aeq,beq,lb,ub,nonlcon)

where
A is a matrix and b is vector that represent inequality constraints of the
form Ax b.
Aeq is a matrix and beq is a vector that represent equality constraints of
the form Aeqx = beq.
lb and ub are vectors representing bound constraints of the form lb x and
x ub, respectively.
nonlcon is a function that returns the nonlinear equality and inequality
vectors, c and ceq, respectively. The function is minimized such that
c(x) 0 and ceq(x) = 0.
You only need to pass in the constraints that are part of the problem. For
example, if there are no bound constraints or a nonlinear constraint function,
use the syntax
[x fval] = patternsearch(@objfun,x0,A,b,Aeq,beq)

Use empty brackets [] for constraint arguments that are not needed for the
problem. For example, if there are no inequality constraints or a nonlinear
constraint function, use the syntax
[x fval] = patternsearch(@objfun,x0,[],[],Aeq,beq,lb,ub)

4-37

Using Direct Search

Additional Output Arguments

To get more information about the performance of the pattern search, you can
call patternsearch with the syntax
[x fval exitflag output] = patternsearch(@objfun,x0)

Besides x and fval, this returns the following additional output arguments:
exitflag Integer indicating whether the algorithm was successful
output Structure containing information about the performance of the
solver
See the reference page for patternsearch for more information about these
arguments.

Setting Options for patternsearch at the Command

Line
You can specify any available patternsearch options by passing an options
structure as an input argument to patternsearch using the syntax
[x fval] = patternsearch(@fitnessfun,nvars, ...
A,b,Aeq,beq,lb,ub,nonlcon,options)

Pass in empty brackets [] for any constraints that do not appear in the
problem.
You create the options structure using the function psoptimset.
options = psoptimset(@patternsearch)

This returns the options structure with the default values for its fields.
options =
TolMesh:
TolCon:
TolX:
TolFun:
TolBind:
MaxIter:

4-38

1.0000e-006
1.0000e-006
1.0000e-006
1.0000e-006
1.0000e-003
'100*numberofvariables'

Performing a Pattern Search from the Command Line

MaxFunEvals:
TimeLimit:
MeshContraction:
MeshExpansion:
MeshAccelerator:
MeshRotate:
InitialMeshSize:
ScaleMesh:
MaxMeshSize:
InitialPenalty:
PenaltyFactor:
PollMethod:
CompletePoll:
PollingOrder:
SearchMethod:
CompleteSearch:
Display:
OutputFcns:
PlotFcns:
PlotInterval:
Cache:
CacheSize:
CacheTol:
Vectorized:
UseParallel:

'2000*numberofvariables'
Inf
0.5000
2
'off'
'on'
1
'on'
Inf
10
100
'gpspositivebasis2n'
'off'
'consecutive'
[]
'off'
'final'
[]
[]
1
'off'
10000
2.2204e-016
'off'
'never'

The patternsearch function uses these default values if you do not pass
in options as an input argument.
The value of each option is stored in a field of the options structure, such as
options.MeshExpansion. You can display any of these values by entering
options followed by the name of the field. For example, to display the mesh
expansion factor for the pattern search, enter
options.MeshExpansion
ans =
2

4-39

Using Direct Search

To create an options structure with a field value that is different from the
default, use psoptimset. For example, to change the mesh expansion factor
to 3 instead of its default value 2, enter
options = psoptimset('MeshExpansion',3);

This creates the options structure with all values set to empty except for
MeshExpansion, which is set to 3. (An empty field causes patternsearch to
use the default value.)
If you now call patternsearch with the argument options, the pattern
search uses a mesh expansion factor of 3.
If you subsequently decide to change another field in the options structure,
such as setting PlotFcns to @psplotmeshsize, which plots the mesh size at
each iteration, call psoptimset with the syntax
options = psoptimset(options, 'PlotFcns', @psplotmeshsize)

This preserves the current values of all fields of options except for PlotFcns,
which is changed to @plotmeshsize. Note that if you omit the options input
argument, psoptimset resets MeshExpansion to its default value, which is 2.
You can also set both MeshExpansion and PlotFcns with the single command
options = psoptimset('MeshExpansion',3,'PlotFcns',@plotmeshsize)

Using Options and Problems from the Optimization

Tool
As an alternative to creating the options structure using psoptimset, you can
set the values of options in the Optimization Tool and then export the options
to a structure in the MATLAB workspace, as described in the Importing and
Exporting Your Work section of the Optimization Toolbox documentation.
If you export the default options in the Optimization Tool, the resulting
options structure has the same settings as the default structure returned
by the command
options = psoptimset

except for the default value of 'Display', which is 'final' when created by
psoptimset, but 'none' when created in the Optimization Tool.

4-40

Performing a Pattern Search from the Command Line

You can also export an entire problem from the Optimization Tool and run it
from the command line. Enter
patternsearch(problem)

where problem is the name of the exported problem.

4-41

Using Direct Search

Pattern Search Examples: Setting Options

In this section...
Poll Method on page 4-42
Complete Poll on page 4-44
Example: Comparing the Efficiency of Poll Options on page 4-48
Using a Search Method on page 4-55
Mesh Expansion and Contraction on page 4-59
Mesh Accelerator on page 4-64
Using Cache on page 4-65
Setting Tolerances for the Solver on page 4-68
Constrained Minimization Using patternsearch on page 4-69
Vectorizing the Objective and Constraint Functions on page 4-72

Note All examples use the generalized pattern search (GPS) algorithm, but
can be applied to the other algorithms as well.

Poll Method
At each iteration, the pattern search polls the points in the current
meshthat is, it computes the objective function at the mesh points to see
if there is one whose function value is less than the function value at the
current point. How Pattern Search Works on page 4-14 provides an example
of polling. You can specify the pattern that defines the mesh by the Poll
method option. The default pattern, GPS Positive basis 2N, consists of the
following 2N directions, where N is the number of independent variables for
the objective function.
[1 0 0...0]
[0 1 0...0]
...
[0 0 0...1]
[1 0 0...0]

4-42

Pattern Search Examples: Setting Options

[0 1 0...0]
[0 0 0...1].
For example, if the objective function has three independent variables, the
GPS Positive basis 2N, consists of the following six vectors.
[1 0 0]
[0 1 0]
[0 0 1]
[1 0 0]
[0 1 0]
[0 0 1].
Alternatively, you can set Poll method to GPS Positive basis NP1, the
pattern consisting of the following N + 1 directions.
[1 0 0...0]
[0 1 0...0]
...
[0 0 0...1]
[1 1 1...1].
For example, if objective function has three independent variables, the GPS
Positive basis Np1, consists of the following four vectors.

[1 0 0]
[0 1 0]
[0 0 1]
[1 1 1].
A pattern search will sometimes run faster using GPS Positive basis Np1
rather than the GPS Positive basis 2N as the Poll method, because
the algorithm searches fewer points at each iteration. Although not being
addressed in this example, the same is true when using the MADS Positive
basis Np1 over the MADS Positive basis 2N, and similarly for GSS. For
example, if you run a pattern search on the example described in Example: A
Linearly Constrained Problem on page 4-26, the algorithm performs 1588
function evaluations with GPS Positive basis 2N, the default Poll method,
but only 877 function evaluations using GPS Positive basis Np1. For more
detail, see Example: Comparing the Efficiency of Poll Options on page 4-48.

4-43

Using Direct Search

However, if the objective function has many local minima, using GPS
Positive basis 2N as the Poll method might avoid finding a local
minimum that is not the global minimum, because the search explores more
points around the current point at each iteration.

Complete Poll
By default, if the pattern search finds a mesh point that improves the value
of the objective function, it stops the poll and sets that point as the current
point for the next iteration. When this occurs, some mesh points might not
get polled. Some of these unpolled points might have an objective function
value that is even lower than the first one the pattern search finds.
For problems in which there are several local minima, it is sometimes
preferable to make the pattern search poll all the mesh points at each
iteration and choose the one with the best objective function value. This
enables the pattern search to explore more points at each iteration and
thereby potentially avoid a local minimum that is not the global minimum.
In the Optimization Tool you can make the pattern search poll the entire
mesh setting Complete poll to On in Poll options. At the command line, use
psoptimset to set the CompletePoll option to 'on'.

Example: Using a Complete Poll in a Generalized Pattern

Search
As an example, consider the following function.

x12 + x22 25
for x12 + x22 25

2
2
f ( x1 , x2 ) = x12 + ( x2 9 ) 16 for x12 + ( x2 9 ) 16
0
otherwise.

The following figure shows a plot of the function.

4-44

Pattern Search Examples: Setting Options

0
5
10
15
10

20
5
25
15

0
10

5
5

10

10

Global minimum at (0,0)

Local minimum at (0,9)

The global minimum of the function occurs at (0, 0), where its value is -25.
However, the function also has a local minimum at (0, 9), where its value is
-16.
To create a file that computes the function, copy and paste the following code
into a new file in the MATLAB Editor.
function z = poll_example(x)
if x(1)^2 + x(2)^2 <= 25
z = x(1)^2 + x(2)^2 - 25;
elseif x(1)^2 + (x(2) - 9)^2 <= 16
z = x(1)^2 + (x(2) - 9)^2 - 16;
else z = 0;
end

Then save the file as poll_example.m in a folder on the MATLAB path.

4-45

Using Direct Search

To run a pattern search on the function, enter the following in the

Optimization Tool:
Set Solver to patternsearch.
Set Objective function to @poll_example.
Set Start point to [0 5].
Set Level of display to Iterative in the Display to command window
options.
Click Start to run the pattern search with Complete poll set to Off, its
default value. The Optimization Tool displays the results in the Run solver
and view results pane, as shown in the following figure.

The pattern search returns the local minimum at (0, 9). At the initial point,
(0, 5), the objective function value is 0. At the first iteration, the search polls
the following mesh points.
f((0, 5) + (1, 0)) = f(1, 5) = 0
f((0, 5) + (0, 1)) = f(0, 6) = -7
As soon as the search polls the mesh point (0, 6), at which the objective
function value is less than at the initial point, it stops polling the current
mesh and sets the current point at the next iteration to (0, 6). Consequently,

4-46

Pattern Search Examples: Setting Options

the search moves toward the local minimum at (0, 9) at the first iteration. You
see this by looking at the first two lines of the command line display.
Iter
0
1

f-count
1
3

f(x)
0
-7

MeshSize
1
2

Method
Successful Poll

Note that the pattern search performs only two evaluations of the objective
function at the first iteration, increasing the total function count from 1 to 3.
Next, set Complete poll to On and click Start. The Run solver and view
results pane displays the following results.

This time, the pattern search finds the global minimum at (0, 0). The
difference between this run and the previous one is that with Complete poll
set to On, at the first iteration the pattern search polls all four mesh points.
f((0, 5) + (1, 0)) = f(1, 5) = 0
f((0, 5) + (0, 1)) = f(0, 6) = -6
f((0, 5) + (-1, 0)) = f(-1, 5) = 0
f((0, 5) + (0, -1)) = f(0, 4) = -9

4-47

Using Direct Search

Because the last mesh point has the lowest objective function value, the
pattern search selects it as the current point at the next iteration. The first
two lines of the command-line display show this.
Iter
0
1

f-count
1
5

f(x)
0
-9

MeshSize
1
2

Method
Successful Poll

In this case, the objective function is evaluated four times at the first iteration.
As a result, the pattern search moves toward the global minimum at (0, 0).
The following figure compares the sequence of points returned when
Complete poll is set to Off with the sequence when Complete poll is On.

14
Initial point
Complete poll off
Complete poll on

12
10
8

Local minimum

6
4

Global minimum

2
0
2
4
6
6

10

12

14

Example: Comparing the Efficiency of Poll Options

This example shows how several poll options interact in terms of iterations
and total function evaluations. The main results are:

4-48

Pattern Search Examples: Setting Options

GSS is more efficient than GPS or MADS for linearly constrained problems.
Whether setting CompletePoll to 'on' increases efficiency or decreases
efficiency is unclear, although it affects the number of iterations.
Similarly, whether having a 2N poll is more or less efficient than having
an Np1 poll is also unclear. The most efficient poll is GSS Positive Basis
Np1 with Complete poll set to on. The least efficient is MADS Positive
Basis Np1 with Complete poll set to on.
Note The efficiency of an algorithm depends on the problem. GSS is
efficient for linearly constrained problems. However, predicting the efficiency
implications of the other poll options is difficult, as is knowing which poll type
works best with other constraints.

Problem setup
The problem is the same as in Performing a Pattern Search on the Example
on page 4-27. This linearly constrained problem uses the lincontest7 file
that comes with the toolbox:
1 Enter the following into your MATLAB workspace:

x0 = [2 1 0 9 1 0];
Aineq = [-8 7 3 -4 9 0];
bineq = 7;
Aeq = [7 1 8 3 3 3; 5 0 -5 1 -5 8; -2 -6 7 1 1 9; 1 -1 2 -2 3 -3];
beq = [84 62 65 1];
2 Open the Optimization Tool by entering optimtool at the command line.
3 Choose the patternsearch solver.
4 Enter the problem and constraints as pictured.

4-49

Using Direct Search

5 Ensure that the Poll method is GPS Positive basis 2N.

Generate the Results

1 Run the optimization.

2 Choose File > Export to Workspace.

4-50

Pattern Search Examples: Setting Options

3 Export the results to a structure named gps2noff.

4 Set Options > Poll > Complete poll to on.

4-51

Using Direct Search

5 Run the optimization.

6 Export the result to a structure named gps2non.
7 Set Options > Poll > Poll method to GPS Positive basis Np1 and set

Complete poll to off.

8 Run the optimization.
9 Export the result to a structure named gpsnp1off.
10 Set Complete poll to on.
11 Run the optimization.
12 Export the result to a structure named gpsnp1on.
13 Continue in a like manner to create solution structures for the other

poll methods with Complete poll set on and off: gss2noff, gss2non,
gssnp1off, gssnp1on, mads2noff, mads2non, madsnp1off, and madsnp1on.

Examine the Results

You have the results of 12 optimization runs. The following table shows the
efficiency of the runs, measured in total function counts and in iterations.
Your MADS results could differ, since MADS is a stochastic algorithm.

4-52

Pattern Search Examples: Setting Options

Algorithm

Function Count

Iterations

GPS2N, complete poll off

1206

106

GPS2N, complete poll on

1362

94

GPSNp1, complete poll off

914

122

GPSNp1, complete poll on

991

106

GSS2N, complete poll off

683

74

GSS2N, complete poll on

889

74

GSSNp1, complete poll off

529

94

GSSNp1, complete poll on

519

74

MADS2N, complete poll off

907

159

MADS2N, complete poll on

1292

160

MADSNp1, complete poll off

1215

208

MADSNp1, complete poll on

1714

207

To obtain, say, the first row in the table, enter gps2noff.output.funccount

and gps2noff.output.iterations. You can also examine a structure in the
Variable Editor by double-clicking the structure in the Workspace Browser,
and then double-clicking the output structure.

4-53

Using Direct Search

The main results gleaned from the table are:

Setting Complete poll to on generally lowers the number of iterations
for GPS and GSS, but the change in number of function evaluations is
unpredictable.
Setting Complete poll to on does not change the number of iterations for
MADS, but substantially increases the number of function evaluations.
The most efficient algorithm/options settings, with efficiency meaning
lowest function count:

4-54

Pattern Search Examples: Setting Options

1 GSS Positive basis Np1 with Complete poll set to on (function count
519)
2 GSS Positive basis Np1 with Complete poll set to off (function
count 529)
3 GSS Positive basis 2N with Complete poll set to off (function count
683)
4 GSS Positive basis 2N with Complete poll set to on (function count
889)
The other poll methods had function counts exceeding 900.
For this problem, the most efficient poll is GSS Positive Basis Np1 with
Complete poll set to on, although the Complete poll setting makes only
a small difference. The least efficient poll is MADS Positive Basis Np1
with Complete poll set to on. In this case, the Complete poll setting
makes a substantial difference.

Using a Search Method

In addition to polling the mesh points, the pattern search algorithm can
perform an optional step at every iteration, called search. At each iteration,
the search step applies another optimization method to the current point. If
this search does not improve the current point, the poll step is performed.
The following example illustrates the use of a search method on the problem
described in Example: A Linearly Constrained Problem on page 4-26. To
set up the example, enter the following commands at the MATLAB prompt to
define the initial point and constraints.
x0 = [2 1 0 9 1 0];
Aineq = [-8 7 3 -4 9 0];
bineq = 7;
Aeq = [7 1 8 3 3 3; 5 0 -5 1 -5 8; -2 -6 7 1 1 9; 1 -1 2 -2 3 -3];
beq = [84 62 65 1];

Then enter the settings shown in the following figures in the Optimization
Tool.

4-55

Using Direct Search

For comparison, click Start to run the example without a search method.
This displays the plots shown in the following figure.

4-56

Pattern Search Examples: Setting Options

To see the effect of using a search method, select MADS Positive Basis 2N
in the Search method field in Search options.

4-57

Using Direct Search

This sets the search method to be a pattern search using the pattern for
MADS Positive Basis 2N. Then click Start to run the pattern search. This
displays the following plots.

Note that using the search method reduces the total function
evaluationsfrom 1206 to 1159and reduces the number of iterations from
106 to 97.

4-58

Pattern Search Examples: Setting Options

Mesh Expansion and Contraction

The Expansion factor and Contraction factor options, in Mesh options,
control how much the mesh size is expanded or contracted at each iteration.
With the default Expansion factor value of 2, the pattern search multiplies
the mesh size by 2 after each successful poll. With the default Contraction
factor value of 0.5, the pattern search multiplies the mesh size by 0.5 after
each unsuccessful poll.
You can view the expansion and contraction of the mesh size during the
pattern search by selecting Mesh size in the Plot functions pane.

To also display the values of the mesh size and objective function at the
command line, set Level of display to Iterative in the Display to
command window options.

When you run the example described in Example: A Linearly Constrained

Problem on page 4-26, the Optimization Tool displays the following plot.

4-59

Using Direct Search

To see the changes in mesh size more clearly, change the y-axis to logarithmic
scaling as follows:
1 Select Axes Properties from the Edit menu in the plot window.
2 In the Properties Editor, select the Y Axis tab.
3 Set Scale to Log.

Updating these settings in the MATLAB Property Editor will show the plot in
the following figure.

4-60

Pattern Search Examples: Setting Options

First unsuccessful poll

The first 5 iterations result in successful polls, so the mesh sizes increase
steadily during this time. You can see that the first unsuccessful poll occurs
at iteration 6 by looking at the command-line display:
Iter
0
1
2
3
4
5
6

f-count
1
2
3
4
5
8
17

f(x)
2434.15
2419.37
2392.12
2346.85
2293.23
1921.94
1921.94

MeshSize
1
2
4
8
16
32
16

Method
Successful Poll
Successful Poll
Successful Poll
Successful Poll
Successful Poll
Refine Mesh

Note that at iteration 5, which is successful, the mesh size doubles for the
next iteration. But at iteration 6, which is unsuccessful, the mesh size is
multiplied 0.5.
To see how Expansion factor and Contraction factor affect the pattern
search, make the following changes:

4-61

Using Direct Search

Set Expansion factor to 3.0.

Set Contraction factor to 0.75.

Then click Start. The Run solver and view results pane shows that
the final point is approximately the same as with the default settings of
Expansion factor and Contraction factor, but that the pattern search
takes longer to reach that point.

4-62

Pattern Search Examples: Setting Options

When you change the scaling of the y-axis to logarithmic, the mesh size plot
appears as shown in the following figure.

4-63

Using Direct Search

Note that the mesh size increases faster with Expansion factor set to 3.0,
as compared with the default value of 2.0, and decreases more slowly with
Contraction factor set to 0.75, as compared with the default value of 0.5.

Mesh Accelerator
The mesh accelerator can make a pattern search converge faster to an
optimal point by reducing the number of iterations required to reach the
mesh tolerance. When the mesh size is below a certain value, the pattern
search contracts the mesh size by a factor smaller than the Contraction
factor factor.
Note It is recommended to only use the mesh accelerator for problems in
which the objective function is not too steep near the optimal point, or you
might lose some accuracy. For differentiable problems, this means that the
absolute value of the derivative is not too large near the solution.

4-64

Pattern Search Examples: Setting Options

To use the mesh accelerator, set Accelerator to On in the Mesh options.

When you run the example described in Example: A Linearly Constrained
Problem on page 4-26, the number of iterations required to reach the mesh
tolerance is 61, as compared with 74 when Accelerator is set to Off.
You can see the effect of the mesh accelerator by setting Level of display
to Iterative in Display to command window. Run the example with
Accelerator set to On, and then run it again with Accelerator set to Off.
The mesh sizes are the same until iteration 56, but differ at iteration 57. The
MATLAB Command Window displays the following lines for iterations 56
and 57 with Accelerator set to Off.
Iter
56
57

f-count
509
521

f(x)
1919.54
1919.54

MeshSize
6.104e-005
3.052e-005

Method
Refine Mesh
Refine Mesh

Note that the mesh size is multiplied by 0.5, the default value of Contraction
factor.
For comparison, the Command Window displays the following lines for the
same iteration numbers with Accelerator set to On.
Iter
56
57

f-count
509
521

f(x)
1919.54
1919.54

MeshSize
6.104e-005
1.526e-005

Method
Refine Mesh
Refine Mesh

In this case the mesh size is multiplied by 0.25.

Using Cache
Typically, at any given iteration of a pattern search, some of the mesh points
might coincide with mesh points at previous iterations. By default, the
pattern search recomputes the objective function at these mesh points even
though it has already computed their values and found that they are not
optimal. If computing the objective function takes a long timesay, several
minutesthis can make the pattern search run significantly longer.
You can eliminate these redundant computations by using a cache, that is,
by storing a history of the points that the pattern search has already visited.
To do so, set Cache to On in Cache options. At each poll, the pattern search
checks to see whether the current mesh point is within a specified tolerance,

4-65

Using Direct Search

Tolerance, of a point in the cache. If so, the search does not compute the
objective function for that point, but uses the cached function value and
moves on to the next point.
Note When Cache is set to On, the pattern search might fail to identify a
point in the current mesh that improves the objective function because it is
within the specified tolerance of a point in the cache. As a result, the pattern
search might run for more iterations with Cache set to On than with Cache
set to Off. It is generally a good idea to keep the value of Tolerance very
small, especially for highly nonlinear objective functions.
To illustrate this, select Best function value and Function count in the
Plot functions pane and run the example described in Example: A Linearly
Constrained Problem on page 4-26 with Cache set to Off. After the pattern
search finishes, the plots appear as shown in the following figure.

4-66

Pattern Search Examples: Setting Options

Note that the total function count is 683.

Now, set Cache to On and run the example again. This time, the plots appear
as shown in the following figure.

4-67

Using Direct Search

This time, the total function count is reduced to 664.

Setting Tolerances for the Solver

Tolerance refers to how small a parameter, such a mesh size, can become
before the search is halted or changed in some way. You can specify the value
of the following tolerances:
Mesh tolerance When the current mesh size is less than the value of
Mesh tolerance, the algorithm halts.

4-68

Pattern Search Examples: Setting Options

X tolerance After a successful poll, if the distance from the previous

best point to the current best point is less than the value of X tolerance,
the algorithm halts.
Function tolerance After a successful poll, if the difference between
the function value at the previous best point and function value at the
current best point is less than the value of Function tolerance, the
algorithm halts.
Nonlinear constraint tolerance The algorithm treats a point to be
feasible if constraint violation is less than TolCon.
Bind tolerance Bind tolerance applies to linearly constrained problems.
It specifies how close a point must get to the boundary of the feasible
region before a linear constraint is considered to be active. When a linear
constraint is active, the pattern search polls points in directions parallel to
the linear constraint boundary as well as the mesh points.
Usually, you should set Bind tolerance to be at least as large as the
maximum of Mesh tolerance, X tolerance, and Function tolerance.

Constrained Minimization Using patternsearch

Suppose you want to minimize the simple objective function of two variables
x1 and x2,

min f ( x) = 4 - 2.1 x12 x14 / 3 x12 + x1 x2 + 4 + 4 x22 x22

x

subject to the following nonlinear inequality constraints and bounds

x1 x2 + x1 x2 + 1.5 0 (nonlinear constraint)

10 x1 x2 0
0 x1 1
0 x2 13

(nonlinear constraint)
(bound)
(bound)

Begin by creating the objective and constraint functions. First, create a file
named simple_objective.m as follows:
function y = simple_objective(x)
y = (4 - 2.1*x(1)^2 + x(1)^4/3)*x(1)^2 + x(1)*x(2) + (-4 + 4*x(2)^2)*x(2)^2;

4-69

Using Direct Search

The pattern search solver assumes the objective function will take one input x
where x has as many elements as number of variables in the problem. The
objective function computes the value of the function and returns that scalar
value in its one return argument y.
Then create a file named simple_constraint.m containing the constraints:
function [c, ceq] = simple_constraint(x)
c = [1.5 + x(1)*x(2) + x(1) - x(2);
-x(1)*x(2) + 10];
ceq = [];

The pattern search solver assumes the constraint function will take one input
x, where x has as many elements as the number of variables in the problem.
The constraint function computes the values of all the inequality and equality
constraints and returns two vectors, c and ceq, respectively.
Next, to minimize the objective function using the patternsearch function,
you need to pass in a function handle to the objective function as well as
specifying a start point as the second argument. Lower and upper bounds
are provided as LB and UB respectively. In addition, you also need to pass a
function handle to the nonlinear constraint function.
ObjectiveFunction = @simple_objective;
X0 = [0 0];

% Starting point

LB = [0 0];

% Lower bound

UB = [1 13];

% Upper bound

ConstraintFunction = @simple_constraint;
[x,fval] = patternsearch(ObjectiveFunction,X0,[],[],[],[],...
LB,UB,ConstraintFunction)
Optimization terminated: mesh size less than options.TolMesh
and constraint violation is less than options.TolCon.
x =
0.8122
fval =
9.1324e+004

4-70

12.3122

Pattern Search Examples: Setting Options

Next, plot the results. Create an options structure using psoptimset that
selects two plot functions. The first plot function psplotbestf plots the
best objective function value at every iteration. The second plot function
psplotmaxconstr plots the maximum constraint violation at every iteration.
Note You can also visualize the progress of the algorithm by displaying
information to the Command Window using the 'Display' option.

options = psoptimset('PlotFcns',{@psplotbestf,@psplotmaxconstr},'Display','iter');
[x,fval] = patternsearch(ObjectiveFunction,X0,[],[],[],[],LB,UB,ConstraintFunction,options)

max
Iter

f-count

f(x)

constraint
10

MeshSize

Method

0.8919

113580

0.001

Increase penalty

24

91324.4

1e-005

Increase penalty

48

91324

1e-007

Increase penalty

Optimization terminated: mesh size less than options.TolMesh

and constraint violation is less than options.TolCon.

x =
0.8122

12.3122

fval =
9.1324e+004

4-71

Using Direct Search

Best Objective Function Value and Maximum Constraint Violation at Each

Iteration

Vectorizing the Objective and Constraint Functions

Direct search often runs faster if you vectorize the objective and nonlinear
constraint functions. This means your functions evaluate all the points in a
poll or search pattern at once, with one function call, without having to loop
through the points one at a time. Therefore, the option Vectorize = 'on'
works only when CompletePoll or CompleteSearch are also set to 'on'.
If there are nonlinear constraints, the objective function and the nonlinear
constraints all need to be vectorized in order for the algorithm to compute in
a vectorized manner.

4-72

Pattern Search Examples: Setting Options

Vectorized Objective Function

A vectorized objective function accepts a matrix as input and generates a
vector of function values, where each function value corresponds to one row
or column of the input matrix. patternsearch resolves the ambiguity in
whether the rows or columns of the matrix represent the points of a pattern
as follows. Suppose the input matrix has m rows and n columns:
If the initial point x0 is a column vector of size m, the objective function
takes each column of the matrix as a point in the pattern and returns a
vector of size n.
If the initial point x0 is a row vector of size n, the objective function takes
each row of the matrix as a point in the pattern and returns a vector of
size m.
If the initial point x0 is a scalar, the matrix has one row (m = 1, the matrix
is a vector), and each entry of the matrix represents one point to evaluate.
Pictorially, the matrix and calculation are represented by the following figure.

4-73

Using Direct Search

Results

The matrix passed to

objective and constraint functions

If x0 is a
column

...
Each column represents
one search/poll point

Results

If x0 is a row

...

Each row represents

one search/poll point

The matrix passed to

objective and constraint functions
Structure of Vectorized Functions

For example, suppose the objective function is

f ( x) = x14 + x24 4 x12 2 x22 + 3 x1 x2 / 2.

If the initial vector x0 is a column vector, such as [0;0], a function for
vectorized evaluation is
function f = vectorizedc(x)
f = x(1,:).^4+x(2,:).^4-4*x(1,:).^2-2*x(2,:).^2 ...
+3*x(1,:)-.5*x(2,:);

4-74

Pattern Search Examples: Setting Options

If the initial vector x0 is a row vector, such as [0,0], a function for vectorized
evaluation is
function f = vectorizedr(x)
f = x(:,1).^4+x(:,2).^4-4*x(:,1).^2-2*x(:,2).^2 ...
+3*x(:,1)-.5*x(:,2);

If you want to use the same objective (fitness) function for both pattern search
and genetic algorithm, write your function to have the points represented
by row vectors, and write x0 as a row vector. The genetic algorithm always
takes individuals as the rows of a matrix. This was a design decisionthe
genetic algorithm does not require a user-supplied population, so needs to
have a default format.
To minimize vectorizedc, enter the following commands:
options=psoptimset('Vectorized','on','CompletePoll','on');
x0=[0;0];
[x fval]=patternsearch(@vectorizedc,x0,...
[],[],[],[],[],[],[],options)

MATLAB returns the following output:

Optimization terminated: mesh size less than options.TolMesh.
x =
-1.5737
1.0575
fval =
-10.0088

Vectorized Constraint Functions

Only nonlinear constraints need to be vectorized; bounds and linear
constraints are handled automatically. If there are nonlinear constraints, the
objective function and the nonlinear constraints all need to be vectorized in
order for the algorithm to compute in a vectorized manner.

4-75

Using Direct Search

The same considerations hold for constraint functions as for objective

functions: the initial point x0 determines the type of points (row or column
vectors) in the poll or search. If the initial point is a row vector of size k, the
matrix x passed to the constraint function has k columns. Similarly, if the
initial point is a column vector of size k, the matrix of poll or search points
has k rows. The figure Structure of Vectorized Functions on page 4-74 may
make this clear.
Your nonlinear constraint function returns two matrices, one for inequality
constraints, and one for equality constraints. Suppose there are nc nonlinear
inequality constraints and nceq nonlinear equality constraints. For row vector
x0, the constraint matrices have nc and nceq columns respectively, and the
number of rows is the same as in the input matrix. Similarly, for a column
vector x0, the constraint matrices have nc and nceq rows respectively, and the
number of columns is the same as in the input matrix. In figure Structure of
Vectorized Functions on page 4-74, Results includes both nc and nceq.

Example of Vectorized Objective and Constraints

Suppose that the nonlinear constraints are

x12 x22
+
1 (the interior of an ellipse),
9
4
x2 cosh ( x1 ) 1.
Write a function for these constraints for row-form x0 as follows:
function [c ceq] = ellipsecosh(x)
c(:,1)=x(:,1).^2/9+x(:,2).^2/4-1;
c(:,2)=cosh(x(:,1))-x(:,2)-1;
ceq=[];

Minimize vectorizedr (defined in Vectorized Objective Function on page

4-73) subject to the constraints ellipsecosh:
x0=[0,0];
options=psoptimset('Vectorized','on','CompletePoll','on');
[x fval]=patternsearch(@vectorizedr,x0,...
[],[],[],[],[],[],@ellipsecosh,options);

4-76

Pattern Search Examples: Setting Options

MATLAB returns the following output:

Optimization terminated: mesh size less than options.TolMesh
and constraint violation is less than options.TolCon.
x =
-1.3516

1.0612

fval =
-9.5394

4-77

4-78

Using Direct Search

5
Using the Genetic
Algorithm
What Is the Genetic Algorithm? on page 5-2
Performing a Genetic Algorithm Optimization on page 5-3
Example: Rastrigins Function on page 5-8
Some Genetic Algorithm Terminology on page 5-17
How the Genetic Algorithm Works on page 5-20
Description of the Nonlinear Constraint Solver on page 5-27
Genetic Algorithm Optimizations Using the Optimization Tool GUI on
page 5-29
Using the Genetic Algorithm from the Command Line on page 5-39
Genetic Algorithm Examples on page 5-49

Using the Genetic Algorithm

What Is the Genetic Algorithm?

The genetic algorithm is a method for solving both constrained and
unconstrained optimization problems that is based on natural selection, the
process that drives biological evolution. The genetic algorithm repeatedly
modifies a population of individual solutions. At each step, the genetic
algorithm selects individuals at random from the current population to be
parents and uses them to produce the children for the next generation. Over
successive generations, the population "evolves" toward an optimal solution.
You can apply the genetic algorithm to solve a variety of optimization
problems that are not well suited for standard optimization algorithms,
including problems in which the objective function is discontinuous,
nondifferentiable, stochastic, or highly nonlinear.
The genetic algorithm uses three main types of rules at each step to create
the next generation from the current population:
Selection rules select the individuals, called parents, that contribute to the
population at the next generation.
Crossover rules combine two parents to form children for the next
generation.
Mutation rules apply random changes to individual parents to form
children.
The genetic algorithm differs from a classical, derivative-based, optimization
algorithm in two main ways, as summarized in the following table.

5-2

Classical Algorithm

Genetic Algorithm

Generates a single point at each

iteration. The sequence of points
approaches an optimal solution.

Generates a population of points at

each iteration. The best point in the
population approaches an optimal
solution.

Selects the next point in the sequence

by a deterministic computation.

Selects the next population by

computation which uses random
number generators.

Performing a Genetic Algorithm Optimization

Performing a Genetic Algorithm Optimization

In this section...
Calling the Function ga at the Command Line on page 5-3
Using the Optimization Tool on page 5-4

Calling the Function ga at the Command Line

To use the genetic algorithm at the command line, call the genetic algorithm
function ga with the syntax
[x fval] = ga(@fitnessfun, nvars, options)

where
@fitnessfun is a handle to the fitness function.
nvars is the number of independent variables for the fitness function.
options is a structure containing options for the genetic algorithm. If you
do not pass in this argument, ga uses its default options.
The results are given by
x Point at which the final value is attained
fval Final value of the fitness function
Using the function ga is convenient if you want to
Return results directly to the MATLAB workspace
Run the genetic algorithm multiple times with different options, by calling
ga from a file
Using the Genetic Algorithm from the Command Line on page 5-39 provides
a detailed description of using the function ga and creating the options
structure.

5-3

Using the Genetic Algorithm

Using the Optimization Tool

To open the Optimization Tool, enter
optimtool('ga')

at the command line, or enter optimtool and then choose ga from the Solver
menu.

5-4

Performing a Genetic Algorithm Optimization

Set options

Expand or contract help

Choose solver

Enter problem
and constraints

Run solver

View results

See final point

You can also start the tool from the MATLAB Start menu as pictured:

5-5

Using the Genetic Algorithm

To use the Optimization Tool, you must first enter the following information:
Fitness function The objective function you want to minimize. Enter
the fitness function in the form @fitnessfun, where fitnessfun.m is a file
that computes the fitness function. Computing Objective Functions on
page 2-2 explains how write this file. The @ sign creates a function handle
to fitnessfun.

5-6

Performing a Genetic Algorithm Optimization

Number of variables The length of the input vector to the fitness

function. For the function my_fun described in Computing Objective
Functions on page 2-2, you would enter 2.
You can enter constraints or a nonlinear constraint function for the problem
in the Constraints pane. If the problem is unconstrained, leave these fields
blank.
To run the genetic algorithm, click the Start button. The tool displays the
results of the optimization in the Run solver and view results pane.
You can change the options for the genetic algorithm in the Options pane.
To view the options in one of the categories listed in the pane, click the +
sign next to it.
For more information,
See the Optimization Tool chapter in the Optimization Toolbox
documentation.
See Example: Rastrigins Function on page 5-8 for an example of using
the tool.

5-7

Using the Genetic Algorithm

Example: Rastrigins Function

In this section...
Rastrigins Function on page 5-8
Finding the Minimum of Rastrigins Function on page 5-10
Finding the Minimum from the Command Line on page 5-12
Displaying Plots on page 5-13

Rastrigins Function
This section presents an example that shows how to find the minimum of
Rastrigins function, a function that is often used to test the genetic algorithm.
For two independent variables, Rastrigins function is defined as

Ras( x) = 20 + x12 + x22 10 ( cos 2x1 + cos 2x2 ) .

Global Optimization Toolbox software contains the rastriginsfcn.m file,
which computes the values of Rastrigins function. The following figure shows
a plot of Rastrigins function.

5-8

Example: Rastrigins Function

Global minimum at [0 0]

As the plot shows, Rastrigins function has many local minimathe valleys
in the plot. However, the function has just one global minimum, which occurs
at the point [0 0] in the x-y plane, as indicated by the vertical line in the
plot, where the value of the function is 0. At any local minimum other than
[0 0], the value of Rastrigins function is greater than 0. The farther the
local minimum is from the origin, the larger the value of the function is at
that point.
Rastrigins function is often used to test the genetic algorithm, because its
many local minima make it difficult for standard, gradient-based methods
to find the global minimum.
The following contour plot of Rastrigins function shows the alternating
maxima and minima.

5-9

Using the Genetic Algorithm

1
0.8
0.6
0.4

Local maxima

0.2
0
0.2
0.4

Local minima

0.6
0.8
1
1

0.5

0.5

Global minimum at [0 0]

Finding the Minimum of Rastrigins Function

This section explains how to find the minimum of Rastrigins function using
the genetic algorithm.
Note Because the genetic algorithm uses random number generators, the
algorithm returns slightly different results each time you run it.
To find the minimum, do the following steps:
1 Enter optimtool('ga') at the command line to open the Optimization

Tool.
2 Enter the following in the Optimization Tool:

5-10

Example: Rastrigins Function

In the Fitness function field, enter @rastriginsfcn.

In the Number of variables field, enter 2, the number of independent
variables for Rastrigins function.
The Fitness function and Number of variables fields should appear
as shown in the following figure.

3 Click the Start button in the Run solver and view results pane, as

shown in the following figure.

While the algorithm is running, the Current iteration field displays

the number of the current generation. You can temporarily pause the
algorithm by clicking the Pause button. When you do so, the button name
changes to Resume. To resume the algorithm from the point at which
you paused it, click Resume.
When the algorithm is finished, the Run solver and view results pane
appears as shown in the following figure.
The Run solver and view results pane displays the following
information:

5-11

Using the Genetic Algorithm

The final value of the fitness function when the algorithm terminated:
Objective function value: 0.05531602101322264

Note that the value shown is very close to the actual minimum value
of Rastrigins function, which is 0. Genetic Algorithm Examples on
page 5-49 describes some ways to get a result that is closer to the actual
minimum.
The reason the algorithm terminated.
Optimization terminated:
average change in the fitness value less than options.TolFun.

The final point, which in this example is [0.001 -0.017].

Finding the Minimum from the Command Line

To find the minimum of Rastrigins function from the command line, enter
[x fval exitflag] = ga(@rastriginsfcn, 2)

This returns
Optimization terminated:
average change in the fitness value less than options.TolFun.
x =

5-12

Example: Rastrigins Function

0.0229

0.0106

fval =
0.1258

exitflag =
1

where
x is the final point returned by the algorithm.
fval is the fitness function value at the final point.
exitflag is integer value corresponding to the reason that the algorithm
terminated.
Note Because the genetic algorithm uses random number generators, the
algorithm returns slightly different results each time you run it.

Displaying Plots
The Plot functions pane enables you to display various plots that provide
information about the genetic algorithm while it is running. This information
can help you change options to improve the performance of the algorithm. For
example, to plot the best and mean values of the fitness function at each
generation, select the box next to Best fitness, as shown in the following
figure.

5-13

Using the Genetic Algorithm

When you click Start, the Optimization Tool displays a plot of the best and
mean values of the fitness function at each generation. When the algorithm
stops, the plot appears as shown in the following figure.

The points at the bottom of the plot denote the best fitness values, while
the points above them denote the averages of the fitness values in each
generation. The plot also displays the best and mean values in the current
generation numerically at the top.
To get a better picture of how much the best fitness values are decreasing, you
can change the scaling of the y-axis in the plot to logarithmic scaling. To do so,

5-14

Example: Rastrigins Function

1 Select Axes Properties from the Edit menu in the plot window to open

the Property Editor attached to your figure window as shown below.

2 Click the Y Axis tab.

3 In the Y Scale pane, select Log.

The plot now appears as shown in the following figure.

5-15

Using the Genetic Algorithm

Best: 0.0067796 Mean: 0.014788

10

10

10

10

10

10

10

20

30

40

50
60
generation

70

80

90

100

Typically, the best fitness value improves rapidly in the early generations,
when the individuals are farther from the optimum. The best fitness value
improves more slowly in later generations, whose populations are closer
to the optimal point.

5-16

Some Genetic Algorithm Terminology

Some Genetic Algorithm Terminology

In this section...
Fitness Functions on page 5-17
Individuals on page 5-17
Populations and Generations on page 5-18
Diversity on page 5-18
Fitness Values and Best Fitness Values on page 5-19
Parents and Children on page 5-19

Fitness Functions
The fitness function is the function you want to optimize. For standard
optimization algorithms, this is known as the objective function. The toolbox
software tries to find the minimum of the fitness function.
Write the fitness function as a file or anonymous function, and pass it as a
function handle input argument to the main genetic algorithm function.

Individuals
An individual is any point to which you can apply the fitness function. The
value of the fitness function for an individual is its score. For example, if
the fitness function is

f ( x1 , x2 , x3 ) = ( 2 x1 + 1) + ( 3 x2 + 4 ) + ( x3 2 ) ,
2

the vector (2, -3, 1), whose length is the number of variables in the problem, is
an individual. The score of the individual (2, 3, 1) is f(2, 3, 1) = 51.
An individual is sometimes referred to as a genome and the vector entries of
an individual as genes.

5-17

Using the Genetic Algorithm

Populations and Generations

A population is an array of individuals. For example, if the size of the
population is 100 and the number of variables in the fitness function is 3,
you represent the population by a 100-by-3 matrix. The same individual can
appear more than once in the population. For example, the individual (2, -3,
1) can appear in more than one row of the array.
At each iteration, the genetic algorithm performs a series of computations
on the current population to produce a new population. Each successive
population is called a new generation.

Diversity
Diversity refers to the average distance between individuals in a population.
A population has high diversity if the average distance is large; otherwise it
has low diversity. In the following figure, the population on the left has high
diversity, while the population on the right has low diversity.

Diversity is essential to the genetic algorithm because it enables the algorithm

to search a larger region of the space.

5-18

Some Genetic Algorithm Terminology

Fitness Values and Best Fitness Values

The fitness value of an individual is the value of the fitness function for that
individual. Because the toolbox software finds the minimum of the fitness
function, the best fitness value for a population is the smallest fitness value
for any individual in the population.

Parents and Children

To create the next generation, the genetic algorithm selects certain individuals
in the current population, called parents, and uses them to create individuals
in the next generation, called children. Typically, the algorithm is more likely
to select parents that have better fitness values.

5-19

Using the Genetic Algorithm

How the Genetic Algorithm Works

In this section...
Outline of the Algorithm on page 5-20
Initial Population on page 5-21
Creating the Next Generation on page 5-22
Plots of Later Generations on page 5-24
Stopping Conditions for the Algorithm on page 5-24

Outline of the Algorithm

The following outline summarizes how the genetic algorithm works:
1 The algorithm begins by creating a random initial population.
2 The algorithm then creates a sequence of new populations. At each step,

the algorithm uses the individuals in the current generation to create the
next population. To create the new population, the algorithm performs
the following steps:
a Scores each member of the current population by computing its fitness

value.
b Scales the raw fitness scores to convert them into a more usable range of

values.
c Selects members, called parents, based on their fitness.
d Some of the individuals in the current population that have lower fitness

are chosen as elite. These elite individuals are passed to the next
population.
e Produces children from the parents. Children are produced either by

making random changes to a single parentmutationor by combining

the vector entries of a pair of parentscrossover.
f

5-20

Replaces the current population with the children to form the next
generation.

How the Genetic Algorithm Works

3 The algorithm stops when one of the stopping criteria is met. See Stopping

Conditions for the Algorithm on page 5-24.

Initial Population
The algorithm begins by creating a random initial population, as shown in
the following figure.

1
Initial population
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
1

0.5

0.5

In this example, the initial population contains 20 individuals, which is the

default value of Population size in the Population options. Note that all
the individuals in the initial population lie in the upper-right quadrant of the
picture, that is, their coordinates lie between 0 and 1, because the default
value of Initial range in the Population options is [0;1].
If you know approximately where the minimal point for a function lies, you
should set Initial range so that the point lies near the middle of that range.
For example, if you believe that the minimal point for Rastrigins function is
near the point [0 0], you could set Initial range to be [-1;1]. However, as
this example shows, the genetic algorithm can find the minimum even with a
less than optimal choice for Initial range.

5-21

Using the Genetic Algorithm

Creating the Next Generation

At each step, the genetic algorithm uses the current population to create the
children that make up the next generation. The algorithm selects a group of
individuals in the current population, called parents, who contribute their
genesthe entries of their vectorsto their children. The algorithm usually
selects individuals that have better fitness values as parents. You can specify
the function that the algorithm uses to select the parents in the Selection
function field in the Selection options.
The genetic algorithm creates three types of children for the next generation:
Elite children are the individuals in the current generation with the
best fitness values. These individuals automatically survive to the next
generation.
Crossover children are created by combining the vectors of a pair of parents.
Mutation children are created by introducing random changes, or
mutations, to a single parent.
The following schematic diagram illustrates the three types of children.

5-22

How the Genetic Algorithm Works

Mutation and Crossover on page 5-63 explains how to specify the number of
children of each type that the algorithm generates and the functions it uses
to perform crossover and mutation.
The following sections explain how the algorithm creates crossover and
mutation children.

Crossover Children
The algorithm creates crossover children by combining pairs of parents in
the current population. At each coordinate of the child vector, the default
crossover function randomly selects an entry, or gene, at the same coordinate
from one of the two parents and assigns it to the child.

Mutation Children
The algorithm creates mutation children by randomly changing the genes of
individual parents. By default, the algorithm adds a random vector from a
Gaussian distribution to the parent.
The following figure shows the children of the initial population, that is, the
population at the second generation, and indicates whether they are elite,
crossover, or mutation children.
1
0.8
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
1

Elite children
Crossover children
Mutation children
0.5

0.5

1.5

5-23

Using the Genetic Algorithm

Plots of Later Generations

The following figure shows the populations at iterations 60, 80, 95, and 100.

Iteration 60

Iteration 80

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0.2

0.2

0.4

0.4

0.6

0.6

0.8
1
1

0.8
0.5

0.5

1
1

0.5

Iteration 95

0.5

0.5

Iteration 100

0.8

0.8

0.6

0.6

0.4

0.4

0.2

0.2

0.2

0.2

0.4

0.4

0.6

0.6

0.8
1
1

0.8
0.5

0.5

1
1

0.5

As the number of generations increases, the individuals in the population get

closer together and approach the minimum point [0 0].

Stopping Conditions for the Algorithm

The genetic algorithm uses the following conditions to determine when to stop:
Generations The algorithm stops when the number of generations
reaches the value of Generations.

5-24

How the Genetic Algorithm Works

Time limit The algorithm stops after running for an amount of time in
seconds equal to Time limit.
Fitness limit The algorithm stops when the value of the fitness function
for the best point in the current population is less than or equal to Fitness
limit.
Stall generations The algorithm stops when the weighted average
change in the fitness function value over Stall generations is less than
Function tolerance.
Stall time limit The algorithm stops if there is no improvement in
the objective function during an interval of time in seconds equal to Stall
time limit.
Function Tolerance The algorithm runs until the weighted average
change in the fitness function value over Stall generations is less than
Function tolerance.
Nonlinear constraint tolerance The Nonlinear constraint
tolerance is not used as stopping criterion. It is used to determine the
feasibility with respect to nonlinear constraints.
The algorithm stops as soon as any one of these conditions is met. You can
specify the values of these criteria in the Stopping criteria pane in the
Optimization Tool. The default values are shown in the pane.

5-25

Using the Genetic Algorithm

When you run the genetic algorithm, the Run solver and view results
panel displays the criterion that caused the algorithm to stop.
The options Stall time limit and Time limit prevent the algorithm from
running too long. If the algorithm stops due to one of these conditions, you
might improve your results by increasing the values of Stall time limit and
Time limit.

5-26

Description of the Nonlinear Constraint Solver

Description of the Nonlinear Constraint Solver

The genetic algorithm uses the Augmented Lagrangian Genetic Algorithm
(ALGA) to solve nonlinear constraint problems. The optimization problem
solved by the ALGA algorithm is

min f ( x)
x

such that

ci ( x) 0, i = 1 m
ceqi ( x) = 0, i = m + 1 mt
Ax b
Aeq x = beq
lb x ub,
where c(x) represents the nonlinear inequality constraints, ceq(x) represents
the equality constraints, m is the number of nonlinear inequality constraints,
and mt is the total number of nonlinear constraints.
The Augmented Lagrangian Genetic Algorithm (ALGA) attempts to solve a
nonlinear optimization problem with nonlinear constraints, linear constraints,
and bounds. In this approach, bounds and linear constraints are handled
separately from nonlinear constraints. A subproblem is formulated by
combining the fitness function and nonlinear constraint function using the
Lagrangian and the penalty parameters. A sequence of such optimization
problems are approximately minimized using the genetic algorithm such that
the linear constraints and bounds are satisfied.
A subproblem formulation is defined as
m

( x, , s, ) = f ( x) i si log(si ci ( x)) +
i=1

mt

i= m +1

i ci ( x) +

mt
ci ( x)2 ,
2 i= m+1

where the components i of the vector are nonnegative and are known as
Lagrange multiplier estimates. The elements si of the vector s are nonnegative

5-27

Using the Genetic Algorithm

shifts, and is the positive penalty parameter. The algorithm begins by using
an initial value for the penalty parameter (InitialPenalty).
The genetic algorithm minimizes a sequence of the subproblem, which is an
approximation of the original problem. When the subproblem is minimized
to a required accuracy and satisfies feasibility conditions, the Lagrangian
estimates are updated. Otherwise, the penalty parameter is increased
by a penalty factor (PenaltyFactor). This results in a new subproblem
formulation and minimization problem. These steps are repeated until the
stopping criteria are met. For a complete description of the algorithm, see the
following references:
[1] Conn, A. R., N. I. M. Gould, and Ph. L. Toint. A Globally Convergent
Augmented Lagrangian Algorithm for Optimization with General Constraints
and Simple Bounds, SIAM Journal on Numerical Analysis, Volume 28,
Number 2, pages 545572, 1991.
[2] Conn, A. R., N. I. M. Gould, and Ph. L. Toint. A Globally Convergent
Augmented Lagrangian Barrier Algorithm for Optimization with General
Inequality Constraints and Simple Bounds, Mathematics of Computation,
Volume 66, Number 217, pages 261288, 1997.

5-28

Genetic Algorithm Optimizations Using the Optimization Tool GUI

Genetic Algorithm Optimizations Using the Optimization

Tool GUI
In this section...
Introduction on page 5-29
Displaying Plots on page 5-29
Example: Creating a Custom Plot Function on page 5-30
Reproducing Your Results on page 5-33
Example: Resuming the Genetic Algorithm from the Final Population
on page 5-34

Introduction
The Optimization Tool GUI is described in the chapter Optimization Tool in
the Optimization Toolbox documentation. This section describes some places
where there are some differences between the use of the genetic algorithm in
the Optimization Tool and the use of other solvers.

Displaying Plots
The Plot functions pane, shown in the following figure, enables you to
display various plots of the results of the genetic algorithm.

5-29

Using the Genetic Algorithm

Select the check boxes next to the plots you want to display. For example, if
you select Best fitness and Best individual, and run the example described
in Example: Rastrigins Function on page 5-8, the tool displays plots similar
to those shown in the following figure.

The upper plot displays the best and mean fitness values in each generation.
The lower plot displays the coordinates of the point with the best fitness value
in the current generation.
Note When you display more than one plot, you can open a larger version
of a plot in a separate window. Right-click (Ctrl-click for Mac) on a blank
area in a plot while ga is running, or after it has stopped, and choose the
sole menu item.
Plot Options on page 9-30 describes the types of plots you can create.

Example: Creating a Custom Plot Function

If none of the plot functions that come with the software is suitable for the
output you want to plot, you can write your own custom plot function, which
the genetic algorithm calls at each generation to create the plot. This example

5-30

Genetic Algorithm Optimizations Using the Optimization Tool GUI

shows how to create a plot function that displays the change in the best fitness
value from the previous generation to the current generation.
This section covers the following topics:
Creating the Custom Plot Function on page 5-31
Using the Plot Function on page 5-32
How the Plot Function Works on page 5-32

Creating the Custom Plot Function

To create the plot function for this example, copy and paste the following code
into a new file in the MATLAB Editor.
function state = gaplotchange(options, state, flag)
% GAPLOTCHANGE Plots the logarithmic change in the best score from the
% previous generation.
%
persistent last_best % Best score in the previous generation
if(strcmp(flag,'init')) % Set up the plot
set(gca,'xlim',[1,options.Generations],'Yscale','log');
hold on;
xlabel Generation
title('Change in Best Fitness Value')
end
best = min(state.Score); % Best score in the current generation
if state.Generation == 0 % Set last_best to best.
last_best = best;
else
change = last_best - best; % Change in best score
last_best=best;
plot(state.Generation, change, '.r');
title(['Change in Best Fitness Value'])
end

Then save the file as gaplotchange.m in a folder on the MATLAB path.

5-31

Using the Genetic Algorithm

Using the Plot Function

To use the custom plot function, select Custom in the Plot functions pane
and enter @gaplotchange in the field to the right. To compare the custom plot
with the best fitness value plot, also select Best fitness. Now, if you run the
example described in Example: Rastrigins Function on page 5-8, the tool
displays plots similar to those shown in the following figure.
Best: 0.0020445 Mean: 0.033263

Fitness value

20
Best fitness
Mean fitness

15
10
5
0

10

20

30

40

10

20

30

40

50
60
70
Generation
Change in Best Fitness Value

80

90

100

80

90

100

10

10

10

10

50
60
Generation

70

Note that because the scale of the y-axis in the lower custom plot is
logarithmic, the plot only shows changes that are greater then 0. The
logarithmic scale enables you to see small changes in the fitness function
that the upper plot does not reveal.

How the Plot Function Works

The plot function uses information contained in the following structures,
which the genetic algorithm passes to the function as input arguments:
options The current options settings
state Information about the current generation

5-32

Genetic Algorithm Optimizations Using the Optimization Tool GUI

flag String indicating the current status of the algorithm

The most important lines of the plot function are the following:
persistent last_best
Creates the persistent variable last_bestthe best score in the previous
generation. Persistent variables are preserved over multiple calls to the
plot function.
set(gca,'xlim',[1,options.Generations],'Yscale','log');
Sets up the plot before the algorithm starts. options.Generations is the
maximum number of generations.
best = min(state.Score)
The field state.Score contains the scores of all individuals in the current
population. The variable best is the minimum score. For a complete
description of the fields of the structure state, see Structure of the Plot
Functions on page 9-32.
change = last_best - best
The variable change is the best score at the previous generation minus the
best score in the current generation.
plot(state.Generation, change, '.r')
Plots the change at the current generation, whose number is contained in
state.Generation.
The code for gaplotchange contains many of the same elements as the code
for gaplotbestf, the function that creates the best fitness plot.

Reproducing Your Results

To reproduce the results of the last run of the genetic algorithm, select the
Use random states from previous run check box. This resets the states of
the random number generators used by the algorithm to their previous values.
If you do not change any other settings in the Optimization Tool, the next time
you run the genetic algorithm, it returns the same results as the previous run.

5-33

Using the Genetic Algorithm

Normally, you should leave Use random states from previous run
unselected to get the benefit of randomness in the genetic algorithm. Select
the Use random states from previous run check box if you want to analyze
the results of that particular run or show the exact results to others. After
the algorithm has run, you can clear your results using the Clear Status
button in the Run solver settings.
Note If you select Include information needed to resume this run,
then selecting Use random states from previous run has no effect on the
initial population created when you import the problem and run the genetic
algorithm on it. The latter option is only intended to reproduce results from
the beginning of a new run, not from a resumed run.

Example: Resuming the Genetic Algorithm from the

Final Population
The following example shows how export a problem so that when you import
it and click Start, the genetic algorithm resumes from the final population
saved with the exported problem. To run the example, enter the following
information in the Optimization Tool:
Set Fitness function to @ackleyfcn, which computes Ackleys function, a
test function provided with the software.
Set Number of variables to 10.
Select Best fitness in the Plot functions pane.
Click Start.
This displays the following plot.

5-34

Genetic Algorithm Optimizations Using the Optimization Tool GUI

Suppose you want to experiment by running the genetic algorithm with other
options settings, and then later restart this run from its final population with
its current options settings. You can do this using the following steps:
1 Click Export to Workspace.
2 In the dialog box that appears,

Select Export problem and options to a MATLAB structure

named.
Enter a name for the problem and options, such as ackley_uniform,
in the text field.
Select Include information needed to resume this run.
The dialog box should now appear as in the following figure.

5-35

Using the Genetic Algorithm

3 Click OK.

This exports the problem and options to a structure in the MATLAB

workspace. You can view the structure in the MATLAB Command Window
by entering
ackley_uniform
ackley_uniform =
fitnessfcn: @ackleyfcn
nvars: 10
Aineq: []
bineq: []
Aeq: []
beq: []
lb: []
ub: []
nonlcon: []
rngstate: []
solver: 'ga'
options: [1x1 struct]

After running the genetic algorithm with different options settings or even a
different fitness function, you can restore the problem as follows:

5-36

Genetic Algorithm Optimizations Using the Optimization Tool GUI

1 Select Import Problem from the File menu. This opens the dialog box

shown in the following figure.

2 Select ackley_uniform.
3 Click Import.

This sets the Initial population and Initial scores fields in the Population
panel to the final population of the run before you exported the problem.
All other options are restored to their setting during that run. When you
click Start, the genetic algorithm resumes from the saved final population.
The following figure shows the best fitness plots from the original run and
the restarted run.

5-37

Using the Genetic Algorithm

Best: 2.9882 Mean: 3.0033

Best: 2.9861 Mean: 3.6454

5.5

5.5
Best fitness
Mean fitness

4.5

4.5
Fitness value

Fitness value

Best fitness
Mean fitness

3.5

3.5

2.5

10

20

30

40

50
60
Generation

First run

70

80

90

100

2.5

10

20

30

40

50
60
Generation

70

80

90

100

Run resumes here

Note If, after running the genetic algorithm with the imported problem,
you want to restore the genetic algorithms default behavior of generating a
random initial population, delete the population in the Initial population
field.

5-38

Using the Genetic Algorithm from the Command Line

Using the Genetic Algorithm from the Command Line

In this section...
Running ga with the Default Options on page 5-39
Setting Options for ga at the Command Line on page 5-40
Using Options and Problems from the Optimization Tool on page 5-43
Reproducing Your Results on page 5-44
Resuming ga from the Final Population of a Previous Run on page 5-45
Running ga From a File on page 5-46

Running ga with the Default Options

To run the genetic algorithm with the default options, call ga with the syntax
[x fval] = ga(@fitnessfun, nvars)

The input arguments to ga are

@fitnessfun A function handle to the file that computes the fitness
function. Computing Objective Functions on page 2-2 explains how to
write this file.
nvars The number of independent variables for the fitness function.
The output arguments are
x The final point
fval The value of the fitness function at x
For a description of additional input and output arguments, see the reference
page for ga.
You can run the example described in Example: Rastrigins Function on
page 5-8 from the command line by entering
[x fval] = ga(@rastriginsfcn, 2)

5-39

Using the Genetic Algorithm

This returns
x =
0.0027

-0.0052

fval =
0.0068

Additional Output Arguments

To get more information about the performance of the genetic algorithm, you
can call ga with the syntax
[x fval exitflag output population scores] = ga(@fitnessfcn, nvars)

Besides x and fval, this function returns the following additional output
arguments:
exitflag Integer value corresponding to the reason the algorithm
terminated
output Structure containing information about the performance of the
algorithm at each generation
population Final population
scores Final scores
See the ga reference page for more information about these arguments.

Setting Options for ga at the Command Line

You can specify any of the options that are available for ga by passing an
options structure as an input argument to ga using the syntax
[x fval] = ga(@fitnessfun, nvars, [],[],[],[],[],[],[],options)

This syntax does not specify any linear equality, linear inequality, or
nonlinear constraints.
You create the options structure using the function gaoptimset.

5-40

Using the Genetic Algorithm from the Command Line

options = gaoptimset(@ga)

This returns the structure options with the default values for its fields.
options =
PopulationType:
PopInitRange:
PopulationSize:
EliteCount:
CrossoverFraction:
ParetoFraction:
MigrationDirection:
MigrationInterval:
MigrationFraction:
Generations:
TimeLimit:
FitnessLimit:
StallGenLimit:
StallTimeLimit:
TolFun:
TolCon:
InitialPopulation:
InitialScores:
InitialPenalty:
PenaltyFactor:
PlotInterval:
CreationFcn:
FitnessScalingFcn:
SelectionFcn:
CrossoverFcn:
MutationFcn:
DistanceMeasureFcn:
HybridFcn:
Display:
PlotFcns:
OutputFcns:
Vectorized:
UseParallel:

'doubleVector'
[2x1 double]
20
2
0.8000
[]
'forward'
20
0.2000
100
Inf
-Inf
50
Inf
1.0000e-006
1.0000e-006
[]
[]
10
100
1
@gacreationuniform
@fitscalingrank
@selectionstochunif
@crossoverscattered
{[1x1 function_handle]
[]
[]
'final'
[]
[]
'off'
'never'

[1]

[1]}

5-41

Using the Genetic Algorithm

The function ga uses these default values if you do not pass in options as an
input argument.
The value of each option is stored in a field of the options structure, such as
options.PopulationSize. You can display any of these values by entering
options. followed by the name of the field. For example, to display the size
of the population for the genetic algorithm, enter
options.PopulationSize
ans =
20

To create an options structure with a field value that is different from the
default for example to set PopulationSize to 100 instead of its default
value 20 enter
options = gaoptimset('PopulationSize', 100)

This creates the options structure with all values set to their defaults except
for PopulationSize, which is set to 100.
If you now enter,
ga(@fitnessfun,nvars,[],[],[],[],[],[],[],options)
ga runs the genetic algorithm with a population size of 100.

If you subsequently decide to change another field in the options structure,

such as setting PlotFcns to @gaplotbestf, which plots the best fitness
function value at each generation, call gaoptimset with the syntax
options = gaoptimset(options, 'PlotFcns', @plotbestf)

This preserves the current values of all fields of options except for PlotFcns,
which is changed to @plotbestf. Note that if you omit the input argument
options, gaoptimset resets PopulationSize to its default value 20.
You can also set both PopulationSize and PlotFcns with the single command

5-42

Using the Genetic Algorithm from the Command Line

options = gaoptimset('PopulationSize',100,'PlotFcns',@plotbestf)

Using Options and Problems from the Optimization

Tool
As an alternative to creating an options structure using gaoptimset, you can
set the values of options in the Optimization Tool and then export the options
to a structure in the MATLAB workspace, as described in the Importing and
Exporting Your Work section of the Optimization Toolbox documentation.
If you export the default options in the Optimization Tool, the resulting
structure options has the same settings as the default structure returned
by the command
options = gaoptimset(@ga)

except that the option 'Display' defaults to 'off' in an exported structure,

and is 'final' in the default at the command line.
If you export a problem structure, ga_problem, from the Optimization Tool,
you can apply ga to it using the syntax
[x fval] = ga(ga_problem)

The problem structure contains the following fields:

fitnessfcn Fitness function
nvars Number of variables for the problem
Aineq Matrix for inequality constraints
Bineq Vector for inequality constraints
Aeq Matrix for equality constraints
Beq Vector for equality constraints
LB Lower bound on x
UB Upper bound on x
nonlcon Nonlinear constraint function
options Options structure

5-43

Using the Genetic Algorithm

Reproducing Your Results

Because the genetic algorithm is stochasticthat is, it makes random
choicesyou get slightly different results each time you run the genetic
algorithm. The algorithm uses the default MATLAB pseudorandom
number stream. For more information about random number streams, see
RandStream. Each time ga calls the stream, its state changes. So that the
next time ga calls the stream, it returns a different random number. This is
why the output of ga differs each time you run it.
If you need to reproduce your results exactly, you can call ga with an output
argument that contains the current state of the default stream, and then reset
the state to this value before running ga again. For example, to reproduce the
output of ga applied to Rastrigins function, call ga with the syntax
[x fval exitflag output] = ga(@rastriginsfcn, 2);

Suppose the results are

x =
0.0027

-0.0052

fval =
0.0068

The state of the stream is stored in output.rngstate:

output =
problemtype:
rngstate:
generations:
funccount:
message:

'unconstrained'
[1x1 struct]
68
1380
'Optimization terminated: average change in
the fitness value less than options.TolFun.'

To reset the state, enter

stream = RandStream.getDefaultStream;
stream.State = output.rngstate.state;

If you now run ga a second time, you get the same results.

5-44

Using the Genetic Algorithm from the Command Line

You can reproduce your run in the Optimization Tool by checking the box Use
random states from previous run in the Run solver and view results
section.

Note If you do not need to reproduce your results, it is better not to set the
state of the stream, so that you get the benefit of the randomness in the
genetic algorithm.

Resuming ga from the Final Population of a Previous

Run
By default, ga creates a new initial population each time you run it. However,
you might get better results by using the final population from a previous run
as the initial population for a new run. To do so, you must have saved the
final population from the previous run by calling ga with the syntax
[x,fval,exitflag,output,final_pop] = ga(@fitnessfcn, nvars);

The last output argument is the final population. To run ga using final_pop
as the initial population, enter
options = gaoptimset('InitialPop', final_pop);
[x,fval,exitflag,output,final_pop2] = ...
ga(@fitnessfcn,nvars,[],[],[],[],[],[],[],options);

You can then use final_pop2, the final population from the second run, as
the initial population for a third run.
In Optimization Tool, you can choose to export a problem in a way that lets
you resume the run. Simply check the box Include information needed
to resume this run when exporting the problem.

5-45

Using the Genetic Algorithm

This saves the final population, which becomes the initial population when
imported.
If you want to run a problem that was saved with the final population, but
would rather not use the initial population, simply delete or otherwise change
the initial population in the Options > Population pane.

Running ga From a File

The command-line interface enables you to run the genetic algorithm many
times, with different options settings, using a file. For example, you can run
the genetic algorithm with different settings for Crossover fraction to see
which one gives the best results. The following code runs the function ga 21
times, varying options.CrossoverFraction from 0 to 1 in increments of
0.05, and records the results.
options = gaoptimset('Generations',300);
strm = RandStream('mt19937ar','Seed',6525);
RandStream.setDefaultStream(strm);
record=[];
for n=0:.05:1
options = gaoptimset(options,'CrossoverFraction', n);
[x fval]=ga(@rastriginsfcn, 10,[],[],[],[],[],[],[],options);
record = [record; fval];
end

5-46

Using the Genetic Algorithm from the Command Line

You can plot the values of fval against the crossover fraction with the
following commands:
plot(0:.05:1, record);
xlabel('Crossover Fraction');
ylabel('fval')

The following plot appears.

The plot suggests that you get the best results by setting
options.CrossoverFraction to a value somewhere between 0.6 and 0.95.
You can get a smoother plot of fval as a function of the crossover fraction by
running ga 20 times and averaging the values of fval for each crossover
fraction. The following figure shows the resulting plot.

5-47

Using the Genetic Algorithm

The plot narrows the range of best choices for options.CrossoverFraction

to values between 0.7 and 0.9.

5-48

Genetic Algorithm Examples

Genetic Algorithm Examples

In this section...
Improving Your Results on page 5-49
Population Diversity on page 5-49
Fitness Scaling on page 5-59
Selection on page 5-62
Reproduction Options on page 5-63
Mutation and Crossover on page 5-63
Setting the Amount of Mutation on page 5-64
Setting the Crossover Fraction on page 5-66
Comparing Results for Varying Crossover Fractions on page 5-70
Example: Global vs. Local Minima with GA on page 5-72
Using a Hybrid Function on page 5-76
Setting the Maximum Number of Generations on page 5-80
Vectorizing the Fitness Function on page 5-81
Constrained Minimization Using ga on page 5-82

Improving Your Results

To get the best results from the genetic algorithm, you usually need to
experiment with different options. Selecting the best options for a problem
involves start and error. This section describes some ways you can change
options to improve results. For a complete description of the available options,
see Genetic Algorithm Options on page 9-29.

Population Diversity
One of the most important factors that determines the performance of the
genetic algorithm performs is the diversity of the population. If the average
distance between individuals is large, the diversity is high; if the average
distance is small, the diversity is low. Getting the right amount of diversity is

5-49

Using the Genetic Algorithm

a matter of start and error. If the diversity is too high or too low, the genetic
algorithm might not perform well.
This section explains how to control diversity by setting the Initial range of
the population. Setting the Amount of Mutation on page 5-64 describes how
the amount of mutation affects diversity.
This section also explains how to set the population size.

Example: Setting the Initial Range

By default, ga creates a random initial population using a creation function.
You can specify the range of the vectors in the initial population in the Initial
range field in Population options.
Note The initial range restricts the range of the points in the initial
population by specifying the lower and upper bounds. Subsequent generations
can contain points whose entries do not lie in the initial range. Set upper and
lower bounds for all generations in the Bounds fields in the Constraints
panel.
If you know approximately where the solution to a problem lies, specify the
initial range so that it contains your guess for the solution. However, the
genetic algorithm can find the solution even if it does not lie in the initial
range, if the population has enough diversity.
The following example shows how the initial range affects the performance
of the genetic algorithm. The example uses Rastrigins function, described
in Example: Rastrigins Function on page 5-8. The minimum value of the
function is 0, which occurs at the origin.
To run the example, open the ga solver in the Optimization Tool by entering
optimtool('ga') at the command line. Set the following:
Set Fitness function to @rastriginsfcn.
Set Number of variables to 2.
Select Best fitness in the Plot functions pane of the Options pane.

5-50

Genetic Algorithm Examples

Select Distance in the Plot functions pane.

Set Initial range in the Population pane of the Options pane to [1;1.1].
Click Start in Run solver and view results. Although the results of genetic
algorithm computations are random, your results are similar to the following
figure, with a best fitness function value of approximately 2.

The upper plot, which displays the best fitness at each generation, shows
little progress in lowering the fitness value. The lower plot shows the average
distance between individuals at each generation, which is a good measure of
the diversity of a population. For this setting of initial range, there is too little
diversity for the algorithm to make progress.

5-51

Using the Genetic Algorithm

Next, try setting Initial range to [1;100] and running the algorithm. This
time the results are more variable. You might obtain a plot with a best fitness
value of 35, as in the following plot. You might obtain different results.

This time, the genetic algorithm makes progress, but because the average
distance between individuals is so large, the best individuals are far from
the optimal solution.
Finally, set Initial range to [1;2] and run the genetic algorithm. Again,
there is variability in the result, but you might obtain a result similar to the
following figure. Run the optimization several times, and you eventually
obtain a final point near [0;0], with a fitness function value near 0.

5-52

Genetic Algorithm Examples

The diversity in this case is better suited to the problem, so ga usually returns
a better result than in the previous two cases.

Example: Linearly Constrained Population and Custom Plot

Function
This example shows how the default creation function for linearly constrained
problems, gacreationlinearfeasible, creates a well-dispersed population
that satisfies linear constraints and bounds. It also contains an example of a
custom plot function.
The problem uses the objective function in lincontest6.m, a quadratic:

5-53

Using the Genetic Algorithm

x12
+ x22 x1 x2 2 x1 6 x2 .
2
To see code for the function, enter type lincontest6. The constraints are
three linear inequalities:
f ( x) =

x1 + x2 2,
x1 + 2x2 2,
2x1 + x2 3.
Also, the variables xi are restricted to be positive.
1 Create a custom plot function file containing the following code:

function state = gaplotshowpopulation2(unused,state,flag,fcn)

% This plot function works in 2-d only
if size(state.Population,2) > 2
return;
end
if nargin < 4 % check to see if fitness function exists
fcn = [];
end
% Dimensions to plot
dimensionsToPlot = [1 2];
switch flag
% Plot initialization
case 'init'
pop = state.Population(:,dimensionsToPlot);
plotHandle = plot(pop(:,1),pop(:,2),'*');
set(plotHandle,'Tag','gaplotshowpopulation2')
title('Population plot in two dimension',...
'interp','none')
xlabelStr = sprintf('%s %s','Variable ',...
num2str(dimensionsToPlot(1)));
ylabelStr = sprintf('%s %s','Variable ',...
num2str(dimensionsToPlot(2)));
xlabel(xlabelStr,'interp','none');
ylabel(ylabelStr,'interp','none');
hold on;

5-54

Genetic Algorithm Examples

% plot the inequalities

plot([0 1.5],[2 0.5],'m-.') % x1 + x2 <= 2
plot([0 1.5],[1 3.5/2],'m-.'); % -x1 + 2*x2 <= 2
plot([0 1.5],[3 0],'m-.'); % 2*x1 + x2 <= 3
% plot lower bounds
plot([0 0], [0 2],'m-.'); % lb = [ 0 0];
plot([0 1.5], [0 0],'m-.'); % lb = [ 0 0];
set(gca,'xlim',[-0.7,2.2])
set(gca,'ylim',[-0.7,2.7])
% Contour plot the objective function
if ~isempty(fcn) % if there is a fitness function
range = [-0.5,2;-0.5,2];
pts = 100;
span = diff(range')/(pts - 1);
x = range(1,1): span(1) : range(1,2);
y = range(2,1): span(2) : range(2,2);
pop = zeros(pts * pts,2);
values = zeros(pts,1);
k = 1;
for i = 1:pts
for j = 1:pts
pop(k,:) = [x(i),y(j)];
values(k) = fcn(pop(k,:));
k = k + 1;
end
end
values = reshape(values,pts,pts);
contour(x,y,values);
colorbar
end
% Pause for three seconds to view the initial plot
pause(3);
case 'iter'
pop = state.Population(:,dimensionsToPlot);
plotHandle = findobj(get(gca,'Children'),'Tag',...
'gaplotshowpopulation2');
set(plotHandle,'Xdata',pop(:,1),'Ydata',pop(:,2));
end

5-55

Using the Genetic Algorithm

The custom plot function plots the lines representing the linear inequalities
and bound constraints, plots level curves of the fitness function, and plots
the population as it evolves. This plot function expects to have not only the
usual inputs (options,state,flag), but also a function handle to the
fitness function, @lincontest6 in this example. To generate level curves,
the custom plot function needs the fitness function.
2 At the command line, enter the constraints as a matrix and vectors:

A = [1,1;-1,2;2,1]; b = [2;2;3]; lb = zeros(2,1);

3 Set options to use gaplotshowpopulation2, and pass in @lincontest6 as

the fitness function handle:

options = gaoptimset('PlotFcn',...
{{@gaplotshowpopulation2,@lincontest6}});
4 Run the optimization using options:

[x,fval] = ga(@lincontest6,2,A,b,[],[],lb,[],[],options);

A plot window appears showing the linear constraints, bounds, level curves of
the objective function, and initial distribution of the population:

5-56

Genetic Algorithm Examples

You can see that the initial population is biased to lie on the constraints.
The population eventually concentrates around the minimum point:

5-57

Using the Genetic Algorithm

Setting the Population Size

The Population size field in Population options determines the size of the
population at each generation. Increasing the population size enables the
genetic algorithm to search more points and thereby obtain a better result.
However, the larger the population size, the longer the genetic algorithm
takes to compute each generation.
Note You should set Population size to be at least the value of Number
of variables, so that the individuals in each population span the space
being searched.
You can experiment with different settings for Population size that return
good results without taking a prohibitive amount of time to run.

5-58

Genetic Algorithm Examples

Fitness Scaling
Fitness scaling converts the raw fitness scores that are returned by the
fitness function to values in a range that is suitable for the selection function.
The selection function uses the scaled fitness values to select the parents of
the next generation. The selection function assigns a higher probability of
selection to individuals with higher scaled values.
The range of the scaled values affects the performance of the genetic
algorithm. If the scaled values vary too widely, the individuals with the
highest scaled values reproduce too rapidly, taking over the population gene
pool too quickly, and preventing the genetic algorithm from searching other
areas of the solution space. On the other hand, if the scaled values vary only a
little, all individuals have approximately the same chance of reproduction and
the search will progress very slowly.
The default fitness scaling option, Rank, scales the raw scores based on the
rank of each individual instead of its score. The rank of an individual is its
position in the sorted scores: the rank of the most fit individual is 1, the next
most fit is 2, and so on. The rank scaling function assigns scaled values so that
The scaled value of an individual with rank n is proportional to 1 / n .
The sum of the scaled values over the entire population equals the number
of parents needed to create the next generation.
Rank fitness scaling removes the effect of the spread of the raw scores.
The following plot shows the raw scores of a typical population of 20
individuals, sorted in increasing order.

5-59

Using the Genetic Algorithm

Raw Scores of Population

140
130
120

Score

110
100
90
80
70
60
50

10
Sorted individuals

15

20

The following plot shows the scaled values of the raw scores using rank
scaling.
Scaled Values Using Rank Scaling
4.5
4

Scaled value

3.5
3
2.5
2
1.5
1
0.5

10
Sorted individuals

15

20

Because the algorithm minimizes the fitness function, lower raw scores have
higher scaled values. Also, because rank scaling assigns values that depend

5-60

Genetic Algorithm Examples

only on an individuals rank, the scaled values shown would be the same for
any population of size 20 and number of parents equal to 32.

Comparing Rank and Top Scaling

To see the effect of scaling, you can compare the results of the genetic
algorithm using rank scaling with one of the other scaling options, such as
Top. By default, top scaling assigns 40 percent of the fittest individuals to the
same scaled value and assigns the rest of the individuals to value 0. Using
the default selection function, only 40 percent of the fittest individuals can
be selected as parents.
The following figure compares the scaled values of a population of size 20 with
number of parents equal to 32 using rank and top scaling.
Comparison of Rank and Top Scaling
Rank scaling
Top scaling

8
7

Scaled value

6
5
4
3
2
1
0
0

10
Sorted individuals

15

20

Because top scaling restricts parents to the fittest individuals, it creates

less diverse populations than rank scaling. The following plot compares the

5-61

Using the Genetic Algorithm

variances of distances between individuals at each generation using rank

and top scaling.
Variance of Distance Between Individuals Using Rank and Top Scaling
0.8
Variance using rank scaling
Variance using top scaling

0.7
0.6

Variance

0.5
0.4
0.3
0.2
0.1
0

10

20

30

40

50
60
Generation

70

80

90

100

Selection
The selection function chooses parents for the next generation based on their
scaled values from the fitness scaling function. An individual can be selected
more than once as a parent, in which case it contributes its genes to more than
one child. The default selection option, Stochastic uniform, lays out a line
in which each parent corresponds to a section of the line of length proportional
to its scaled value. The algorithm moves along the line in steps of equal size.
At each step, the algorithm allocates a parent from the section it lands on.
A more deterministic selection option is Remainder, which performs two steps:
In the first step, the function selects parents deterministically according
to the integer part of the scaled value for each individual. For example,
if an individuals scaled value is 2.3, the function selects that individual
twice as a parent.

5-62

Genetic Algorithm Examples

In the second step, the selection function selects additional parents using
the fractional parts of the scaled values, as in stochastic uniform selection.
The function lays out a line in sections, whose lengths are proportional to
the fractional part of the scaled value of the individuals, and moves along
the line in equal steps to select the parents.
Note that if the fractional parts of the scaled values all equal 0, as can
occur using Top scaling, the selection is entirely deterministic.

Reproduction Options
Reproduction options control how the genetic algorithm creates the next
generation. The options are
Elite count The number of individuals with the best fitness values
in the current generation that are guaranteed to survive to the next
generation. These individuals are called elite children. The default value
of Elite count is 2.
When Elite count is at least 1, the best fitness value can only decrease
from one generation to the next. This is what you want to happen, since the
genetic algorithm minimizes the fitness function. Setting Elite count to a
high value causes the fittest individuals to dominate the population, which
can make the search less effective.
Crossover fraction The fraction of individuals in the next generation,
other than elite children, that are created by crossover. Setting the
Crossover Fraction on page 5-66 describes how the value of Crossover
fraction affects the performance of the genetic algorithm.

Mutation and Crossover

The genetic algorithm uses the individuals in the current generation to create
the children that make up the next generation. Besides elite children, which
correspond to the individuals in the current generation with the best fitness
values, the algorithm creates
Crossover children by selecting vector entries, or genes, from a pair of
individuals in the current generation and combines them to form a child
Mutation children by applying random changes to a single individual in the
current generation to create a child

5-63

Using the Genetic Algorithm

Both processes are essential to the genetic algorithm. Crossover enables the
algorithm to extract the best genes from different individuals and recombine
them into potentially superior children. Mutation adds to the diversity of
a population and thereby increases the likelihood that the algorithm will
generate individuals with better fitness values.
See Creating the Next Generation on page 5-22 for an example of how the
genetic algorithm applies mutation and crossover.
You can specify how many of each type of children the algorithm creates as
follows:
Elite count, in Reproduction options, specifies the number of elite
children.
Crossover fraction, in Reproduction options, specifies the fraction of
the population, other than elite children, that are crossover children.
For example, if the Population size is 20, the Elite count is 2, and the
Crossover fraction is 0.8, the numbers of each type of children in the next
generation are as follows:
There are two elite children.
There are 18 individuals other than elite children, so the algorithm rounds
0.8*18 = 14.4 to 14 to get the number of crossover children.
The remaining four individuals, other than elite children, are mutation
children.

Setting the Amount of Mutation

The genetic algorithm applies mutations using the option that you specify
on the Mutation function pane. The default mutation option, Gaussian,
adds a random number, or mutation, chosen from a Gaussian distribution,
to each entry of the parent vector. Typically, the amount of mutation, which
is proportional to the standard deviation of the distribution, decreases at
each new generation. You can control the average amount of mutation that
the algorithm applies to a parent in each generation through the Scale and
Shrink options:

5-64

Genetic Algorithm Examples

Scale controls the standard deviation of the mutation at the first

generation, which is Scale multiplied by the range of the initial population,
which you specify by the Initial range option.
Shrink controls the rate at which the average amount of mutation
decreases. The standard deviation decreases linearly so that its final
value equals 1 - Shrink times its initial value at the first generation. For
example, if Shrink has the default value of 1, then the amount of mutation
decreases to 0 at the final step.
You can see the effect of mutation by selecting the plot options Distance and
Range, and then running the genetic algorithm on a problem such as the
one described in Example: Rastrigins Function on page 5-8. The following
figure shows the plot.
Average Distance between individuals
4
3
2
1
0

10

20

30

40
50
60
70
generation
Best, Worst, and Mean scores

80

90

100

200
150
100
50
0

20

40

60
generation

80

100

120

The upper plot displays the average distance between points in each
generation. As the amount of mutation decreases, so does the average distance
between individuals, which is approximately 0 at the final generation. The
lower plot displays a vertical line at each generation, showing the range
from the smallest to the largest fitness value, as well as mean fitness value.
As the amount of mutation decreases, so does the range. These plots show

5-65

Using the Genetic Algorithm

that reducing the amount of mutation decreases the diversity of subsequent

generations.
For comparison, the following figure shows the plots for Distance and Range
when you set Shrink to 0.5.
Average Distance between individuals
4
3
2
1
0

10

20

30

40

50
60
70
generation
Best, Worst, and Mean scores

80

90

100

250
200
150
100
50
0

20

40

60
generation

80

100

120

With Shrink set to 0.5, the average amount of mutation decreases by a

factor of 1/2 by the final generation. As a result, the average distance between
individuals decreases by approximately the same factor.

Setting the Crossover Fraction

The Crossover fraction field, in the Reproduction options, specifies the
fraction of each population, other than elite children, that are made up of
crossover children. A crossover fraction of 1 means that all children other than
elite individuals are crossover children, while a crossover fraction of 0 means
that all children are mutation children. The following example show that
neither of these extremes is an effective strategy for optimizing a function.

5-66

Genetic Algorithm Examples

The example uses the fitness function whose value at a point is the sum of the
absolute values of the coordinates at the points. That is,

f ( x1 , x2 ,..., xn ) = x1 + x2 +  + xn .
You can define this function as an anonymous function by setting Fitness
function to
@(x) sum(abs(x))

To run the example,

Set Fitness function to @(x) sum(abs(x)).
Set Number of variables to 10.
Set Initial range to [-1; 1].
Select Best fitness and Distance in the Plot functions pane.
Run the example with the default value of 0.8 for Crossover fraction, in
the Options > Reproduction pane. This returns the best fitness value of
approximately 0.2 and displays the following plots.

5-67

Using the Genetic Algorithm

Best: 0.23492 Mean: 0.48445

Fitness value

10
8
6
4
2
0

10

20

10

20

30

40

50
60
70
80
Generation
Average Distance between individuals

30

40

90

100

90

100

50
60
generation

70

80

Crossover Without Mutation

To see how the genetic algorithm performs when there is no mutation, set
Crossover fraction to 1.0 and click Start. This returns the best fitness
value of approximately 1.3 and displays the following plots.

5-68

Genetic Algorithm Examples

Best: 1.3161 Mean: 1.3161

Fitness value

5
4
3
2
1

10

20

10

20

30

40

50
60
70
80
Generation
Average Distance between individuals

30

40

90

100

90

100

2.5
2
1.5
1
0.5
0

50
60
generation

70

80

In this case, the algorithm selects genes from the individuals in the initial
population and recombines them. The algorithm cannot create any new genes
because there is no mutation. The algorithm generates the best individual
that it can using these genes at generation number 8, where the best fitness
plot becomes level. After this, it creates new copies of the best individual,
which are then are selected for the next generation. By generation number 17,
all individuals in the population are the same, namely, the best individual.
When this occurs, the average distance between individuals is 0. Since the
algorithm cannot improve the best fitness value after generation 8, it stalls
after 50 more generations, because Stall generations is set to 50.

Mutation Without Crossover

To see how the genetic algorithm performs when there is no crossover, set
Crossover fraction to 0 and click Start. This returns the best fitness value
of approximately 3.5 and displays the following plots.

5-69

Using the Genetic Algorithm

Best: 3.493 Mean: 11.2376

Fitness value

25
20
15
10
5
0

10

20

10

20

30

40

50
60
70
80
Generation
Average Distance between individuals

30

40

90

100

90

100

14
12
10
8
6
4

50
60
generation

70

80

In this case, the random changes that the algorithm applies never improve the
fitness value of the best individual at the first generation. While it improves
the individual genes of other individuals, as you can see in the upper plot by
the decrease in the mean value of the fitness function, these improved genes
are never combined with the genes of the best individual because there is no
crossover. As a result, the best fitness plot is level and the algorithm stalls at
generation number 50.

Comparing Results for Varying Crossover Fractions

The demo deterministicstudy.m, which is included in the software,
compares the results of applying the genetic algorithm to Rastrigins function
with Crossover fraction set to 0, .2, .4, .6, .8, and 1. The demo runs for
10 generations. At each generation, the demo plots the means and standard
deviations of the best fitness values in all the preceding generations, for each
value of the Crossover fraction.

5-70

Genetic Algorithm Examples

To run the demo, enter

deterministicstudy

at the MATLAB prompt. When the demo is finished, the plots appear as in
the following figure.
After 10 iIterations

Iteration

2
4
6
8
10
0

0.2

0.4
0.6
CrossoverFraction

0.8

0.2

0.4
0.6
CrossoverFraction

0.8

Score Mean and Std

60

40

20

The lower plot shows the means and standard deviations of the best fitness
values over 10 generations, for each of the values of the crossover fraction.
The upper plot shows a color-coded display of the best fitness values in each
generation.
For this fitness function, setting Crossover fraction to 0.8 yields the
best result. However, for another fitness function, a different setting for
Crossover fraction might yield the best result.

5-71

Using the Genetic Algorithm

Example: Global vs. Local Minima with GA

Sometimes the goal of an optimization is to find the global minimum or
maximum of a functiona point where the function value is smaller or larger
at any other point in the search space. However, optimization algorithms
sometimes return a local minimuma point where the function value is
smaller than at nearby points, but possibly greater than at a distant point
in the search space. The genetic algorithm can sometimes overcome this
deficiency with the right settings.
As an example, consider the following function
2

x
exp

for x 20,
20
f ( x) =

exp(1) + ( x 20)( x 22) for x > 20.

The following figure shows a plot of the function.

3
2.5
2
1.5
1
0.5
0
0.5
1
1.5
10

10

15

20

25

The function has two local minima, one at x = 0, where the function value is
1, and the other at x = 21, where the function value is 1 1/e. Since the
latter value is smaller, the global minimum occurs at x = 21.

5-72

Genetic Algorithm Examples

Running the Genetic Algorithm on the Example

To run the genetic algorithm on this example,
1 Copy and paste the following code into a new file in the MATLAB Editor.

function y = two_min(x)
if x<=20
y = -exp(-(x/20).^2);
else
y = -exp(-1)+(x-20)*(x-22);
end
2 Save the file as two_min.m in a folder on the MATLAB path.
3 In the Optimization Tool,

Set Fitness function to @two_min.

Set Number of variables to 1.
Click Start.
The genetic algorithm returns a point very close to the local minimum at x = 0.

The following custom plot shows why the algorithm finds the local minimum
rather than the global minimum. The plot shows the range of individuals in
each generation and the best individual.

5-73

Using the Genetic Algorithm

Best: 0.00028487
2.5
2
1.5

Best individual

1
0.5
0
0.5
1
1.5
2

10

20

30

40

50
60
Generation

70

80

90

100

Note that all individuals are between -2 and 2.5. While this range is larger
than the default Initial range of [0;1], due to mutation, it is not large
enough to explore points near the global minimum at x = 21.
One way to make the genetic algorithm explore a wider range of pointsthat
is, to increase the diversity of the populationsis to increase the Initial
range. The Initial range does not have to include the point x = 21, but it
must be large enough so that the algorithm generates individuals near x = 21.
Set Initial range to [0;15] as shown in the following figure.

5-74

Genetic Algorithm Examples

Then click Start. The genetic algorithm returns a point very close to 21.

5-75

Using the Genetic Algorithm

This time, the custom plot shows a much wider range of individuals. By the
second generation there are individuals greater than 21, and by generation
12, the algorithm finds a best individual that is approximately equal to 21.
Best: 20.9876
80

60

Best individual

40

20

20

40

60

10

20

30

40

50
60
Generation

70

80

90

100

Using a Hybrid Function

A hybrid function is an optimization function that runs after the genetic
algorithm terminates in order to improve the value of the fitness function.
The hybrid function uses the final point from the genetic algorithm as its
initial point. You can specify a hybrid function in Hybrid function options.
This example uses Optimization Toolbox function fminunc, an unconstrained
minimization function. The example first runs the genetic algorithm to find a
point close to the optimal point and then uses that point as the initial point
for fminunc.
The example finds the minimum of Rosenbrocks function, which is defined by

f ( x1 , x2 ) = 100 x2 x12

5-76

+ (1 x1 )2 .

Genetic Algorithm Examples

The following figure shows a plot of Rosenbrocks function.

3000
2500
2000
1500
1000
500
0
3
2

2
1

1
1

Minimum at (1,1)

Global Optimization Toolbox software contains the dejong2fcn.m file, which

computes Rosenbrocks function. To see a demo of this example, enter
hybriddemo

at the MATLAB prompt.

To explore the example, first enter optimtool('ga') to open the Optimization
Tool to the ga solver. Enter the following settings:
Set Fitness function to @dejong2fcn.
Set Number of variables to 2.
Set Population size to 10.

5-77

Using the Genetic Algorithm

Before adding a hybrid function, click Start to run the genetic algorithm by
itself. The genetic algorithm displays the following results in the Run solver
and view results pane:

The final point is somewhat close to the true minimum at (1, 1). You can
improve this result by setting Hybrid function to fminunc (in the Hybrid
function options).

fminunc uses the final point of the genetic algorithm as its initial point.
It returns a more accurate result, as shown in the Run solver and view
results pane.

5-78

Genetic Algorithm Examples

You can set options for the hybrid function separately from the calling
function. Use optimset (or psoptimset for the patternsearch hybrid
function) to create the options structure. For example:
hybridopts = optimset('display','iter','LargeScale','off');

In the Optimization Tool enter the name of your options structure in the
Options box under Hybrid function:

At the command line, the syntax is as follows:

options = gaoptimset('HybridFcn',{@fminunc,hybridopts});
hybridopts must exist before you set options.

5-79

Using the Genetic Algorithm

Setting the Maximum Number of Generations

The Generations option in Stopping criteria determines the maximum
number of generations the genetic algorithm runs forsee Stopping
Conditions for the Algorithm on page 5-24. Increasing the Generations
option often improves the final result.
As an example, change the settings in the Optimization Tool as follows:
Set Fitness function to @rastriginsfcn.
Set Number of variables to 10.
Select Best fitness in the Plot functions pane.
Set Generations to Inf.
Set Stall generations to Inf.
Set Stall time to Inf.
Run the genetic algorithm for approximately 300 generations and click
Stop. The following figure shows the resulting best fitness plot after 300
generations.
Best: 5.0444 Mean: 48.7926
100
90
80

Fitness value

70
60
50
40
30
20
10
0

50

100

150
Generation

200

Genetic algorithm stalls

5-80

250

300

Genetic Algorithm Examples

Note that the algorithm stalls at approximately generation number 170that

is, there is no immediate improvement in the fitness function after generation
170. If you restore Stall generations to its default value of 50, the algorithm
would terminate at approximately generation number 230. If the genetic
algorithm stalls repeatedly with the current setting for Generations, you
can try increasing both the Generations and Stall generations options
to improve your results. However, changing other options might be more
effective.
Note When Mutation function is set to Gaussian, increasing the value
of Generations might actually worsen the final result. This can occur
because the Gaussian mutation function decreases the average amount of
mutation in each generation by a factor that depends on the value specified
in Generations. Consequently, the setting for Generations affects the
behavior of the algorithm.

Vectorizing the Fitness Function

The genetic algorithm usually runs faster if you vectorize the fitness function.
This means that the genetic algorithm only calls the fitness function once, but
expects the fitness function to compute the fitness for all individuals in the
current population at once. To vectorize the fitness function,
Write the file that computes the function so that it accepts a matrix with
arbitrarily many rows, corresponding to the individuals in the population.
For example, to vectorize the function

f ( x1 , x2 ) = x12 2 x1 x2 + 6 x1 + x22 6 x2
write the file using the following code:
z =x(:,1).^2 - 2*x(:,1).*x(:,2) + 6*x(:,1) + x(:,2).^2 - 6*x(:,2);

The colon in the first entry of x indicates all the rows of x, so that x(:, 1)
is a vector. The .^ and .* operators perform element-wise operations on
the vectors.
In the User function evaluation pane, set the Evaluate fitness and
constraint functions option to vectorized.

5-81

Using the Genetic Algorithm

Note The fitness function must accept an arbitrary number of rows to use
the Vectorize option.
The following comparison, run at the command line, shows the improvement
in speed with Vectorize set to On.
tic;ga(@rastriginsfcn,20);toc
elapsed_time =
4.3660
options=gaoptimset('Vectorize','on');
tic;ga(@rastriginsfcn,20,[],[],[],[],[],[],[],options);toc
elapsed_time =
0.5810

If there are nonlinear constraints, the objective function and the nonlinear
constraints all need to be vectorized in order for the algorithm to compute in
a vectorized manner.

Constrained Minimization Using ga

Suppose you want to minimize the simple fitness function of two variables
x1 and x2,

min f ( x) = 100 x12 x2

x

+ (1 x1 )2

subject to the following nonlinear inequality constraints and bounds

x1 x2 + x1 x2 + 1.5 0
10 x1 x2 0
0 x1 1
0 x2 13

5-82

(nonlinear constraiint)
(nonlinear constraint)
(bound)
(bound)

Genetic Algorithm Examples

Begin by creating the fitness and constraint functions. First, create a file
named simple_fitness.m as follows:
function y = simple_fitness(x)
y = 100*(x(1)^2 - x(2))^2 + (1 - x(1))^2;

The genetic algorithm function, ga, assumes the fitness function will take one
input x, where x has as many elements as the number of variables in the
problem. The fitness function computes the value of the function and returns
that scalar value in its one return argument, y.
Then create a file, simple_constraint.m, containing the constraints
function [c, ceq] = simple_constraint(x)
c = [1.5 + x(1)*x(2) + x(1) - x(2);...
-x(1)*x(2) + 10];
ceq = [];

The ga function assumes the constraint function will take one input x, where
x has as many elements as the number of variables in the problem. The
constraint function computes the values of all the inequality and equality
constraints and returns two vectors, c and ceq, respectively.
To minimize the fitness function, you need to pass a function handle to the
fitness function as the first argument to the ga function, as well as specifying
the number of variables as the second argument. Lower and upper bounds
are provided as LB and UB respectively. In addition, you also need to pass a
function handle to the nonlinear constraint function.
ObjectiveFunction = @simple_fitness;
nvars = 2;

% Number of variables

LB = [0 0];

% Lower bound

UB = [1 13];

% Upper bound

ConstraintFunction = @simple_constraint;
[x,fval] = ga(ObjectiveFunction,nvars,[],[],[],[],LB,UB,ConstraintFunction)
Optimization terminated: average change in the fitness value
less than options.TolFun and constraint violation is
less than options.TolCon.
x =

5-83

Using the Genetic Algorithm

0.8122

12.3122

fval =
1.3578e+004

The genetic algorithm solver handles linear constraints and bounds differently
from nonlinear constraints. All the linear constraints and bounds are satisfied
throughout the optimization. However, ga may not satisfy all the nonlinear
constraints at every generation. If ga converges to a solution, the nonlinear
constraints will be satisfied at that solution.
ga uses the mutation and crossover functions to produce new individuals at
every generation. ga satisfies linear and bound constraints by using mutation
and crossover functions that only generate feasible points. For example, in
the previous call to ga, the mutation function mutationguassian does not
necessarily obey the bound constraints. So when there are bound or linear
constraints, the default ga mutation function is mutationadaptfeasible.
If you provide a custom mutation function, this custom function must only
generate points that are feasible with respect to the linear and bound
constraints. All the included crossover functions generate points that satisfy
the linear constraints and bounds except the crossoverheuristic function.

To see the progress of the optimization, use the gaoptimset function to create
an options structure that selects two plot functions. The first plot function is
gaplotbestf, which plots the best and mean score of the population at every
generation. The second plot function is gaplotmaxconstr, which plots the
maximum constraint violation of nonlinear constraints at every generation.
You can also visualize the progress of the algorithm by displaying information
to the command window using the 'Display' option.
options = gaoptimset('PlotFcns',{@gaplotbestf,@gaplotmaxconstr},'Display','iter');

Rerun the ga solver.

[x,fval] = ga(ObjectiveFunction,nvars,[],[],[],[],...
LB,UB,ConstraintFunction,options)
Best
Generation

5-84

f-count

f(x)

max

Stall

constraint

Generations

849

14915.8

1567

13578.3

2334

13578.3

Genetic Algorithm Examples

3043

13578.3

3752

13578.3

Optimization terminated: average change in the fitness value

less than options.TolFun and constraint violation is
less than options.TolCon.
x =
0.8122

12.3123

fval =
1.3578e+004

You can provide a start point for the minimization to the ga function by
specifying the InitialPopulation option. The ga function will use the
first individual from InitialPopulation as a start point for a constrained
minimization.
X0 = [0.5 0.5]; % Start point (row vector)
options = gaoptimset(options,'InitialPopulation',X0);

Now, rerun the ga solver.

[x,fval] = ga(ObjectiveFunction,nvars,[],[],[],[],...
LB,UB,ConstraintFunction,options)

5-85

Using the Genetic Algorithm

Best
Generation

f-count

f(x)

max

Stall

constraint

Generations

965

13579.6

1728

13578.2

1.776e-015

2422

13578.2

Optimization terminated: average change in the fitness value

less than options.TolFun and constraint violation is
less than options.TolCon.
x =
0.8122

12.3122

fval =
1.3578e+004

Vectorized Constraints
If there are nonlinear constraints, the objective function and the nonlinear
constraints all need to be vectorized in order for the algorithm to compute in
a vectorized manner.
Vectorizing the Objective and Constraint Functions on page 4-72 contains
an example of how to vectorize both for the solver patternsearch. The syntax
is nearly identical for ga. The only difference is that patternsearch can have
its patterns appear as either row or column vectors; the corresponding vectors
for ga are the population vectors, which are always rows.

5-86

6
Using Simulated Annealing
What Is Simulated Annealing? on page 6-2
Performing a Simulated Annealing Optimization on page 6-3
Example: Minimizing De Jongs Fifth Function on page 6-7
Some Simulated Annealing Terminology on page 6-10
How Simulated Annealing Works on page 6-12
Using Simulated Annealing from the Command Line on page 6-14
Simulated Annealing Examples on page 6-19

Using Simulated Annealing

What Is Simulated Annealing?

Simulated annealing is a method for solving unconstrained and
bound-constrained optimization problems. The method models the physical
process of heating a material and then slowly lowering the temperature to
decrease defects, thus minimizing the system energy.
At each iteration of the simulated annealing algorithm, a new point is
randomly generated. The distance of the new point from the current point, or
the extent of the search, is based on a probability distribution with a scale
proportional to the temperature. The algorithm accepts all new points that
lower the objective, but also, with a certain probability, points that raise the
objective. By accepting points that raise the objective, the algorithm avoids
being trapped in local minima, and is able to explore globally for more possible
solutions. An annealing schedule is selected to systematically decrease the
temperature as the algorithm proceeds. As the temperature decreases, the
algorithm reduces the extent of its search to converge to a minimum.

6-2

Performing a Simulated Annealing Optimization

Performing a Simulated Annealing Optimization

Calling simulannealbnd at the Command Line
To call the simulated annealing function at the command line, use the syntax
[x fval] = simulannealbnd(@objfun,x0,lb,ub,options)

where
@objfun is a function handle to the objective function.
x0 is an initial guess for the optimizer.
lb and ub are lower and upper bound constraints, respectively, on x.
options is a structure containing options for the algorithm. If you do not
pass in this argument, simulannealbnd uses its default options.
The results are given by:
x Final point returned by the solver
fval Value of the objective function at x
The command-line function simulannealbnd is convenient if you want to
Return results directly to the MATLAB workspace.
Run the simulated annealing algorithm multiple times with different
options by calling simulannealbnd from a file.
Using Simulated Annealing from the Command Line on page 6-14 provides
a detailed description of using the function simulannealbnd and creating
the options structure.

Using the Optimization Tool

To open the Optimization Tool, enter
optimtool('simulannealbnd')

6-3

Using Simulated Annealing

at the command line, or enter optimtool and then choose simulannealbnd

from the Solver menu.
Set options

Expand or contract help

Choose solver

Enter problem
and constraints

Run solver

View results

See final point

You can also start the tool from the MATLAB Start menu as pictured:

6-4

Performing a Simulated Annealing Optimization

To use the Optimization Tool, you must first enter the following information:
Objective function The objective function you want to minimize. Enter
the fitness function in the form @fitnessfun, where fitnessfun.m is a file
that computes the objective function. Computing Objective Functions on
page 2-2 explains how write this file. The @ sign creates a function handle
to fitnessfun.

6-5

Using Simulated Annealing

Number of variables The length of the input vector to the fitness

function. For the function my_fun described in Computing Objective
Functions on page 2-2, you would enter 2.
You can enter bounds for the problem in the Constraints pane. If the
problem is unconstrained, leave these fields blank.
To run the simulated annealing algorithm, click the Start button. The tool
displays the results of the optimization in the Run solver and view results
pane.
You can change the options for the simulated annealing algorithm in the
Options pane. To view the options in one of the categories listed in the pane,
click the + sign next to it.
For more information,
See the Optimization Tool chapter in the Optimization Toolbox
documentation.
See Minimizing Using the Optimization Tool on page 6-8 for an example
of using the tool with the function simulannealbnd.

6-6

Example: Minimizing De Jongs Fifth Function

Example: Minimizing De Jongs Fifth Function

In this section...
Description on page 6-7
Minimizing at the Command Line on page 6-8
Minimizing Using the Optimization Tool on page 6-8

Description
This section presents an example that shows how to find the minimum of the
function using simulated annealing.
De Jongs fifth function is a two-dimensional function with many (25) local
minima:
dejong5fcn

6-7

Using Simulated Annealing

Many standard optimization algorithms get stuck in local minima. Because

the simulated annealing algorithm performs a wide random search, the
chance of being trapped in local minima is decreased.
Note Because simulated annealing uses random number generators, each
time you run this algorithm you can get different results. See Reproducing
Your Results on page 6-17 for more information.

Minimizing at the Command Line

To run the simulated annealing algorithm without constraints, call
simulannealbnd at the command line using the objective function in
dejong5fcn.m, referenced by anonymous function pointer:
fun = @dejong5fcn;
[x fval] = simulannealbnd(fun, [0 0])

This returns
x =
-31.9779
fval =
0.9980

-31.9595

where
x is the final point returned by the algorithm.
fval is the objective function value at the final point.

Minimizing Using the Optimization Tool

To run the minimization using the Optimization Tool,
1 Set up your problem as pictured in the Optimization Tool

6-8

Example: Minimizing De Jongs Fifth Function

2 Click Start under Run solver and view results:

6-9

Using Simulated Annealing

Some Simulated Annealing Terminology

In this section...
Objective Function on page 6-10
Temperature on page 6-10
Annealing Schedule on page 6-10
Reannealing on page 6-10

Objective Function
The objective function is the function you want to optimize. Global
Optimization Toolbox algorithms attempt to find the minimum of the objective
function. Write the objective function as a file or anonymous function, and
pass it to the solver as a function handle.

Temperature
The temperature is the control parameter in simulated annealing that is
decreased gradually as the algorithm proceeds. It determines the probability
of accepting a worse solution at any step and is used to limit the extent of the
search in a given dimension. You can specify the initial temperature as an
integer in the InitialTemperature option, and the annealing schedule as a
function in the TemperatureFcn option.

Annealing Schedule
The annealing schedule is the rate by which the temperature is decreased
as the algorithm proceeds. The slower the rate of decrease, the better the
chances are of finding an optimal solution, but the longer the run time.
You can specify the temperature schedule as a function handle with the
TemperatureFcn option.

Reannealing
Annealing is the technique of closely controlling the temperature when cooling
a material to ensure that it is brought to an optimal state. Reannealing raises
the temperature after a certain number of new points have been accepted,

6-10

Some Simulated Annealing Terminology

and starts the search again at the higher temperature. Reannealing avoids
getting caught at local minima. You specify the reannealing schedule with the
ReannealInterval option.

6-11

Using Simulated Annealing

How Simulated Annealing Works

In this section...
Outline of the Algorithm on page 6-12
Stopping Conditions for the Algorithm on page 6-12

Outline of the Algorithm

The following is an outline of the steps performed for the simulated annealing
algorithm:
1 The algorithm begins by randomly generating a new point. The distance

of the new point from the current point, or the extent of the search, is
determined by a probability distribution with a scale proportional to the
current temperature.
2 The algorithm determines whether the new point is better or worse than

the current point. If the new point is better than the current point, it
becomes the next point. If the new point is worse than the current point,
the algorithm may still make it the next point. The algorithm accepts a
worse point based on an acceptance probability.
3 The algorithm systematically lowers the temperature, storing the best

point found so far.

4 Reannealing is performed after a certain number of points

(ReannealInterval) are accepted by the solver. Reannealing raises the

temperature in each dimension, depending on sensitivity information. The
search is resumed with the new temperature values.
5 The algorithm stops when the average change in the objective function is

very small, or when any other stopping criterion is met. See Stopping
Conditions for the Algorithm on page 6-12.

Stopping Conditions for the Algorithm

The simulated annealing algorithm uses the following conditions to determine
when to stop:

6-12

How Simulated Annealing Works

TolFun The algorithm runs until the average change in value of the
objective function in StallIterLim iterations is less than TolFun. The
default value is 1e-6.
MaxIter The algorithm stops if the number of iterations exceeds this
maximum number of iterations. You can specify the maximum number of
iterations as a positive integer or Inf. Inf is the default.
MaxFunEval specifies the maximum number of evaluations of the objective
function. The algorithm stops if the number of function evaluations exceeds
the maximum number of function evaluations. The default maximum is
3000*numberofvariables.
TimeLimit specifies the maximum time in seconds the algorithm runs
before stopping.
ObjectiveLimit The algorithm stops if the best objective function value
is less than or equal to the value of ObjectiveLimit.

6-13

Using Simulated Annealing

Using Simulated Annealing from the Command Line

In this section...
Running simulannealbnd With the Default Options on page 6-14
Setting Options for simulannealbnd at the Command Line on page 6-15
Reproducing Your Results on page 6-17

Running simulannealbnd With the Default Options

To run the simulated annealing algorithm with the default options, call
simulannealbnd with the syntax
[x,fval] = simulannealbnd(@objfun,x0)

The input arguments to simulannealbnd are

@objfun A function handle to the file that computes the objective
function. Computing Objective Functions on page 2-2 explains how to
write this file.
x0 The initial guess of the optimal argument to the objective function.
The output arguments are
x The final point.
fval The value of the objective function at x.
For a description of additional input and output arguments, see the reference
pages for simulannealbnd.
You can run the example described in Example: Minimizing De Jongs Fifth
Function on page 6-7 from the command line by entering
[x,fval] = simulannealbnd(@dejong5fcn, [0 0])

This returns
x =
-31.9564

6-14

-15.9755

Using Simulated Annealing from the Command Line

fval =
5.9288

Additional Output Arguments

To get more information about the performance of the algorithm, you can
call simulannealbnd with the syntax
[x,fval,exitflag,output] = simulannealbnd(@objfun,x0)

Besides x and fval, this function returns the following additional output
arguments:
exitflag Flag indicating the reason the algorithm terminated
output Structure containing information about the performance of the
algorithm
See the simulannealbnd reference pages for more information about these
arguments.

Setting Options for simulannealbnd at the Command

Line
You can specify options by passing an options structure as an input argument
to simulannealbnd using the syntax
[x,fval] = simulannealbnd(@objfun,x0,[],[],options)

This syntax does not specify any lower or upper bound constraints.
You create the options structure using the saoptimset function:
options = saoptimset('simulannealbnd')

This returns the structure options with the default values for its fields:
options =
AnnealingFcn:
TemperatureFcn:
AcceptanceFcn:
TolFun:

@annealingfast
@temperatureexp
@acceptancesa
1.0000e-006

6-15

Using Simulated Annealing

StallIterLimit:
MaxFunEvals:
TimeLimit:
MaxIter:
ObjectiveLimit:
Display:
DisplayInterval:
HybridFcn:
HybridInterval:
PlotFcns:
PlotInterval:
OutputFcns:
InitialTemperature:
ReannealInterval:
DataType:

'500*numberofvariables'
'3000*numberofvariables'
Inf
Inf
-Inf
'final'
10
[]
'end'
[]
1
[]
100
100
'double'

The value of each option is stored in a field of the options structure, such as
options.ReannealInterval. You can display any of these values by entering
options followed by the name of the field. For example, to display the interval
for reannealing used for the simulated annealing algorithm, enter
options.ReannealInterval
ans =
100

To create an options structure with a field value that is different from the
defaultfor example, to set ReannealInterval to 300 instead of its default
value 100enter
options = saoptimset('ReannealInterval', 300)

This creates the options structure with all values set to their defaults, except
for ReannealInterval, which is set to 300.
If you now enter
simulannealbnd(@objfun,x0,[],[],options)
simulannealbnd runs the simulated annealing algorithm with a reannealing

interval of 300.

6-16

Using Simulated Annealing from the Command Line

If you subsequently decide to change another field in the options structure,

such as setting PlotFcns to @saplotbestf, which plots the best objective
function value at each iteration, call saoptimset with the syntax
options = saoptimset(options,'PlotFcns',@saplotbestf)

This preserves the current values of all fields of options except for PlotFcns,
which is changed to @saplotbestf. Note that if you omit the input argument
options, saoptimset resets ReannealInterval to its default value 100.
You can also set both ReannealInterval and PlotFcns with the single
command
options = saoptimset('ReannealInterval',300, ...
'PlotFcns',@saplotbestf)

Reproducing Your Results

Because the simulated annealing algorithm is stochasticthat is, it makes
random choicesyou get slightly different results each time you run it. The
algorithm uses the default MATLAB pseudorandom number stream. For
more information about random number streams, see RandStream. Each
time the algorithm calls the stream, its state changes. So the next time the
algorithm calls the stream, it returns a different random number.
If you need to reproduce your results exactly, call simulannealbnd with
the output argument. The output structure contains the current random
number generator state in the output.rngstate field. Reset the state before
running the function again.
For example, to reproduce the output of simulannealbnd applied to De Jongs
fifth function, call simulannealbnd with the syntax
[x,fval,exitflag,output] = simulannealbnd(@dejong5fcn,[0 0]);

Suppose the results are

x =
31.9361

-31.9457

fval =
4.9505

6-17

Using Simulated Annealing

The state of the random number generator, rngstate, is stored in output:

output =
iterations: 1754
funccount: 1769
message: 'Optimization terminated:...
change in best function value less than options.TolFun.'
rngstate: [1x1 struct]
problemtype: 'unconstrained'
temperature: [2x1 double]
totaltime: 1.1094

Reset the stream by entering

stream = RandStream.getDefaultStream;
stream.State = output.rngstate.state;

If you now run simulannealbnd a second time, you get the same results.
You can reproduce your run in the Optimization Tool by checking the box Use
random states from previous run in the Run solver and view results
section.

Note If you do not need to reproduce your results, it is better not to set the
states of RandStream, so that you get the benefit of the randomness in these
algorithms.

6-18

Simulated Annealing Examples

Simulated Annealing Examples

If you are viewing this documentation in the Help browser, click the following
link to see the demo Minimization Using Simulated Annealing. Or, from the
MATLAB command line, type echodemo('saobjective').

6-19

6-20

Using Simulated Annealing

7
Multiobjective Optimization
What Is Multiobjective Optimization? on page 7-2
Using gamultiobj on page 7-5
References on page 7-14

Multiobjective Optimization

What Is Multiobjective Optimization?

You might need to formulate problems with more than one objective, since a
single objective with several constraints may not adequately represent the
problem being faced. If so, there is a vector of objectives,
F(x) = [F1(x), F2(x),...,Fm(x)],
that must be traded off in some way. The relative importance of these
objectives is not generally known until the systems best capabilities are
determined and tradeoffs between the objectives fully understood. As the
number of objectives increases, tradeoffs are likely to become complex and
less easily quantified. The designer must rely on his or her intuition and
ability to express preferences throughout the optimization cycle. Thus,
requirements for a multiobjective design strategy must enable a natural
problem formulation to be expressed, and be able to solve the problem and
enter preferences into a numerically tractable and realistic design problem.
Multiobjective optimization is concerned with the minimization of a vector of
objectives F(x) that can be the subject of a number of constraints or bounds:

minn F ( x), subject to

xR

Gi ( x) = 0, i = 1,..., ke ; Gi ( x) 0, i = ke + 1,..., k; l x u.
Note that because F(x) is a vector, if any of the components of F(x) are
competing, there is no unique solution to this problem. Instead, the concept
of noninferiority [4] (also called Pareto optimality [1] and [2]) must be used
to characterize the objectives. A noninferior solution is one in which an
improvement in one objective requires a degradation of another. To define
this concept more precisely, consider a feasible region, , in the parameter
space. x is an element of the n-dimensional real numbers x R n that satisfies
all the constraints, i.e.,

= xRn ,
subject to

7-2

What Is Multiobjective Optimization?

Gi ( x) = 0, i = 1,..., ke ,
Gi ( x) 0, i = ke + 1,..., k,
l x u.
This allows definition of the corresponding feasible region for the objective
function space :

= y R m : y = F ( x), x .
The performance vector F(x) maps parameter space into objective function
space, as represented in two dimensions in the figure Mapping from
Parameter Space into Objective Function Space on page 7-3.

Figure 7-1:

Mapping from Parameter Space into Objective Function Space

A noninferior solution point can now be defined.

Definition: Point x* is a noninferior solution if for some neighborhood of
x* there does not exist a x such that ( x * + x ) and

Fi ( x * + x ) Fi ( x*), i = 1,..., m, and

F j ( x * + x ) < F j ( x*) for at least one j.
In the two-dimensional representation of the figure Set of Noninferior
Solutions on page 7-4, the set of noninferior solutions lies on the curve
between C and D. Points A and B represent specific noninferior points.

7-3

Multiobjective Optimization

Figure 7-2:

Set of Noninferior Solutions

A and B are clearly noninferior solution points because an improvement

in one objective, F1, requires a degradation in the other objective, F2, i.e.,
F1B < F1A, F2B > F2A.
Since any point in that is an inferior point represents a point in which
improvement can be attained in all the objectives, it is clear that such a point
is of no value. Multiobjective optimization is, therefore, concerned with the
generation and selection of noninferior solution points.
Noninferior solutions are also called Pareto optima. A general goal in
multiobjective optimization is constructing the Pareto optima. The algorithm
used in gamultiobj is described in [3].

7-4

Using gamultiobj

Using gamultiobj
In this section...
Problem Formulation on page 7-5
Using gamultiobj with Optimization Tool on page 7-6
Example: Multiobjective Optimization on page 7-7
Options and Syntax: Differences With ga on page 7-13

Problem Formulation
The gamultiobj solver attempts to create a set of Pareto optima for a
multiobjective minimization. You may optionally set bounds and linear
constraints on variables. gamultiobj uses the genetic algorithm for finding
local Pareto optima. As in the ga function, you may specify an initial
population, or have the solver generate one automatically.
The fitness function for use in gamultiobj should return a vector of type
double. The population may be of type double, a bit string vector, or can be
a custom-typed vector. As in ga, if you use a custom population type, you
must write your own creation, mutation, and crossover functions that accept
inputs of that population type, and specify these functions in the following
fields, respectively:
Creation function (CreationFcn)
Mutation function (MutationFcn)
Crossover function (CrossoverFcn)
You can set the initial population in a variety of ways. Suppose that you
choose a population of size m. (The default population size is 15 times the
number of variables n.) You can set the population:
As an m-by-n matrix, where the rows represent m individuals.
As a k-by-n matrix, where k < m. The remaining m k individuals are
generated by a creation function.
The entire population can be created by a creation function.

7-5

Multiobjective Optimization

Using gamultiobj with Optimization Tool

You can access gamultiobj from the Optimization Tool GUI. Enter
optimtool('gamultiobj')

at the command line, or enter optimtool and then choose gamultiobj from
the Solver menu. You can also start the tool from the MATLAB Start menu
as pictured:

7-6

Using gamultiobj

If the Quick Reference help pane is closed, you can open it by clicking
the >> button on the upper right of the GUI:
. All the options
available are explained briefly in the help pane.
You can create an options structure in the Optimization Tool, export it to the
MATLAB workspace, and use the structure at the command line. For details,
see Importing and Exporting Your Work in the Optimization Toolbox
documentation.

Example: Multiobjective Optimization

This example has a two-objective fitness function f(x), where x is also
two-dimensional:
function f = mymulti1(x)
f(1) = x(1)^4 - 10*x(1)^2+x(1)*x(2) + x(2)^4 -(x(1)^2)*(x(2)^2);
f(2) = x(2)^4 - (x(1)^2)*(x(2)^2) + x(1)^4 + x(1)*x(2);

Create this function file before proceeding.

Performing the Optimization with Optimization Tool

1 To define the optimization problem, start the Optimization Tool, and set it

as pictured.

7-7

Multiobjective Optimization

2 Set the options for the problem as pictured.

3 Run the optimization by clicking Start under Run solver and view

results.
A plot appears in a figure window.

7-8

Using gamultiobj

This plot shows the tradeoff between the two components of f. It is plotted
in objective function space; see the figure Set of Noninferior Solutions on
page 7-4.
The results of the optimization appear in the following table containing both
objective function values and the value of the variables.

7-9

Multiobjective Optimization

You can sort the table by clicking a heading. Click the heading again to sort
it in the reverse order. The following figures show the result of clicking the
heading f1.

7-10

Using gamultiobj

Performing the Optimization at the Command Line

To perform the same optimization at the command line:
1 Set the options:

options = gaoptimset('PopulationSize',60,...
'ParetoFraction',0.7,'PlotFcn',@gaplotpareto);
2 Run the optimization using the options:

[x fval flag output population] = gamultiobj(@mymulti1,2,...

[],[],[],[],[-5,-5],[5,5],options);

Alternate Views
There are other ways of regarding the problem. The following figure contains
a plot of the level curves of the two objective functions, the Pareto frontier
calculated by gamultiobj (boxes), and the x-values of the true Pareto frontier
(diamonds connected by a nearly-straight line). The true Pareto frontier
points are where the level curves of the objective functions are parallel. They

7-11

Multiobjective Optimization

were calculated by finding where the gradients of the objective functions are
parallel. The figure is plotted in parameter space; see the figure Mapping
from Parameter Space into Objective Function Space on page 7-3.

= Frontier points calculated by

gamultiobj
= True Pareto frontier

Contours of objective functions, and Pareto frontier

7-12

Using gamultiobj

gamultiobj found the ends of the line segment, meaning it found the full

extent of the Pareto frontier.

Options and Syntax: Differences With ga

The syntax and options for gamultiobj are similar to those for ga, with the
following differences:
gamultiobj does not have nonlinear constraints, so its syntax has fewer
inputs.
gamultiobj takes an option DistanceMeasureFcn, a function that assigns
a distance measure to each individual with respect to its neighbors.
gamultiobj takes an option ParetoFraction, a number between 0 and 1
that specifies the fraction of the population on the best Pareto frontier to
be kept during the optimization. If there is only one Pareto frontier, this
option is ignored.
gamultiobj uses only the Tournament selection function.
gamultiobj uses elite individuals differently than ga. It sorts noninferior
individuals above inferior ones, so it uses elite individuals automatically.
gamultiobj has only one hybrid function, fgoalattain.
gamultiobj does not have a stall time limit.
gamultiobj has different plot functions available.
gamultiobj does not have a choice of scaling function.

7-13

Multiobjective Optimization

References
[1] Censor, Y., Pareto Optimality in Multiobjective Problems, Appl. Math.
Optimiz., Vol. 4, pp 4159, 1977.
[2] Da Cunha, N.O. and E. Polak, Constrained Minimization Under
Vector-Valued Criteria in Finite Dimensional Spaces, J. Math. Anal. Appl.,
Vol. 19, pp 103124, 1967.
[3] Deb, Kalyanmoy, Multi-Objective Optimization using Evolutionary
Algorithms, John Wiley & Sons, Ltd, Chichester, England, 2001.
[4] Zadeh, L.A., Optimality and Nonscalar-Valued Performance Criteria,
IEEE Trans. Automat. Contr., Vol. AC-8, p. 1, 1963.

7-14

8
Parallel Processing
Background on page 8-2
How to Use Parallel Processing on page 8-11

Parallel Processing

Background
Types of Parallel Processing in Global Optimization
Toolbox
Parallel processing is an attractive way to speed optimization algorithms. To
use parallel processing, you must have a Parallel Computing Toolbox license,
and have a MATLAB worker pool open (matlabpool). For more information,
see How to Use Parallel Processing on page 8-11.
Global Optimization Toolbox solvers use parallel computing in various ways.
Solver

Parallel?

Parallel Characteristics

GlobalSearch

No parallel functionality. However, fmincon can use parallel

gradient estimation when run in GlobalSearch. See Using
Parallel Computing with fmincon, fgoalattain, and fminimax
in the Optimization Toolbox documentation.
Start points distributed to multiple processors. From these
points, local solvers run to completion. For more details, see
MultiStart on page 8-5 and How to Use Parallel Processing
on page 8-11.

MultiStart

For fmincon, no parallel gradient estimation with parallel

MultiStart.
ga, gamultiobj

Population evaluated in parallel, which occurs once per

iteration. For more details, see Genetic Algorithm on page 8-8
and How to Use Parallel Processing on page 8-11.
No vectorization of fitness or constraint functions.
Poll points evaluated in parallel, which occurs once per
iteration. For more details, see Pattern Search on page 8-7
and How to Use Parallel Processing on page 8-11.

patternsearch

No vectorization of fitness or constraint functions.

simulannealbnd

8-2

No parallel functionality. However, simulannealbnd can

use a hybrid function that runs in parallel. See Simulated
Annealing on page 8-10.

Background

In addition, several solvers have hybrid functions that run after they finish.
Some hybrid functions can run in parallel. Also, most patternsearch search
methods can run in parallel. For more information, see Parallel Search
Functions or Hybrid Functions on page 8-14.

How Toolbox Functions Distribute Processes

parfor Characteristics and Caveats on page 8-3
MultiStart on page 8-5
GlobalSearch on page 8-6
Pattern Search on page 8-7
Genetic Algorithm on page 8-8
Parallel Computing with gamultiobj on page 8-9
Simulated Annealing on page 8-10

parfor Characteristics and Caveats

No Nested parfor Loops. Solvers employ the Parallel Computing Toolbox
parfor function to perform parallel computations.
Note parfor does not work in parallel when called from within another
parfor loop.
Suppose, for example, your objective function userfcn calls parfor, and you
want to call fmincon using MultiStart and parallel processing. Suppose also
that the conditions for parallel gradient evaluation of fmincon are satisfied,
as given in Parallel Optimization Functionality. The figure When parfor
Runs In Parallel on page 8-4 shows three cases:
1 The outermost loop is parallel MultiStart. Only that loop runs in parallel.
2 The outermost parfor loop is in fmincon. Only fmincon runs in parallel.

8-3

Parallel Processing

3 The outermost parfor loop is in userfcn. In this case, userfcn can use

parfor in parallel.

Bold indicates the function that runs in parallel

1

...
problem = createOptimProblem(fmincon,'objective',@userfcn,...)
ms = MultiStart('UseParallel','always');
x = run(ms,problem,10)
Only the outermost parfor loop
...
runs in parallel
If fmincon UseParallel option = 'always'
fmincon estimates gradients in parallel

...
x = fmincon(@userfcn,...)
...
If fmincon UseParallel option = 'never'
userfcn can use parfor in parallel
...
x = fmincon(@userfcn,...)
...

When parfor Runs In Parallel

Parallel Random Numbers Are Not Reproducible. Random number

sequences in MATLAB are pseudorandom, determined from a seed, or an
initial setting. Parallel computations use seeds that are not necessarily
controllable or reproducible. For example, each instance of MATLAB has a
default global setting that determines the current seed for random sequences.
For patternsearch, if you select MADS as a poll or search method, parallel
pattern search does not have reproducible runs. If you select the genetic
algorithm or Latin hypercube as search methods, parallel pattern search does
not have reproducible runs.
For ga and gamultiobj, parallel population generation gives nonreproducible
results.
MultiStart is different. You can have reproducible runs from parallel
MultiStart. Runs are reproducible because MultiStart generates

pseudorandom start points locally, and then distributes the start points to
parallel processors. Therefore, the parallel processors do not use random

8-4

Background

numbers. For more details, see Parallel Processing and Random Number
Streams on page 3-61.
Limitations and Performance Considerations. More caveats related to
parfor appear in the Limitations section of the Parallel Computing Toolbox
documentation.
For information on factors that affect the speed of parallel computations,
and factors that affect the results of parallel computations, see Improving
Performance with Parallel Computing in the Optimization Toolbox
documentation. The same considerations apply to parallel computing with
Global Optimization Toolbox functions.

MultiStart
MultiStart can automatically distribute a problem and start points
to multiple processes or processors. The problems run independently,
and MultiStart combines the distinct local minima into a vector of
GlobalOptimSolution objects. MultiStart uses parallel computing when
you:

Have a license for Parallel Computing Toolbox software.

Enable parallel computing with matlabpool, a Parallel Computing Toolbox
function.
Set the UseParallel property to 'always' in the MultiStart object:
ms = MultiStart('UseParallel','always');

When these conditions hold, MultiStart distributes a problem and start

points to processes or processors one at a time. The algorithm halts when it
reaches a stopping condition or runs out of start points to distribute. If the
MultiStart Display property is 'iter', then MultiStart displays:
Running the local solvers in parallel.

For an example of parallel MultiStart, see Example: Parallel MultiStart

on page 3-74.

8-5

Parallel Processing

Implementation Issues in Parallel MultiStart. fmincon cannot estimate

gradients in parallel when used with parallel MultiStart. This lack of
parallel gradient estimation is due to the limitation of parfor described in
No Nested parfor Loops on page 8-3.
fmincon can take longer to estimate gradients in parallel rather than in
serial. In this case, using MultiStart with parallel gradient estimation in
fmincon amplifies the slowdown. For example, suppose the ms MultiStart
object has UseParallel set to 'never'. Suppose fmincon takes 1 s longer
to solve problem with problem.options.UseParallel set to 'always'.
Then run(ms,problem,200) takes 200 s longer than the same run with
problem.options.UseParallel set to 'never'

Note When executing serially, parfor loops run slower than for loops.
Therefore, for best performance, set your local solver UseParallel option to
'never' when the MultiStart UseParallel property is 'always'.

GlobalSearch
GlobalSearch does not distribute a problem and start points to multiple
processes or processors. However, when GlobalSearch runs the fmincon local
solver, fmincon can estimate gradients by parallel finite differences. fmincon

uses parallel computing when you:

Have a license for Parallel Computing Toolbox software.
Enable parallel computing with matlabpool, a Parallel Computing Toolbox
function.
Set the UseParallel option to 'always' with optimset. Set this option in
the problem structure:
opts = optimset('UseParallel','always','Algorithm','sqp');
problem = createOptimProblem('fmincon','objective',@myobj,...
'x0',startpt,'options',opts);

For more details, see Using Parallel Computing with fmincon, fgoalattain,
and fminimax in the Optimization Toolbox documentation.

8-6

Background

Pattern Search
patternsearch can automatically distribute the evaluation of objective and
constraint functions associated with the points in a pattern to multiple
processes or processors. patternsearch uses parallel computing when you:

Have a license for Parallel Computing Toolbox software.

Enable parallel computing with matlabpool, a Parallel Computing Toolbox
function.
Set the following options using psoptimset or the Optimization Tool:

CompletePoll is 'on'.
Vectorized is 'off' (default).
UseParallel is 'always'.

When these conditions hold, the solver computes the objective function and
constraint values of the pattern search in parallel during a poll. Furthermore,
patternsearch overrides the setting of the Cache option, and uses the default
'off' setting.
Parallel Search Function. patternsearch can optionally call a search
function at each iteration. The search is parallel when you:
Set CompleteSearch to 'on'.
Do not set the search method to @searchneldermead or custom.
Set the search method to a patternsearch poll method or Latin hypercube
search, and set UseParallel to 'always'.
Or, if you set the search method to ga, create a search method option
structure with UseParallel set to 'always'.
Implementation Issues in Parallel Pattern Search. The limitations
on patternsearch options, listed in Pattern Search on page 8-7, arise
partly from the limitations of parfor, and partly from the nature of parallel
processing:
Cache is overridden to be 'off' patternsearch implements Cache as a
persistent variable. parfor does not handle persistent variables, because
the variable could have different settings at different processors.

8-7

Parallel Processing

CompletePoll is 'on' CompletePoll determines whether a poll stops as

soon as patternsearch finds a better point. When searching in parallel,
parfor schedules all evaluations simultaneously, and patternsearch
continues after all evaluations complete. patternsearch cannot halt
evaluations after they start.
Vectorized is 'off' Vectorized determines whether patternsearch
evaluates all points in a pattern with one function call in a vectorized
fashion. If Vectorized is 'on', patternsearch does not distribute the
evaluation of the function, so does not use parfor.

Genetic Algorithm
ga and gamultiobj can automatically distribute the evaluation of objective
and nonlinear constraint functions associated with a population to multiple
processors. ga uses parallel computing when you:

Have a license for Parallel Computing Toolbox software.

Enable parallel computing with matlabpool, a Parallel Computing Toolbox
function.
Set the following options using gaoptimset or the Optimization Tool:

Vectorized is 'off' (default).

UseParallel is 'always'.

When these conditions hold, the solver computes the objective function and
nonlinear constraint values of the individuals in a population in parallel.
Implementation Issues in Parallel Genetic Algorithm. The limitations
on options, listed in Genetic Algorithm on page 8-8, arise partly from
limitations of parfor, and partly from the nature of parallel processing:
Vectorized is 'off' Vectorized determines whether ga evaluates
an entire population with one function call in a vectorized fashion. If
Vectorized is 'on', ga does not distribute the evaluation of the function,
so does not use parfor.
ga can have a hybrid function that runs after it finishes; see Using a Hybrid

Function on page 5-76. If you want the hybrid function to take advantage
of parallel computation, set its options separately so that UseParallel is

8-8

Background

'always'. If the hybrid function is patternsearch, set CompletePoll to 'on'

so that patternsearch runs in parallel.

If the hybrid function is fmincon, set the following options with optimset to
have parallel gradient estimation:
GradObj must not be 'on' it can be 'off' or [].
Or, if there is a nonlinear constraint function, GradConstr must not be
'on' it can be 'off' or [].
To find out how to write options for the hybrid function, see Parallel Hybrid
Functions on page 8-15.

Parallel Computing with gamultiobj

Parallel computing with gamultiobj works almost the same as with ga. For
detailed information, see Genetic Algorithm on page 8-8.
The difference between parallel computing with gamultiobj and ga has to
do with the hybrid function. gamultiobj allows only one hybrid function,
fgoalattain. This function optionally runs after gamultiobj finishes its run.
Each individual in the calculated Pareto frontier, that is, the final population
found by gamultiobj, becomes the starting point for an optimization using
fgoalattain. These optimizations run in parallel. The number of processors
performing these optimizations is the smaller of the number of individuals
and the size of your matlabpool.
For fgoalattain to run in parallel, set its options correctly:
fgoalopts = optimset('UseParallel','always')
gaoptions = gaoptimset('HybridFcn',{@fgoalattain,fgoalopts});

Run gamultiobj with gaoptions, and fgoalattain runs in parallel. For

more information about setting the hybrid function, see Hybrid Function
Options on page 9-47.
gamultiobj calls fgoalattain using a parfor loop, so fgoalattain does

not estimate gradients in parallel when used as a hybrid function with

gamultiobj. For more information, see No Nested parfor Loops on page 8-3.

8-9

Parallel Processing

Simulated Annealing
simulannealbnd does not run in parallel automatically. However, it can
call hybrid functions that take advantage of parallel computing. To find out
how to write options for the hybrid function, see Parallel Hybrid Functions
on page 8-15.

8-10

How to Use Parallel Processing

How to Use Parallel Processing

In this section...
Multicore Processors on page 8-11
Processor Network on page 8-12
Parallel Search Functions or Hybrid Functions on page 8-14

Multicore Processors
If you have a multicore processor, you might see speedup using parallel
processing. You can establish a matlabpool of several parallel workers with a
Parallel Computing Toolbox license. For a description of Parallel Computing
Toolbox software, and the maximum number of parallel workers, see Product
Overview.
Suppose you have a dual-core processor, and want to use parallel computing:
Enter
matlabpool open 2

at the command line. The 2 specifies the number of MATLAB processes

to start.
Set your solver to use parallel processing.
MultiStart

Patternsearch

GA

ms =
MultiStart('UseParallel',
'always');

options =
psoptimset('UseParallel',
'always', 'CompletePoll',
'on', 'Vectorized', 'off');

options =
gaoptimset('UseParallel',
'always', 'Vectorized',
'off');

For Optimization Tool:

For Optimization Tool:

Options > User function

evaluation > Evaluate

Options > User function

evaluation > Evaluate

or
ms.UseParallel = 'always'

8-11

Parallel Processing

MultiStart

Patternsearch
objective and constraint
functions > in parallel

GA
fitness and constraint
functions > in parallel

Options > Complete poll

> on

When you run an applicable solver with options, applicable solvers

automatically use parallel computing.
To stop computing optimizations in parallel, set UseParallel to 'never',
or set the Optimization Tool not to compute in parallel. To halt all parallel
computation, enter
matlabpool close

Processor Network
If you have multiple processors on a network, use Parallel Computing Toolbox
functions and MATLAB Distributed Computing Server software to
establish parallel computation. Here are the steps to take:
1 Make sure your system is configured properly for parallel computing.

Check with your systems administrator, or refer to the Parallel Computing

Toolbox documentation, or the Administrator Guide documentation for
MATLAB Distributed Computing Server.
To perform a basic check:
a At the command line, enter

matlabpool open conf

or
matlabpool open conf n

where conf is your configuration, and n is the number of processors

you want to use.
b If network_file_path is the network path to your objective or constraint

functions, enter

8-12

How to Use Parallel Processing

pctRunOnAll('addpath network_file_path')

so the worker processors can access your objective or constraint files.

c Check whether a file is on the path of every worker by entering

pctRunOnAll('which filename')

If any worker does not have a path to the file, it reports

filename not found.

2 Set your solver to use parallel processing.

MultiStart

Patternsearch

GA

ms =
MultiStart('UseParallel',
'always');

options =
psoptimset('UseParallel',
'always', 'CompletePoll',
'on', 'Vectorized', 'off');

options =
gaoptimset('UseParallel',
'always', 'Vectorized',
'off');

For Optimization Tool:

For Optimization Tool:

Options > User function

evaluation > Evaluate
objective and constraint
functions > in parallel

Options > User function

evaluation > Evaluate
fitness and constraint
functions > in parallel

or
ms.UseParallel = 'always'

Options > Complete poll

> on
After you establish your parallel computing environment, applicable solvers
automatically use parallel computing whenever you call them with options.
To stop computing optimizations in parallel, set UseParallel to 'never',
or set the Optimization Tool not to compute in parallel. To halt all parallel
computation, enter
matlabpool close

8-13

Parallel Processing

Parallel Search Functions or Hybrid Functions

To have a patternsearch search function run in parallel, or a hybrid function
for ga or simulannealbnd run in parallel, do the following.
1 Set up parallel processing as described in Multicore Processors on page

8-11 or Processor Network on page 8-12.

2 Ensure that your search function or hybrid function has the conditions

outlined in these sections:

patternsearch Search Function on page 8-14
Parallel Hybrid Functions on page 8-15

patternsearch Search Function

patternsearch uses a parallel search function under the following conditions:

CompleteSearch is 'on'.
The search method is not @searchneldermead or custom.
If the search method is a patternsearch poll method or Latin hypercube
search, UseParallel is 'always'. Set at the command line with
psoptimset:
options = psoptimset('UseParallel','always',...
'CompleteSearch','on','SearchMethod',@GPSPositiveBasis2N);

Or you can use the Optimization Tool.

8-14

How to Use Parallel Processing

If the search method is ga, the search method option structure has
UseParallel set to 'always'. Set at the command line with psoptimset
and gaoptimset:
iterlim = 1; % iteration limit, specifies # ga runs
gaopt = gaoptimset('UseParallel','always');
options = psoptimset('SearchMethod',...
{@searchga,iterlim,gaopt});

In the Optimization Tool, first create the gaopt structure as above, and then
use these settings:

For more information about search functions, see Using a Search Method
on page 4-55.

Parallel Hybrid Functions

ga and simulannealbnd can have other solvers run after or interspersed

with their iterations. These other solvers are called hybrid functions. For
information on using a hybrid function with gamultiobj, see Parallel
Computing with gamultiobj on page 8-9. Both patternsearch and fmincon
can be hybrid functions. You can set options so that patternsearch runs in
parallel, or fmincon estimates gradients in parallel.
Set the options for the hybrid function as described in Hybrid Function
Options on page 9-47 for ga, or Hybrid Function Options on page 9-57
for simulannealbnd. To summarize:

8-15

Parallel Processing

If your hybrid function is patternsearch

1 Create a patternsearch options structure:

hybridopts = psoptimset('UseParallel','always',...
'CompletePoll','on');
2 Set the ga or simulannealbnd options to use patternsearch as a hybrid

function:
options = gaoptimset('UseParallel','always'); % for ga
options = gaoptimset(options,...
'HybridFcn',{@patternsearch,hybridopts});
% or, for simulannealbnd:
options = saoptimset('HybridFcn',{@patternsearch,hybridopts});

Or use the Optimization Tool.

For more information on parallel patternsearch, see Pattern Search

on page 8-7.
If your hybrid function is fmincon:
1 Create a fmincon options structure:

hybridopts = optimset('UseParallel','always',...
'Algorithm','interior-point');
% You can use any Algorithm except trust-region-reflective
2 Set the ga or simulannealbnd options to use fmincon as a hybrid

function:
options = gaoptimset('UseParallel','always');
options = gaoptimset(options,'HybridFcn',{@fmincon,hybridopts});

8-16

How to Use Parallel Processing

% or, for simulannealbnd:

options = saoptimset('HybridFcn',{@fmincon,hybridopts});

Or use the Optimization Tool.

For more information on parallel fmincon, see Parallel Computing for

Optimization in the Optimization Toolbox documentation.

8-17

8-18

Parallel Processing

9
Options Reference
GlobalSearch and MultiStart Properties (Options) on page 9-2
Pattern Search Options on page 9-7
Genetic Algorithm Options on page 9-29
Simulated Annealing Options on page 9-53

Options Reference

GlobalSearch and MultiStart Properties (Options)

In this section...
How to Set Properties on page 9-2
Properties of Both Objects on page 9-2
GlobalSearch Properties on page 9-4
MultiStart Properties on page 9-6

How to Set Properties

To set a property of an existing GlobalSearch or MultiStart object, use dot
addressing. For example, if ms is a MultiStart object, and you want to set the
Display property to 'iter', enter
ms.Display = 'iter';

To set multiple properties simultaneously, use the constructor (GlobalSearch

or MultiStart) with parameter-value pairs. For example, to set the Display
property to 'iter' and the MaxTime property to 100, enter
ms = MultiStart(ms,'Display','iter','MaxTime',100);

For more information on setting properties, see Changing Global Options

on page 3-57.

Properties of Both Objects

You can create a MultiStart object from a GlobalSearch object and
vice-versa.
The syntax for creating a new object from an existing object is:
ms = MultiStart(gs);
or
gs = GlobalSearch(ms);

The new object contains the properties that apply of the old object. This
section describes those shared properties:

9-2

GlobalSearch and MultiStart Properties (Options)

Display on page 9-3

MaxTime on page 9-3
StartPointsToRun on page 9-3
TolX on page 9-3
TolFun on page 9-4

Display
Values for the Display property are:
'final' (default) Summary results to command line after last solver
run.
'off' No output to command line.
'iter' Summary results to command line after each local solver run.

MaxTime
The MaxTime property describes a tolerance on the number of seconds since
the solver began its run. Solvers halt when they see MaxTime seconds have
passed since the beginning of the run. Time means wall clock as opposed to
processor cycles. The default is Inf.

StartPointsToRun
The StartPointsToRun property directs the solver to exclude certain start
points from being run:
all Accept all start points.
bounds Reject start points that do not satisfy bounds.
bounds-ineqs Reject start points that do not satisfy bounds or inequality
constraints.

TolX
The TolX property describes how close two points must be for solvers to
consider them identical for creating the vector of local solutions. Set TolX to 0
to obtain the results of every local solver run. Set TolX to a larger value to

9-3

Options Reference

have fewer results. Solvers compute the distance between a pair of points
with norm, the Euclidean distance.
Solvers consider two solutions identical if they are within TolX distance of
each other and have objective function values within TolFun of each other. If
both conditions are not met, solvers report the solutions as distinct.

TolFun
The TolFun property describes how close two objective function values must
be for solvers to consider them identical for creating the vector of local
solutions. Set TolFun to 0 to obtain the results of every local solver run. Set
TolFun to a larger value to have fewer results.
Solvers consider two solutions identical if they are within TolX distance of
each other and have objective function values within TolFun of each other. If
both conditions are not met, solvers report the solutions as distinct.

GlobalSearch Properties
NumTrialPoints on page 9-4
NumStageOnePoints on page 9-5
MaxWaitCycle on page 9-5
BasinRadiusFactor on page 9-5
DistanceThresholdFactor on page 9-5
PenaltyThresholdFactor on page 9-5

NumTrialPoints
Number of potential start points to examine in addition to x0 from the
problem structure. GlobalSearch runs only those potential start points that
pass several tests. For more information, see GlobalSearch Algorithm on
page 3-37.
Default: 1000

9-4

GlobalSearch and MultiStart Properties (Options)

NumStageOnePoints
Number of start points in Stage 1. For details, see Obtain Stage 1 Start
Point, Run on page 3-38.
Default: 200

MaxWaitCycle
A positive integer tolerance appearing in several points in the algorithm.
If the observed penalty function of MaxWaitCycle consecutive trial points
is at least the penalty threshold, then raise the penalty threshold (see
PenaltyThresholdFactor on page 9-5).
If MaxWaitCycle consecutive trial points are in a basin, then update that
basins radius (see BasinRadiusFactor on page 9-5).
Default: 20

BasinRadiusFactor
A basin radius decreases after MaxWaitCycle consecutive start points
are within the basin. The basin radius decreases by a factor of
1 BasinRadiusFactor.
Default: 0.2

DistanceThresholdFactor
A multiplier for determining whether a trial point is in an existing basin of
attraction. For details, see Examine Stage 2 Trial Point to See if fmincon
Runs on page 3-39. Default: 0.75

PenaltyThresholdFactor
Determines increase in penalty threshold. For details, see React to Large
Counter Values.
Default: 0.2

9-5

Options Reference

MultiStart Properties
UseParallel
The UseParallel property determines whether the solver distributes start
points to multiple processors:
'never' (default) Do not run in parallel.
'always' Run in parallel.
For the solver to run in parallel you must set up a parallel environment with
matlabpool. For details, see How to Use Parallel Processing on page 8-11.

9-6

Pattern Search Options

Pattern Search Options

In this section...
Optimization Tool vs. Command Line on page 9-7
Plot Options on page 9-8
Poll Options on page 9-10
Search Options on page 9-13
Mesh Options on page 9-17
Algorithm Settings on page 9-19
Cache Options on page 9-19
Stopping Criteria on page 9-20
Output Function Options on page 9-21
Display to Command Window Options on page 9-23
Vectorize and Parallel Options (User Function Evaluation) on page 9-24
Options Table for Pattern Search Algorithms on page 9-25

Optimization Tool vs. Command Line

There are two ways to specify options for pattern search, depending on
whether you are using the Optimization Tool or calling the function
patternsearch at the command line:
If you are using the Optimization Tool, you specify the options by selecting
an option from a drop-down list or by entering the value of the option in
the text field.
If you are calling patternsearch from the command line, you specify the
options by creating an options structure using the function psoptimset,
as follows:
options = psoptimset('Param1',value1,'Param2',value2,...);

See Setting Options for patternsearch at the Command Line on page

4-38 for examples.

9-7

Options Reference

In this section, each option is listed in two ways:

By its label, as it appears in the Optimization Tool
By its field name in the options structure
For example:
Poll method refers to the label of the option in the Optimization Tool.
PollMethod refers to the corresponding field of the options structure.

Plot Options
Plot options enable you to plot data from the pattern search while it is
running. When you select plot functions and run the pattern search, a plot
window displays the plots on separate axes. You can stop the algorithm at
any time by clicking the Stop button on the plot window.
Plot interval (PlotInterval) specifies the number of iterations between
consecutive calls to the plot function.
You can select any of the following plots in the Plot functions pane.
Best function value (@psplotbestf) plots the best objective function
value.
Function count (@psplotfuncount) plots the number of function
evaluations.
Mesh size (@psplotmeshsize) plots the mesh size.
Best point (@psplotbestx) plots the current best point.
Max constraint (@psplotmaxconstr) plots the maximum nonlinear
constraint violation.
Custom enables you to use your own plot function. To specify the plot
function using the Optimization Tool,

9-8

Select Custom function.

Enter @myfun in the text box, where myfun is the name of your function.

Pattern Search Options

Structure of the Plot Functions on page 9-9 describes the structure of

a plot function.
To display a plot when calling patternsearch from the command line, set the
PlotFcns field of options to be a function handle to the plot function. For
example, to display the best function value, set options as follows
options = psoptimset('PlotFcns', @psplotbestf);

To display multiple plots, use the syntax

options = psoptimset('PlotFcns', {@plotfun1, @plotfun2, ...});

where @plotfun1, @plotfun2, and so on are function handles to the plot

functions (listed in parentheses in the preceding list).

Structure of the Plot Functions

The first line of a plot function has the form
function stop = plotfun(optimvalues, flag)

The input arguments to the function are

optimvalues Structure containing information about the current state
of the solver. The structure contains the following fields:

x Current point
iteration Iteration number
fval Objective function value
meshsize Current mesh size
funccount Number of function evaluations
method Method used in last iteration
TolFun Tolerance on function value in last iteration
TolX Tolerance on x value in last iteration
nonlinineq Nonlinear inequality constraints, displayed only when a
nonlinear constraint function is specified

9-9

Options Reference

nonlineq Nonlinear equality constraints, displayed only when a

nonlinear constraint function is specified

flag Current state in which the plot function is called. The possible
values for flag are

init Initialization state

iter Iteration state
interrupt Intermediate stage
done Final state

Passing Extra Parameters in the Optimization Toolbox documentation

explains how to provide additional parameters to the function.
The output argument stop provides a way to stop the algorithm at the current
iteration. stop can have the following values:
false The algorithm continues to the next iteration.
true The algorithm terminates at the current iteration.

Poll Options
Poll options control how the pattern search polls the mesh points at each
iteration.
Poll method (PollMethod) specifies the pattern the algorithm uses to create
the mesh. There are two patterns for each of the classes of direct search
algorithms: the generalized pattern search (GPS) algorithm, the generating
set search (GSS) algorithm, and the mesh adaptive direct search (MADS)
algorithm. These patterns are the Positive basis 2N and the Positive basis
N+1:
The default pattern, GPS Positive basis 2N (GPSPositiveBasis2N),
consists of the following 2N vectors, where N is the number of independent
variables for the objective function.
[1 0 0...0]
[0 1 0...0]
...

9-10

Pattern Search Options

[0 0 0...1]
[1 0 0...0]
[0 1 0...0]
[0 0 0...1].
For example, if the optimization problem has three independent variables,
the pattern consists of the following six vectors.
[1 0 0]
[0 1 0]
[0 0 1]
[1 0 0]
[0 1 0]
[0 0 1].
The GSS Positive basis 2N pattern (GSSPositiveBasis2N) is similar to
GPS Positive basis 2N, but adjusts the basis vectors to account for linear
constraints. GSS Positive basis 2N is more efficient than GPS Positive
basis 2N when the current point is near a linear constraint boundary.
The MADS Positive basis 2N pattern (MADSPositiveBasis2N) consists
of 2N randomly generated vectors, where N is the number of independent
variables for the objective function. This is done by randomly generating
N vectors which form a linearly independent set, then using this first set
and the negative of this set gives 2N vectors. As shown above, the GPS
Positive basis 2N pattern is formed using the positive and negative
of the linearly independent identity, however, with the MADS Positive
basis 2N, the pattern is generated using a random permutation of an
N-by-N linearly independent lower triangular matrix that is regenerated
at each iteration.
The GPS Positive basis NP1 pattern consists of the following N + 1
vectors.
[1 0 0...0]
[0 1 0...0]
...
[0 0 0...1]
[1 1 1...1].

9-11

Options Reference

For example, if the objective function has three independent variables, the
pattern consists of the following four vectors.
[1 0 0]
[0 1 0]
[0 0 1]
[1 1 1].
The GSS Positive basis Np1 pattern (GSSPositiveBasisNp1) is similar
to GPS Positive basis Np1, but adjusts the basis vectors to account for
linear constraints. GSS Positive basis Np1 is more efficient than GPS
Positive basis Np1 when the current point is near a linear constraint
boundary.
The MADS Positive basis Np1 pattern (MADSPositiveBasisNp1) consists
of N randomly generated vectors to form the positive basis, where N is
the number of independent variables for the objective function. Then, one
more random vector is generated, giving N+1 randomly generated vectors.
Each iteration generates a new pattern when the MADS Positive basis
N+1 is selected.
Complete poll (CompletePoll) specifies whether all the points in the
current mesh must be polled at each iteration. Complete Poll can have
the values On or Off.
If you set Complete poll to On, the algorithm polls all the points in the
mesh at each iteration and chooses the point with the smallest objective
function value as the current point at the next iteration.
If you set Complete poll to Off, the default value, the algorithm stops the
poll as soon as it finds a point whose objective function value is less than
that of the current point. The algorithm then sets that point as the current
point at the next iteration.
Polling order (PollingOrder) specifies the order in which the algorithm
searches the points in the current mesh. The options are
Random The polling order is random.
Success The first search direction at each iteration is the direction in
which the algorithm found the best point at the previous iteration. After

9-12

Pattern Search Options

the first point, the algorithm polls the mesh points in the same order as
Consecutive.
Consecutive The algorithm polls the mesh points in consecutive order,
that is, the order of the pattern vectors as described in Poll Method on
page 4-42.

Search Options
Search options specify an optional search that the algorithm can perform at
each iteration prior to the polling. If the search returns a point that improves
the objective function, the algorithm uses that point at the next iteration and
omits the polling. Please note, if you have selected the same Search method
and Poll method, only the option selected in the Poll method will be used,
although both will be used when the options selected are different.
Complete search (CompleteSearch) applies when you set Search method
to GPS Positive basis Np1, GPS Positive basis 2N, GSS Positive basis
Np1, GSS Positive basis 2N, MADS Positive basis Np1, MADS Positive
basis 2N, or Latin hypercube. Complete search can have the values On or
Off.
For GPS Positive basis Np1, MADS Positive basis Np1, GPS Positive
basis 2N, or MADS Positive basis 2N, Complete search has the same
meaning as the poll option Complete poll.
Search method (SearchMethod) specifies the optional search step. The
options are
None ([]) (the default) specifies no search step.
GPS Positive basis 2N (@GPSPositiveBasis2N)
GPS Positive basis Np1 (@GPSPositiveBasisNp1)
GSS Positive basis 2N (@GSSPositiveBasis2N)
GSS Positive basis Np1 (@GSSPositiveBasisNp1)
MADS Positive basis 2N (@MADSPositiveBasis2N)
MADS Positive basis Np1 (@MADSPositiveBasisNp1)

9-13

Options Reference

Genetic Algorithm (@searchga) specifies a search using the genetic

algorithm. If you select Genetic Algorithm, two other options appear:

Iteration limit Positive integer specifying the number of iterations of

the pattern search for which the genetic algorithm search is performed.
The default for Iteration limit is 1.

Options Options structure for the genetic algorithm, which you can
set using gaoptimset.

To change the default values of Iteration limit and Options at the

command line, use the syntax
options = psoptimset('SearchMethod',...
{@searchga,iterlim,optionsGA})

where iterlim is the value of Iteration limit and optionsGA is the

genetic algorithm options structure.
Latin hypercube (@searchlhs) specifies a Latin hypercube search.
patternsearch generates each point for the search as follows. For each
component, take a random permutation of the vector [1,2,...,k] minus
rand(1,k), divided by k. (k is the number of points.) This yields k points,
with each component close to evenly spaced. The resulting points are then
scaled to fit any bounds. Latin hypercube uses default bounds of -1 and 1.
The way the search is performed depends on the setting for Complete
search:

If you set Complete search to On, the algorithm polls all the points that
are randomly generated at each iteration by the Latin hypercube search
and chooses the one with the smallest objective function value.

If you set Complete search to Off (the default), the algorithm stops
the poll as soon as it finds one of the randomly generated points whose
objective function value is less than that of the current point, and
chooses that point for the next iteration.

If you select Latin hypercube, two other options appear:

9-14

Iteration limit Positive integer specifying the number of iterations

of the pattern search for which the Latin hypercube search is performed.
The default for Iteration limit is 1.

Pattern Search Options

Design level The Design level is the number of points

patternsearch searches, a positive integer. The default for Design
level is 15 times the number of dimensions.

To change the default values of Iteration limit and Design level at the
command line, use the syntax
options=psoptimset('SearchMethod', {@searchlhs,iterlim,level})

where iterlim is the value of Iteration limit and level is the value of
Design level.
Nelder-Mead (@searchneldermead) specifies a search using fminsearch,
which uses the Nelder-Mead algorithm. If you select Nelder-Mead, two
other options appear:

Iteration limit Positive integer specifying the number of iterations

of the pattern search for which the Nelder-Mead search is performed.
The default for Iteration limit is 1.

Options Options structure for the function fminsearch, which you

can create using the function optimset.

To change the default values of Iteration limit and Options at the

command line, use the syntax
options=psoptimset('SearchMethod',...
{@searchga,iterlim,optionsNM})

where iterlim is the value of Iteration limit and optionsNM is the

options structure.
Custom enables you to write your own search function. To specify the
search function using the Optimization Tool,

Set Search function to Custom.

Set Function name to @myfun, where myfun is the name of your
function.

If you are using patternsearch, set

options = psoptimset('SearchMethod', @myfun);

To see a template that you can use to write your own search function, enter

9-15

Options Reference

edit searchfcntemplate

The following section describes the structure of the search function.

Structure of the Search Function

Your search function must have the following calling syntax.
function [successSearch,xBest,fBest,funccount] =
searchfcntemplate(fun,x,A,b,Aeq,beq,lb,ub, ...
optimValues,options)

The search function has the following input arguments:

fun Objective function
x Current point
A,b Linear inequality constraints
Aeq,beq Linear equality constraints
lb,ub Lower and upper bound constraints
optimValues Structure that enables you to set search options. The
structure contains the following fields:

x Current point
fval Objective function value at x
iteration Current iteration number
funccount Counter for user function evaluation
scale Scale factor used to scale the design points
problemtype Flag passed to the search routines, indicating
whether the problem is 'unconstrained', 'boundconstraints', or
'linearconstraints'. This field is a subproblem type for nonlinear
constrained problems.
meshsize Current mesh size used in search step
method Method used in last iteration

options Pattern search options structure

9-16

Pattern Search Options

The function has the following output arguments:

successSearch A Boolean identifier indicating whether the search is
successful or not
xBest,fBest Best point and best function value found by search method
Note If you set Search method to Genetic algorithm or Nelder-Mead,
we recommend that you leave Iteration limit set to the default value 1,
because performing these searches more than once is not likely to improve
results.
funccount Number of user function evaluation in search method
See Using a Search Method on page 4-55 for an example.

Mesh Options
Mesh options control the mesh that the pattern search uses. The following
options are available.
Initial size (InitialMeshSize) specifies the size of the initial mesh, which is
the length of the shortest vector from the initial point to a mesh point. Initial
size should be a positive scalar. The default is 1.0.
Max size (MaxMeshSize) specifies a maximum size for the mesh. When the
maximum size is reached, the mesh size does not increase after a successful
iteration. Max size must be a positive scalar, and is only used when a GPS or
GSS algorithm is selected as the Poll or Search method. The default value is
Inf. MADS uses a maximum size of 1.
Accelerator (MeshAccelerator) specifies whether, when the mesh size is
small, the Contraction factor is multiplied by 0.5 after each unsuccessful
iteration. Accelerator can have the values On or Off, the default.
Accelerator applies to the GPS and GSS algorithms.
Rotate (MeshRotate) is only applied when Poll method is set to GPS
Positive basis Np1 or GSS Positive basis Np1. It specifies whether the

9-17

Options Reference

mesh vectors are multiplied by 1 when the mesh size is less than 1/100 of the
mesh tolerance (minimum mesh size TolMesh) after an unsuccessful poll. In
other words, after the first unsuccessful poll with small mesh size, instead of
polling in directions ei (unit positive directions) and ei, the algorithm polls
in directions ei and ei. Rotate can have the values Off or On (the default).
When the problem has equality constraints, Rotate is disabled.
Rotate is especially useful for discontinuous functions.
Note Changing the setting of Rotate has no effect on the poll when Poll
method is set to GPS Positive basis 2N, GSS Positive basis 2N, MADS
Positive basis 2N, or MADS Positive basis Np1.
Scale (ScaleMesh) specifies whether the algorithm scales the mesh points
by carefully multiplying the pattern vectors by constants proportional to the
logarithms of the absolute values of components of the current point (or, for
unconstrained problems, of the initial point). Scale can have the values
Off or On (the default). When the problem has equality constraints, Scale
is disabled.
Expansion factor (MeshExpansion) specifies the factor by which the mesh
size is increased after a successful poll. The default value is 2.0, which
means that the size of the mesh is multiplied by 2.0 after a successful poll.
Expansion factor must be a positive scalar and is only used when a GPS or
GSS method is selected as the Poll or Search method. MADS uses a factor of
4.0.
Contraction factor (MeshContraction) specifies the factor by which the
mesh size is decreased after an unsuccessful poll. The default value is
0.5, which means that the size of the mesh is multiplied by 0.5 after an
unsuccessful poll. Contraction factor must be a positive scalar and is only
used when a GPS or GSS method is selected as the Poll or Search method.
MADS uses a factor of 0.25.
See Mesh Expansion and Contraction on page 4-59 for more information.

9-18

Pattern Search Options

Algorithm Settings
Algorithm settings refer to constraint parameters. For details, see
Description of the Nonlinear Constraint Solver on page 4-24.
Parameters that can be specified for a nonlinear constraint algorithm include:
Initial penalty (InitialPenalty) Specifies an initial value of the
penalty parameter that is used by the algorithm. Initial penalty must be
greater than or equal to 1.
Penalty factor (PenaltyFactor) Increases the penalty parameter when
the problem is not solved to required accuracy and constraints are not
satisfied. Penalty factor must be greater than 1.
Bind tolerance (TolBind) specifies the tolerance for the distance from the
current point to the boundary of the feasible region with respect to linear
constraints. This means Bind tolerance specifies when a linear constraint is
active. Bind tolerance is not a stopping criterion. Active linear constraints
change the pattern of points patternsearch uses for polling or searching.
patternsearch always uses points that satisfy linear constraints to within
Bind tolerance. The default value of Bind tolerance is 1e-3.

Cache Options
The pattern search algorithm can keep a record of the points it has already
polled, so that it does not have to poll the same point more than once. If the
objective function requires a relatively long time to compute, the cache option
can speed up the algorithm. The memory allocated for recording the points is
called the cache. This option should only be used for deterministic objective
functions, but not for stochastic ones.
Cache (Cache) specifies whether a cache is used. The options are On and Off,
the default. When you set Cache to On, the algorithm does not evaluate the
objective function at any mesh points that are within Tolerance of a point
in the cache.
Tolerance (CacheTol) specifies how close a mesh point must be to a point in
the cache for the algorithm to omit polling it. Tolerance must be a positive
scalar. The default value is eps.

9-19

Options Reference

Size (CacheSize) specifies the size of the cache. Size must be a positive
scalar. The default value is 1e4.
See Using Cache on page 4-65 for more information.

Stopping Criteria
Stopping criteria determine what causes the pattern search algorithm to stop.
Pattern search uses the following criteria:
Mesh tolerance (TolMesh) specifies the minimum tolerance for mesh size.
The GPS and GSS algorithms stop if the mesh size becomes smaller than
Mesh tolerance. MADS 2N stops when the mesh size becomes smaller than
TolMesh^2. MADS Np1 stops when the mesh size becomes smaller than
(TolMesh/nVar)^2, where nVar is the number of elements of x0. The default
value of TolMesh is 1e-6.
Max iteration (MaxIter) specifies the maximum number of iterations the
algorithm performs. The algorithm stops if the number of iterations reaches
Max iteration. You can select either
100*numberOfVariables Maximum number of iterations is 100 times
the number of independent variables (the default).
Specify A positive integer for the maximum number of iterations
Max function evaluations (MaxFunEval) specifies the maximum number
of evaluations of the objective function. The algorithm stops if the number
of function evaluations reaches Max function evaluations. You can select
either
2000*numberOfVariables Maximum number of function evaluations
is 2000 times the number of independent variables.
Specify A positive integer for the maximum number of function
evaluations
Time limit (TimeLimit) specifies the maximum time in seconds the pattern
search algorithm runs before stopping. This also includes any specified pause
time for pattern search algorithms.

9-20

Pattern Search Options

X tolerance (TolX) specifies the minimum distance between the current

points at two consecutive iterations. The algorithm stops if the distance
between two consecutive points is less than X tolerance. The default value
is 1e-6.
Function tolerance (TolFun) specifies the minimum tolerance for the
objective function. After a successful poll, if the difference between the
function value at the previous best point and function value at the current
best point is less than TolFun, the algorithm halts. The default value is 1e-6.
See Setting Tolerances for the Solver on page 4-68 for an example.
Nonlinear constraint tolerance (TolCon) The Nonlinear constraint
tolerance is not used as stopping criterion. It is used to determine the
feasibility with respect to nonlinear constraints.

Output Function Options

Output functions are functions that the pattern search algorithm calls at each
iteration. The following options are available:
History to new window (@psoutputhistory) displays the history of
points computed by the algorithm in the MATLAB Command Window at
each multiple of Interval iterations.
Custom enables you to write your own output function. To specify the
output function using the Optimization Tool,

Select Custom function.

For multiple output functions, enter a cell array of output function

handles: {@myfun1,@myfun2,...}.

Enter @myfun in the text box, where myfun is the name of your function.
To pass extra parameters in the output function, use Anonymous
Functions.

At the command line, set

options = psoptimset('OutputFcn',@myfun);

For multiple output functions, enter a cell array:

9-21

Options Reference

options = psoptimset('OutputFcn',{@myfun1,@myfun2,...});

To see a template that you can use to write your own output function, enter
edit psoutputfcntemplate

at the MATLAB command prompt.

The following section describes the structure of the output function.

Structure of the Output Function

Your output function must have the following calling syntax:
[stop,options,optchanged] =
psoutputhistory(optimvalues,options,flag,interval)

The function has the following input arguments:

optimvalues Structure containing information about the current state
of the solver. The structure contains the following fields:

x Current point
iteration Iteration number
fval Objective function value
meshsize Current mesh size
funccount Number of function evaluations
method Method used in last iteration
TolFun Tolerance on function value in last iteration
TolX Tolerance on x value in last iteration
nonlinineq Nonlinear inequality constraints, displayed only when a
nonlinear constraint function is specified
nonlineq Nonlinear equality constraints, displayed only when a

nonlinear constraint function is specified

options Options structure

9-22

Pattern Search Options

flag Current state in which the output function is called. The possible
values for flag are

init Initialization state

iter Iteration state
interrupt Intermediate stage
done Final state

interval Optional interval argument

Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the output function.
The output function returns the following arguments to ga:
stop Provides a way to stop the algorithm at the current iteration. stop
can have the following values.

false The algorithm continues to the next iteration.

true The algorithm terminates at the current iteration.

options Options structure.

optchanged Flag indicating changes to options.

Display to Command Window Options

Level of display ('Display') specifies how much information is displayed
at the command line while the pattern search is running. The available
options are
Off ('off') No output is displayed.
Iterative ('iter') Information is displayed for each iteration.
Diagnose ('diagnose') Information is displayed for each iteration. In
addition, the diagnostic lists some problem information and the options
that are changed from the defaults.
Final ('final') The reason for stopping is displayed.
Both Iterative and Diagnose display the following information:

9-23

Options Reference

Iter Iteration number

FunEval Cumulative number of function evaluations
MeshSize Current mesh size
FunVal Objective function value of the current point
Method Outcome of the current poll (with no nonlinear constraint
function specified). With a nonlinear constraint function, Method displays
the update method used after a subproblem is solved.
Max Constraint Maximum nonlinear constraint violation (displayed
only when a nonlinear constraint function has been specified)
The default value of Level of display is
Off in the Optimization Tool
'final' in an options structure created using psoptimset

Vectorize and Parallel Options (User Function

Evaluation)
You can choose to have your objective and constraint functions evaluated in
serial, parallel, or in a vectorized fashion. These options are available in the
User function evaluation section of the Options pane of the Optimization
Tool, or by setting the 'Vectorized' and 'UseParallel' options with
psoptimset.
When Evaluate objective and constraint functions ('Vectorized') is
in serial ('Off'), patternsearch calls the objective function on one point
at a time as it loops through all of the mesh points. (At the command line,
this assumes 'UseParallel' is at its default value of 'never'.)
When Evaluate objective and constraint functions ('Vectorized') is
vectorized ('On'), patternsearch calls the objective function on all the
points in the mesh at once, i.e., in a single call to the objective function if
either Complete Poll or Complete Search is On.
If there are nonlinear constraints, the objective function and the nonlinear
constraints all need to be vectorized in order for the algorithm to compute
in a vectorized manner.

9-24

Pattern Search Options

When Evaluate objective and constraint functions (UseParallel)

is in parallel ('always') patternsearch calls the objective function in
parallel, using the parallel environment you established (see How to Use
Parallel Processing on page 8-11). At the command line, set UseParallel
to 'never' to compute serially.
Note You cannot simultaneously use vectorized and parallel computations.
If you set 'UseParallel' to 'always' and 'Vectorized' to 'On',
patternsearch evaluates your objective and constraint functions in a
vectorized manner, not in parallel.

How Objective and Constraint Functions Are Evaluated

Vectorized = Off

Vectorized = On

UseParallel = 'Never'

Serial

Vectorized

UseParallel = 'Always'

Parallel

Vectorized

Options Table for Pattern Search Algorithms

Option Availability Table for All Algorithms

Option

Description

Cache

With Cache set to 'on',

patternsearch keeps
a history of the mesh
points it polls and does
not poll points close to
them again at subsequent
iterations. Use this option
if patternsearch runs
slowly because it is taking
a long time to compute
the objective function. If
the objective function is

Algorithm
Availability
All

9-25

Options Reference

Option Availability Table for All Algorithms (Continued)

Option

Description

Algorithm
Availability

stochastic, it is advised not

to use this option.

9-26

CacheSize

Size of the cache, in number

of points.

All

CacheTol

Positive scalar specifying

how close the current mesh
point must be to a point
in the cache in order for
patternsearch to avoid
polling it. Available if
'Cache' option is set to
'on'.

All

CompletePoll

Complete poll around

current iterate. Evaluate
all the points in a poll step.

All

CompleteSearch

Complete search around

current iterate. Evaluate
all the points in a search
step.

All

Display

Level of display to
Command Window.

All

InitialMeshSize

Initial mesh size used in

pattern search algorithms.

All

InitialPenalty

Initial value of the penalty

parameter.

All

MaxFunEvals

Maximum number
of objective function
evaluations.

All

MaxIter

Maximum number of
iterations.

All

Pattern Search Options

Option Availability Table for All Algorithms (Continued)

Algorithm
Availability

Option

Description

MaxMeshSize

Maximum mesh size used

in a poll/search step.

GPS and GSS

MeshAccelerator

Accelerate mesh size

contraction.

GPS and GSS

MeshContraction

Mesh contraction factor,

used when iteration is
unsuccessful.

GPS and GSS

MeshExpansion

Mesh expansion factor,

expands mesh when
iteration is successful.

GPS and GSS

MeshRotate

Rotate the pattern before

declaring a point to be
optimum.

GPS Np1 and GSS

Np1

OutputFcn

User-specified function that

a pattern search calls at
each iteration.

All

PenaltyFactor

Penalty update parameter.

All

PlotFcn

Specifies function to plot at

runtime.

All

PlotInterval

Specifies that plot functions

will be called at every
interval.

All

PollingOrder

Order in which search

directions are polled.

GPS and GSS

PollMethod

Polling strategy used in

pattern search.

All

ScaleMesh

Automatic scaling of
variables.

All

9-27

Options Reference

Option Availability Table for All Algorithms (Continued)

9-28

Algorithm
Availability

Option

Description

SearchMethod

Specifies search method

used in pattern search.

All

TimeLimit

Total time (in seconds)

allowed for optimization.
Also includes any specified
pause time for pattern
search algorithms.

All

TolBind

Binding tolerance used

to determine if linear
constraint is active.

All

TolCon

Tolerance on nonlinear
constraints.

All

TolFun

Tolerance on function
value.

All

TolMesh

Tolerance on mesh size.

All

TolX

Tolerance on independent
variable.

All

UseParallel

When 'always', compute

objective functions of a
poll or search in parallel.
Disable by setting to
'never'.

All

Vectorized

Specifies whether objective

and constraint functions
are vectorized.

All

Genetic Algorithm Options

Genetic Algorithm Options

In this section...
Optimization Tool vs. Command Line on page 9-29
Plot Options on page 9-30
Population Options on page 9-33
Fitness Scaling Options on page 9-36
Selection Options on page 9-37
Reproduction Options on page 9-39
Mutation Options on page 9-40
Crossover Options on page 9-42
Migration Options on page 9-45
Algorithm Settings on page 9-46
Multiobjective Options on page 9-47
Hybrid Function Options on page 9-47
Stopping Criteria Options on page 9-48
Output Function Options on page 9-48
Display to Command Window Options on page 9-50
Vectorize and Parallel Options (User Function Evaluation) on page 9-51

Optimization Tool vs. Command Line

There are two ways to specify options for the genetic algorithm, depending on
whether you are using the Optimization Tool or calling the functions ga or
at the command line:
If you are using the Optimization Tool (optimtool), select an option from a
drop-down list or enter the value of the option in a text field.
If you are calling ga or gamultiobj from the command line, create an
options structure using the function gaoptimset, as follows:
options = gaoptimset('Param1', value1, 'Param2', value2, ...);

9-29

Options Reference

See Setting Options for ga at the Command Line on page 5-40 for
examples.
In this section, each option is listed in two ways:
By its label, as it appears in the Optimization Tool
By its field name in the options structure
For example:
Population type is the label of the option in the Optimization Tool.
PopulationType is the corresponding field of the options structure.

Plot Options
Plot options enable you to plot data from the genetic algorithm while it is
running. When you select plot functions and run the genetic algorithm, a plot
window displays the plots on separate axes. Click on any subplot to view
a larger version of the plot in a separate figure window. You can stop the
algorithm at any time by clicking the Stop button on the plot window.
Plot interval (PlotInterval) specifies the number of generations between
consecutive calls to the plot function.
You can select any of the following plot functions in the Plot functions pane:
Best fitness (@gaplotbestf) plots the best function value versus
generation.
Expectation (@gaplotexpectation) plots the expected number of children
versus the raw scores at each generation.
Score diversity (@gaplotscorediversity) plots a histogram of the scores
at each generation.
Stopping (@gaplotstopping) plots stopping criteria levels.
Best individual (@gaplotbestindiv) plots the vector entries of the
individual with the best fitness function value in each generation.

9-30

Genetic Algorithm Options

Genealogy (@gaplotgenealogy) plots the genealogy of individuals. Lines

from one generation to the next are color-coded as follows:

Red lines indicate mutation children.

Blue lines indicate crossover children.
Black lines indicate elite individuals.

Scores (@gaplotscores) plots the scores of the individuals at each

generation.
Max constraint (@gaplotmaxconstr) plots the maximum nonlinear
constraint violation at each generation.
Distance (@gaplotdistance) plots the average distance between
individuals at each generation.
Range (@gaplotrange) plots the minimum, maximum, and mean fitness
function values in each generation.
Selection (@gaplotselection) plots a histogram of the parents.
Custom function enables you to use plot functions of your own. To specify
the plot function if you are using the Optimization Tool,

Select Custom function.

Enter @myfun in the text box, where myfun is the name of your function.

See Structure of the Plot Functions on page 9-32.

To display a plot when calling ga from the command line, set the PlotFcns
field of options to be a function handle to the plot function. For example, to
display the best fitness plot, set options as follows
options = gaoptimset('PlotFcns', @gaplotbestf);

To display multiple plots, use the syntax

options =gaoptimset('PlotFcns', {@plotfun1, @plotfun2, ...});

where @plotfun1, @plotfun2, and so on are function handles to the plot

functions.

9-31

Options Reference

Structure of the Plot Functions

The first line of a plot function has the form
function state = plotfun(options, state, flag)

The input arguments to the function are

options Structure containing all the current options settings.
state Structure containing information about the current generation.
The State Structure on page 9-32 describes the fields of state.
flag String that tells what stage the algorithm is currently in.
Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the function.

The State Structure

The state structure, which is an input argument to plot, mutation, and output
functions, contains the following fields:
Population Population in the current generation
Score Scores of the current population
Generation Current generation number
StartTime Time when genetic algorithm started
StopFlag String containing the reason for stopping
Selection Indices of individuals selected for elite, crossover and
mutation
Expectation Expectation for selection of individuals
Best Vector containing the best score in each generation
LastImprovement Generation at which the last improvement in fitness
value occurred
LastImprovementTime Time at which last improvement occurred
NonlinIneq Nonlinear inequality constraints, displayed only when a
nonlinear constraint function is specified

9-32

Genetic Algorithm Options

NonlinEq Nonlinear equality constraints, displayed only when a

nonlinear constraint function is specified

Population Options
Population options enable you to specify the parameters of the population that
the genetic algorithm uses.
Population type (PopulationType) specifies the data type of the input to
the fitness function. You can set Population type to be one of the following:
Double Vector ('doubleVector') Use this option if the individuals in
the population have type double. This is the default.
Bit string ('bitstring') Use this option if the individuals in the
population are bit strings.
Custom ('custom') Use this option to create a population whose data
type is neither of the preceding.
If you use a custom population type, you must write your own creation,
mutation, and crossover functions that accept inputs of that population
type, and specify these functions in the following fields, respectively:

Creation function (CreationFcn)

Mutation function (MutationFcn)
Crossover function (CrossoverFcn)

Population type (PopulationType) specifies the type of the input to the

fitness function. Types and their restrictions:
Double vector ('doubleVector') Use this option if the individuals in
the population have type double. This is the default.
Bit string ('bitstring') Use this option if the individuals in the
population are bit strings. For Creation function (CreationFcn) and
Mutation function (MutationFcn), use Uniform (mutationuniform)
or Custom. For Crossover function (CrossoverFcn), use Scattered
(@crossoverscattered), Single point (@crossoversinglepoint), Two
point (@crossovertwopoint), or Custom. You cannot use a Hybrid
function or Nonlinear constraint function.

9-33

Options Reference

Custom For Crossover function and Mutation function, use

Custom. For Creation function, either use Custom, or provide an
Initial population. You cannot use a Hybrid function or Nonlinear
constraint function.
Population size (PopulationSize) specifies how many individuals there
are in each generation. With a large population size, the genetic algorithm
searches the solution space more thoroughly, thereby reducing the chance
that the algorithm will return a local minimum that is not a global minimum.
However, a large population size also causes the algorithm to run more slowly.
If you set Population size to a vector, the genetic algorithm creates multiple
subpopulations, the number of which is the length of the vector. The size of
each subpopulation is the corresponding entry of the vector.
Creation function (CreationFcn) specifies the function that creates the
initial population for ga. You can choose from the following functions:
Uniform (@gacreationuniform) creates a random initial population with
a uniform distribution. This is the default if there are no constraints or
bound constraints.
Feasible population (@gacreationlinearfeasible) creates a random
initial population that satisfies all bounds and linear constraints. It is
biased to create individuals that are on the boundaries of the constraints,
and to create well-dispersed populations. This is the default if there are
linear constraints.
Custom enables you to write your own creation function, which must
generate data of the type that you specify in Population type. To specify
the creation function if you are using the Optimization Tool,

Set Creation function to Custom.

Set Function name to @myfun, where myfun is the name of your
function.

If you are using ga, set

options = gaoptimset('CreationFcn', @myfun);

Your creation function must have the following calling syntax.

9-34

Genetic Algorithm Options

function Population = myfun(GenomeLength, FitnessFcn, options)

The input arguments to the function are

Genomelength Number of independent variables for the fitness

FitnessFcn Fitness function

function
options Options structure

The function returns Population, the initial population for the genetic
algorithm.
Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the function.
Initial population (InitialPopulation) specifies an initial population for
the genetic algorithm. The default value is [], in which case ga uses the
default Creation function to create an initial population. If you enter a
nonempty array in the Initial population field, the array must have no more
than Population size rows, and exactly Number of variables columns. In
this case, the genetic algorithm calls a Creation function to generate the
remaining individuals, if required.
Initial scores (InitialScores) specifies initial scores for the initial
population. The initial scores can also be partial.
Initial range (PopInitRange) specifies the range of the vectors in the initial
population that is generated by a creation function. You can set Initial
range to be a matrix with two rows and Number of variables columns,
each column of which has the form [lb; ub], where lb is the lower bound
and ub is the upper bound for the entries in that coordinate. If you specify
Initial range to be a 2-by-1 vector, each entry is expanded to a constant
row of length Number of variables.
See Example: Setting the Initial Range on page 5-50 for an example.

9-35

Options Reference

Fitness Scaling Options

Fitness scaling converts the raw fitness scores that are returned by the fitness
function to values in a range that is suitable for the selection function. You
can specify options for fitness scaling in the Fitness scaling pane.
Scaling function (FitnessScalingFcn) specifies the function that performs
the scaling. The options are
Rank (@fitscalingrank) The default fitness scaling function, Rank,
scales the raw scores based on the rank of each individual instead of its
score. The rank of an individual is its position in the sorted scores. The
rank of the most fit individual is 1, the next most fit is 2, and so on. Rank
fitness scaling removes the effect of the spread of the raw scores.
Proportional (@fitscalingprop) Proportional scaling makes the scaled
value of an individual proportional to its raw fitness score.
Top (@fitscalingtop) Top scaling scales the top individuals equally.
Selecting Top displays an additional field, Quantity, which specifies the
number of individuals that are assigned positive scaled values. Quantity
can be an integer between 1 and the population size or a fraction between 0
and 1 specifying a fraction of the population size. The default value is 0.4.
Each of the individuals that produce offspring is assigned an equal scaled
value, while the rest are assigned the value 0. The scaled values have the
form [0 1/n 1/n 0 0 1/n 0 0 1/n ...].
To change the default value for Quantity at the command line, use the
following syntax
options = gaoptimset('FitnessScalingFcn', {@fitscalingtop,
quantity})

where quantity is the value of Quantity.

Shift linear (@fitscalingshiftlinear) Shift linear scaling scales
the raw scores so that the expectation of the fittest individual is equal to
a constant multiplied by the average score. You specify the constant in
the Max survival rate field, which is displayed when you select Shift
linear. The default value is 2.
To change the default value of Max survival rate at the command line,
use the following syntax

9-36

Genetic Algorithm Options

options = gaoptimset('FitnessScalingFcn',
{@fitscalingshiftlinear, rate})

where rate is the value of Max survival rate.

Custom enables you to write your own scaling function. To specify the
scaling function using the Optimization Tool,

Set Scaling function to Custom.

Set Function name to @myfun, where myfun is the name of your
function.

If you are using ga at the command line, set

options = gaoptimset('FitnessScalingFcn', @myfun);

Your scaling function must have the following calling syntax:

function expectation = myfun(scores, nParents)

The input arguments to the function are

scores A vector of scalars, one for each member of the population

nParents The number of parents needed from this population

The function returns expectation, a row vector of scalars of the same

length as scores, giving the scaled values of each member of the
population. The sum of the entries of expectation must equal nParents.
Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the function.
See Fitness Scaling on page 5-59 for more information.

Selection Options
Selection options specify how the genetic algorithm chooses parents for the
next generation. You can specify the function the algorithm uses in the
Selection function (SelectionFcn) field in the Selection options pane.
The options are

9-37

Options Reference

Stochastic uniform (@selectionstochunif) The default selection

function, Stochastic uniform, lays out a line in which each parent
corresponds to a section of the line of length proportional to its scaled value.
The algorithm moves along the line in steps of equal size. At each step, the
algorithm allocates a parent from the section it lands on. The first step is a
uniform random number less than the step size.
Remainder (@selectionremainder) Remainder selection assigns parents
deterministically from the integer part of each individuals scaled value and
then uses roulette selection on the remaining fractional part. For example,
if the scaled value of an individual is 2.3, that individual is listed twice as
a parent because the integer part is 2. After parents have been assigned
according to the integer parts of the scaled values, the rest of the parents
are chosen stochastically. The probability that a parent is chosen in this
step is proportional to the fractional part of its scaled value.
Uniform (@selectionuniform) Uniform selection chooses parents using
the expectations and number of parents. Uniform selection is useful for
debugging and testing, but is not a very effective search strategy.
Roulette (@selectionroulette) Roulette selection chooses parents
by simulating a roulette wheel, in which the area of the section of the
wheel corresponding to an individual is proportional to the individuals
expectation. The algorithm uses a random number to select one of the
sections with a probability equal to its area.
Tournament (@selectiontournament) Tournament selection chooses
each parent by choosing Tournament size players at random and then
choosing the best individual out of that set to be a parent. Tournament
size must be at least 2. The default value of Tournament size is 4.
To change the default value of Tournament size at the command line,
use the syntax
options = gaoptimset('SelectionFcn',...
{@selecttournament,size})

where size is the value of Tournament size.

Custom enables you to write your own selection function. To specify the
selection function using the Optimization Tool,

9-38

Set Selection function to Custom.

Genetic Algorithm Options

Set Function name to @myfun, where myfun is the name of your

function.

If you are using ga at the command line, set

options = gaoptimset('SelectionFcn', @myfun);

Your selection function must have the following calling syntax:

function parents = myfun(expectation, nParents, options)

The input arguments to the function are

expectation Expected number of children for each member of the

nParents Number of parents to select

population
options Genetic algorithm options structure

The function returns parents, a row vector of length nParents containing

the indices of the parents that you select.
Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the function.
See Selection on page 5-62 for more information.

Reproduction Options
Reproduction options specify how the genetic algorithm creates children for
the next generation.
Elite count (EliteCount) specifies the number of individuals that are
guaranteed to survive to the next generation. Set Elite count to be a positive
integer less than or equal to the population size. The default value is 2.
Crossover fraction (CrossoverFraction) specifies the fraction of the next
generation, other than elite children, that are produced by crossover. Set
Crossover fraction to be a fraction between 0 and 1, either by entering the
fraction in the text box or moving the slider. The default value is 0.8.
See Setting the Crossover Fraction on page 5-66 for an example.

9-39

Options Reference

Mutation Options
Mutation options specify how the genetic algorithm makes small random
changes in the individuals in the population to create mutation children.
Mutation provides genetic diversity and enable the genetic algorithm to
search a broader space. You can specify the mutation function in the
Mutation function (MutationFcn) field in the Mutation options pane. You
can choose from the following functions:
Gaussian (mutationgaussian) The default mutation function, Gaussian,
adds a random number taken from a Gaussian distribution with mean
0 to each entry of the parent vector. The standard deviation of this
distribution is determined by the parameters Scale and Shrink, which
are displayed when you select Gaussian, and by the Initial range setting
in the Population options.

The Scale parameter determines the standard deviation at the first

generation. If you set Initial range to be a 2-by-1 vector v, the initial
standard deviation is the same at all coordinates of the parent vector,
and is given by Scale*(v(2) - v(1)).
If you set Initial range to be a vector v with two rows and Number of
variables columns, the initial standard deviation at coordinate i of the
parent vector is given by Scale*(v(i,2) - v(i,1)).

The Shrink parameter controls how the standard deviation shrinks

as generations go by. If you set Initial range to be a 2-by-1 vector,
the standard deviation at the kth generation, k, is the same at all
coordinates of the parent vector, and is given by the recursive formula

k = k1 1 Shrink
.
Generations

If you set Initial range to be a vector with two rows and Number of
variables columns, the standard deviation at coordinate i of the parent
vector at the kth generation, i,k, is given by the recursive formula

i,k = i,k1 1 Shrink

.
Generations

If you set Shrink to 1, the algorithm shrinks the standard deviation

in each coordinate linearly until it reaches 0 at the last generation is

9-40

Genetic Algorithm Options

reached. A negative value of Shrink causes the standard deviation to

grow.
The default value of both Scale and Shrink is 1. To change the default
values at the command line, use the syntax
options = gaoptimset('MutationFcn', ...
{@mutationgaussian, scale, shrink})

where scale and shrink are the values of Scale and Shrink, respectively.
Uniform (mutationuniform) Uniform mutation is a two-step process.
First, the algorithm selects a fraction of the vector entries of an individual
for mutation, where each entry has a probability Rate of being mutated.
The default value of Rate is 0.01. In the second step, the algorithm
replaces each selected entry by a random number selected uniformly from
the range for that entry.
To change the default value of Rate at the command line, use the syntax
options = gaoptimset('MutationFcn', {@mutationuniform, rate})

where rate is the value of Rate.

Adaptive Feasible (mutationadaptfeasible) randomly generates
directions that are adaptive with respect to the last successful or
unsuccessful generation. The feasible region is bounded by the constraints
and inequality constraints. A step length is chosen along each direction so
that linear constraints and bounds are satisfied.
Custom enables you to write your own mutation function. To specify the
mutation function using the Optimization Tool,

Set Mutation function to Custom.

Set Function name to @myfun, where myfun is the name of your
function.

If you are using ga, set

options = gaoptimset('MutationFcn', @myfun);

Your mutation function must have this calling syntax:

function mutationChildren = myfun(parents, options, nvars,

9-41

Options Reference

FitnessFcn, state, thisScore, thisPopulation)

The arguments to the function are

parents Row vector of parents chosen by the selection function

options Options structure
nvars Number of variables
FitnessFcn Fitness function
state Structure containing information about the current generation.
The State Structure on page 9-32 describes the fields of state.
thisScore Vector of scores of the current population
thisPopulation Matrix of individuals in the current population

The function returns mutationChildrenthe mutated offspringas a

matrix whose rows correspond to the children. The number of columns of
the matrix is Number of variables.
Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the function.

Crossover Options
Crossover options specify how the genetic algorithm combines two individuals,
or parents, to form a crossover child for the next generation.
Crossover function (CrossoverFcn) specifies the function that performs the
crossover. You can choose from the following functions:
Scattered (@crossoverscattered), the default crossover function, creates
a random binary vector and selects the genes where the vector is a 1 from
the first parent, and the genes where the vector is a 0 from the second
parent, and combines the genes to form the child. For example, if p1 and p2
are the parents
p1 = [a b c d e f g h]
p2 = [1 2 3 4 5 6 7 8]

and the binary vector is [1 1 0 0 1 0 0 0], the function returns the following
child:

9-42

Genetic Algorithm Options

child1 = [a b 3 4 e 6 7 8]

Single point (@crossoversinglepoint) chooses a random integer n

between 1 and Number of variables and then

Selects vector entries numbered less than or equal to n from the first
parent.

Selects vector entries numbered greater than n from the second parent.
Concatenates these entries to form a child vector.
For example, if p1 and p2 are the parents
p1 = [a b c d e f g h]
p2 = [1 2 3 4 5 6 7 8]

and the crossover point is 3, the function returns the following child.
child = [a b c 4 5 6 7 8]

Two point (@crossovertwopoint) selects two random integers m and n

between 1 and Number of variables. The function selects

Vector entries numbered less than or equal to m from the first parent
Vector entries numbered from m+1 to n, inclusive, from the second parent
Vector entries numbered greater than n from the first parent.

The algorithm then concatenates these genes to form a single gene. For
example, if p1 and p2 are the parents
p1 = [a b c d e f g h]
p2 = [1 2 3 4 5 6 7 8]

and the crossover points are 3 and 6, the function returns the following
child.
child = [a b c 4 5 6 g h]

Intermediate (@crossoverintermediate) creates children by taking a

weighted average of the parents. You can specify the weights by a single
parameter, Ratio, which can be a scalar or a row vector of length Number

9-43

Options Reference

of variables. The default is a vector of all 1s. The function creates the
child from parent1 and parent2 using the following formula.
child = parent1 + rand * Ratio * ( parent2 - parent1)

If all the entries of Ratio lie in the range [0, 1], the children produced are
within the hypercube defined by placing the parents at opposite vertices. If
Ratio is not in that range, the children might lie outside the hypercube. If
Ratio is a scalar, then all the children lie on the line between the parents.
To change the default value of Ratio at the command line, use the syntax
options = gaoptimset('CrossoverFcn', ...
{@crossoverintermediate, ratio});

where ratio is the value of Ratio.

Heuristic (@crossoverheuristic) returns a child that lies on the line
containing the two parents, a small distance away from the parent with the
better fitness value in the direction away from the parent with the worse
fitness value. You can specify how far the child is from the better parent
by the parameter Ratio, which appears when you select Heuristic. The
default value of Ratio is 1.2. If parent1 and parent2 are the parents, and
parent1 has the better fitness value, the function returns the child
child = parent2 + R * (parent1 - parent2);

To change the default value of Ratio at the command line, use the syntax
options=gaoptimset('CrossoverFcn',...
{@crossoverheuristic,ratio});

where ratio is the value of Ratio.

Arithmetic (@crossoverarithmetic) creates children that are the
weighted arithmetic mean of two parents. Children are always feasible
with respect to linear constraints and bounds.
Custom enables you to write your own crossover function. To specify the
crossover function using the Optimization Tool,

9-44

Set Crossover function to Custom.

Genetic Algorithm Options

Set Function name to @myfun, where myfun is the name of your

function.

If you are using ga, set

options = gaoptimset('CrossoverFcn',@myfun);

Your crossover function must have the following calling syntax.

xoverKids = myfun(parents, options, nvars, FitnessFcn, ...
unused,thisPopulation)

The arguments to the function are

parents Row vector of parents chosen by the selection function

options options structure
nvars Number of variables
FitnessFcn Fitness function
unused Placeholder not used
thisPopulation Matrix representing the current population. The
number of rows of the matrix is Population size and the number of
columns is Number of variables.

The function returns xoverKidsthe crossover offspringas a matrix

whose rows correspond to the children. The number of columns of the
matrix is Number of variables.
Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the function.

Migration Options
Migration options specify how individuals move between subpopulations.
Migration occurs if you set Population size to be a vector of length greater
than 1. When migration occurs, the best individuals from one subpopulation
replace the worst individuals in another subpopulation. Individuals that
migrate from one subpopulation to another are copied. They are not removed
from the source subpopulation.

9-45

Options Reference

You can control how migration occurs by the following three fields in the
Migration options pane:
Direction (MigrationDirection) Migration can take place in one or
both directions.

If you set Direction to Forward ('forward'), migration takes place

toward the last subpopulation. That is, the nth subpopulation migrates
into the (n+1)th subpopulation.

If you set Direction to Both ('both'), the nth subpopulation migrates

into both the (n1)th and the (n+1)th subpopulation.

Migration wraps at the ends of the subpopulations. That is, the last
subpopulation migrates into the first, and the first may migrate into the
last.
Interval (MigrationInterval) Specifies how many generation pass
between migrations. For example, if you set Interval to 20, migration
takes place every 20 generations.
Fraction (MigrationFraction) Specifies how many individuals move
between subpopulations. Fraction specifies the fraction of the smaller of
the two subpopulations that moves. For example, if individuals migrate
from a subpopulation of 50 individuals into a subpopulation of 100
individuals and you set Fraction to 0.1, the number of individuals that
migrate is 0.1 * 50 = 5.

Algorithm Settings
Algorithm settings refer to the nonlinear constraint solver. For details, see
Description of the Nonlinear Constraint Solver on page 5-27.
Parameters that can be specified for a nonlinear constraint algorithm include
Initial penalty (InitialPenalty) Specifies an initial value of the
penalty parameter that is used by the algorithm. Initial penalty must be
greater than or equal to 1.
Penalty factor (PenaltyFactor) Increases the penalty parameter when
the problem is not solved to required accuracy and constraints are not
satisfied. Penalty factor must be greater than 1.

9-46

Genetic Algorithm Options

Multiobjective Options
Multiobjective options define parameters characteristic of the multiobjective
genetic algorithm. You can specify the following parameters:
DistanceMeasureFcn Defines a handle to the function that computes
distance measure of individuals, computed in decision variable or design
space (genotype) or in function space (phenotype). For example, the default
distance measure function is distancecrowding in function space, or
{@distancecrowding,'phenotype'}.
ParetoFraction Sets the fraction of individuals to keep on the first
Pareto front while the solver selects individuals from higher fronts. This
option is a scalar between 0 and 1.

Hybrid Function Options

A hybrid function is another minimization function that runs after the genetic
algorithm terminates. You can specify a hybrid function in Hybrid function
(HybridFcn) options. The choices are
[] No hybrid function.
fminsearch (@fminsearch) Uses the MATLAB function fminsearch to
perform unconstrained minimization.
patternsearch (@patternsearch) Uses a pattern search to perform
constrained or unconstrained minimization.
fminunc (@fminunc) Uses the Optimization Toolbox function fminunc to
perform unconstrained minimization.
fmincon (@fmincon) Uses the Optimization Toolbox function fmincon to
perform constrained minimization.
You can set a separate options structure for the hybrid function. Use
psoptimset or optimset to create the structure, depending on whether the
hybrid function is patternsearch or not:
hybridopts = optimset('display','iter','LargeScale','off');

Include the hybrid options in the Genetic Algorithm options structure as

follows:

9-47

Options Reference

options = gaoptimset(options,'HybridFcn',{@fminunc,hybridopts});
hybridopts must exist before you set options.

See Using a Hybrid Function on page 5-76 for an example.

Stopping Criteria Options

Stopping criteria determine what causes the algorithm to terminate. You can
specify the following options:
Generations (Generations) Specifies the maximum number of
iterations for the genetic algorithm to perform. The default is 100.
Time limit (TimeLimit) Specifies the maximum time in seconds the
genetic algorithm runs before stopping.
Fitness limit (FitnessLimit) The algorithm stops if the best fitness
value is less than or equal to the value of Fitness limit.
Stall generations (StallGenLimit) The algorithm stops if the weighted
average change in the fitness function value over Stall generations is less
than Function tolerance.
Stall time limit (StallTimeLimit) The algorithm stops if there is no
improvement in the best fitness value for an interval of time in seconds
specified by Stall time.
Function tolerance (TolFun) The algorithm runs until the cumulative
change in the fitness function value over Stall generations is less than
or equal to Function Tolerance.
Nonlinear constraint tolerance (TolCon) The Nonlinear constraint
tolerance is not used as stopping criterion. It is used to determine the
feasibility with respect to nonlinear constraints.
See Setting the Maximum Number of Generations on page 5-80 for an
example.

Output Function Options

Output functions are functions that the genetic algorithm calls at each
generation. The following options are available:

9-48

Genetic Algorithm Options

History to new window (@gaoutputgen) displays the history of points

computed by the algorithm in a new window at each multiple of Interval
iterations.
Custom enables you to write your own output function. To specify the output
function using the Optimization Tool,
Select Custom function.
Enter @myfun in the text box, where myfun is the name of your function.
To pass extra parameters in the output function, use Anonymous
Functions.
For multiple output functions, enter a cell array of output function handles:
{@myfun1,@myfun2,...}.
At the command line, set
options = gaoptimset('OutputFcn',@myfun);

For multiple output functions, enter a cell array:

options = gaoptimset('OutputFcn',{@myfun1,@myfun2,...});

To see a template that you can use to write your own output functions, enter
edit gaoutputfcntemplate

at the MATLAB command line.

Structure of the Output Function

The output function has the following calling syntax.
[state,options,optchanged] = myfun(options,state,flag,interval)

The function has the following input arguments:

options Options structure
state Structure containing information about the current generation.
The State Structure on page 9-32 describes the fields of state.

9-49

Options Reference

flag String indicating the current status of the algorithm as follows:

'init' Initial stage

'iter' Algorithm running
'interrupt' Intermediate stage
'done' Algorithm terminated

interval Optional interval argument

Passing Extra Parameters in the Optimization Toolbox documentation
explains how to provide additional parameters to the function.
The output function returns the following arguments to ga:
state Structure containing information about the current generation
options Options structure modified by the output function. This
argument is optional.
optchanged Flag indicating changes to options

Display to Command Window Options

Level of display ('Display') specifies how much information is displayed
at the command line while the genetic algorithm is running. The available
options are
Off ('off') No output is displayed.
Iterative ('iter') Information is displayed at each iteration.
Diagnose ('diagnose') Information is displayed at each iteration. In
addition, the diagnostic lists some problem information and the options
that have been changed from the defaults.
Final ('final') The reason for stopping is displayed.
Both Iterative and Diagnose display the following information:
Generation Generation number
f-count Cumulative number of fitness function evaluations

9-50

Genetic Algorithm Options

Best f(x) Best fitness function value

Mean f(x) Mean fitness function value
Stall generations Number of generations since the last improvement
of the fitness function
When a nonlinear constraint function has been specified, Iterative and
Diagnose will not display the Mean f(x), but will additionally display:
Max Constraint Maximum nonlinear constraint violation
The default value of Level of display is
Off in the Optimization Tool
'final' in an options structure created using gaoptimset

Vectorize and Parallel Options (User Function

Evaluation)
You can choose to have your fitness and constraint functions evaluated in
serial, parallel, or in a vectorized fashion. These options are available in the
User function evaluation section of the Options pane of the Optimization
Tool, or by setting the 'Vectorized' and 'UseParallel' options with
gaoptimset.
When Evaluate fitness and constraint functions ('Vectorized') is in
serial ('Off'), ga calls the fitness function on one individual at a time
as it loops through the population. (At the command line, this assumes
'UseParallel' is at its default value of 'never'.)
When Evaluate fitness and constraint functions ('Vectorized') is
vectorized ('On'), ga calls the fitness function on the entire population at
once, i.e., in a single call to the fitness function.
If there are nonlinear constraints, the fitness function and the nonlinear
constraints all need to be vectorized in order for the algorithm to compute
in a vectorized manner.
See Vectorizing the Fitness Function on page 5-81 for an example.
When Evaluate fitness and constraint functions (UseParallel) is in
parallel ('always'), ga calls the fitness function in parallel, using the

9-51

Options Reference

parallel environment you established (see How to Use Parallel Processing

on page 8-11). At the command line, set UseParallel to 'never' to
compute serially.
Note You cannot simultaneously use vectorized and parallel computations. If
you set 'UseParallel' to 'always' and 'Vectorized' to 'On', ga evaluates
your fitness and constraint functions in a vectorized manner, not in parallel.

How Fitness and Constraint Functions Are Evaluated

9-52

Vectorized = Off

Vectorized = On

UseParallel = 'Never'

Serial

Vectorized

UseParallel = 'Always'

Parallel

Vectorized

Simulated Annealing Options

Simulated Annealing Options

In this section...
saoptimset At The Command Line on page 9-53
Plot Options on page 9-53
Temperature Options on page 9-55
Algorithm Settings on page 9-56
Hybrid Function Options on page 9-57
Stopping Criteria Options on page 9-58
Output Function Options on page 9-59
Display Options on page 9-61

saoptimset At The Command Line

Specify options by creating an options structure using the saoptimset
function as follows:
options = saoptimset('Param1',value1,'Param2',value2, ...);

See Setting Options for simulannealbnd at the Command Line on page

6-15 for examples.
Each option in this section is listed by its field name in the options structure.
For example, InitialTemperature refers to the corresponding field of the
options structure.

Plot Options
Plot options enable you to plot data from the simulated annealing solver while
it is running. When you specify plot functions and run the algorithm, a plot
window displays the plots on separate axes. Right-click on any subplot to view
a larger version of the plot in a separate figure window.
PlotInterval specifies the number of iterations between consecutive calls to

the plot function.

9-53

Options Reference

To display a plot when calling simulannealbnd from the command line, set
the PlotFcns field of options to be a function handle to the plot function. You
can specify any of the following plots:
@saplotbestf plots the best objective function value.
@saplotbestx plots the current best point.
@saplotf plots the current function value.
@saplotx plots the current point.
@saplotstopping plots stopping criteria levels.
@saplottemperature plots the temperature at each iteration.
@myfun plots a custom plot function, where myfun is the name of your
function. See Structure of the Plot Functions on page 9-9 for a description
of the syntax.
For example, to display the best objective plot, set options as follows
options = saoptimset('PlotFcns',@saplotbestf);

To display multiple plots, use the cell array syntax

options = saoptimset('PlotFcns',{@plotfun1,@plotfun2, ...});

where @plotfun1, @plotfun2, and so on are function handles to the plot

functions.

Structure of the Plot Functions

The first line of a plot function has the form
function stop = plotfun(options,optimvalues,flag)

The input arguments to the function are

options Options structure created using saoptimset.
optimvalues Structure containing information about the current state
of the solver. The structure contains the following fields:

9-54

x Current point

Simulated Annealing Options

fval Objective function value at x

bestx Best point found so far
bestfval Objective function value at best point
temperature Current temperature
iteration Current iteration
funccount Number of function evaluations
t0 Start time for algorithm
k Annealing parameter

flag Current state in which the plot function is called. The possible
values for flag are

'init' Initialization state

'iter' Iteration state
'done' Final state

The output argument stop provides a way to stop the algorithm at the current
iteration. stop can have the following values:
false The algorithm continues to the next iteration.
true The algorithm terminates at the current iteration.

Temperature Options
Temperature options specify how the temperature will be lowered at each
iteration over the course of the algorithm.
InitialTemperature Initial temperature at the start of the algorithm.
The default is 100.
TemperatureFcn Function used to update the temperature schedule. Let
i denote the iteration number. The options are:

@temperatureexp The temperature is equal to InitialTemperature *

@temperaturefast The temperature is equal to InitialTemperature / i.

0.95^i. This is the default.

9-55

Options Reference

@temperatureboltz The temperature is equal to InitialTemperature /

@myfun Uses a custom function, myfun, to update temperature. The

ln(i).
syntax is:
temperature = myfun(optimValues,options)

where optimValues is a structure described in Structure of the Plot

Functions on page 9-54. options is either the structure created with
saoptimset, or the structure of default options, if you did not create an
options structure. For example, the function temperaturefast is:
temperature = options.InitialTemperature./optimValues.k;

ReannealInterval Number of points accepted before reannealing. The

default value is 100.

Algorithm Settings
Algorithm settings define algorithmic specific parameters used in generating
new points at each iteration.
Parameters that can be specified for simulannealbnd are:
DataType Type of data to use in the objective function. Choices:

'double' (default) A vector of type double.

'custom' Any other data type. You must provide a 'custom'

annealing function. You cannot use a hybrid function.

AnnealingFcn Function used to generate new points for the next

iteration. The choices are:

@annealingfast The step has length temperature, with direction

uniformly at random. This is the default.

@annealingboltz The step has length square root of temperature,

with direction uniformly at random.

@myfun Uses a custom annealing algorithm, myfun. The syntax is:

newx = myfun(optimValues,problem)

9-56

Simulated Annealing Options

where optimValues is a structure described in Structure of the Output

Function on page 9-59, and problem is a structure containing the
following information:
objective:
x0:

function handle to the objective function

the start point

nvar:

number of decision variables

lb:

lower bound on decision variables

ub:

upper bound on decision variables

For example, the current position is optimValues.x, and the current

objective function value is problem.objective(optimValues.x).
AcceptanceFcn Function used to determine whether a new point is

accepted or not. The choices are:

@acceptancesa Simulated annealing acceptance function. The default
for simulannealbnd.
@myfun A custom acceptance function, myfun. The syntax is:
acceptpoint = myfun(optimValues,newx,newfval);

where optimValues is a structure described in Structure of the Output

Function on page 9-59, newx is the point being evaluated for acceptance,
and newfval is the objective function at newx. acceptpoint is a Boolean,
with value true to accept newx, and false to reject newx.

Hybrid Function Options

A hybrid function is another minimization function that runs during or at the
end of iterations of the solver. HybridInterval specifies the interval (if not
never or end) at which the hybrid function is called. You can specify a hybrid
function using the HybridFcn option. The choices are:
[] No hybrid function.
@fminsearch Uses the MATLAB function fminsearch to perform
unconstrained minimization.

9-57

Options Reference

@patternsearch Uses patternsearch to perform constrained or

unconstrained minimization.
@fminunc Uses the Optimization Toolbox function fminunc to perform
unconstrained minimization.
@fmincon Uses the Optimization Toolbox function fmincon to perform
constrained minimization.
You can set a separate options structure for the hybrid function. Use
psoptimset or optimset to create the structure, depending on whether the
hybrid function is patternsearch or not:
hybridopts = optimset('display','iter','LargeScale','off');

Include the hybrid options in the simulannealbnd options structure as

follows:
options = saoptimset(options,'HybridFcn',{@fminunc,hybridopts});
hybridopts must exist before you set options.

See Using a Hybrid Function on page 5-76 for an example.

Stopping Criteria Options

Stopping criteria determine what causes the algorithm to terminate. You can
specify the following options:
TolFun The algorithm runs until the average change in value of the
objective function in StallIterLim iterations is less than TolFun. The
default value is 1e-6.
MaxIter The algorithm stops if the number of iterations exceeds this
maximum number of iterations. You can specify the maximum number of
iterations as a positive integer or Inf. Inf is the default.
MaxFunEval specifies the maximum number of evaluations of the objective
function. The algorithm stops if the number of function evaluations exceeds
the maximum number of function evaluations. The allowed maximum is
3000*numberofvariables.

9-58

Simulated Annealing Options

TimeLimit specifies the maximum time in seconds the algorithm runs

before stopping.
ObjectiveLimit The algorithm stops if the best objective function value
is less than or equal to the value of ObjectiveLimit.

Output Function Options

Output functions are functions that the algorithm calls at each iteration.
The default value is to have no output function, []. You must first create
an output function using the syntax described in Structure of the Output
Function on page 9-59.
Using the Optimization Tool:
Specify Output function as @myfun, where myfun is the name of your
function.
To pass extra parameters in the output function, use Anonymous
Functions.
For multiple output functions, enter a cell array of output function handles:
{@myfun1,@myfun2,...}.
At the command line:

options = saoptimset('OutputFcns',@myfun);

For multiple output functions, enter a cell array:

options = saoptimset('OutputFcn',{@myfun1,@myfun2,...});

To see a template that you can use to write your own output functions, enter
edit saoutputfcntemplate

at the MATLAB command line.

Structure of the Output Function

The output function has the following calling syntax.

9-59

Options Reference

[stop,options,optchanged] = myfun(options,optimvalues,flag)

The function has the following input arguments:

options Options structure created using saoptimset.
optimvalues Structure containing information about the current state
of the solver. The structure contains the following fields:

x Current point
fval Objective function value at x
bestx Best point found so far
bestfval Objective function value at best point
temperature Current temperature
iteration Current iteration
funccount Number of function evaluations
t0 Start time for algorithm
k Annealing parameter

flag Current state in which the output function is called. The possible
values for flag are

'init' Initialization state

'iter' Iteration state
'done' Final state

Passing Extra Parameters in the Optimization Toolbox documentation

explains how to provide additional parameters to the output function.
The output function returns the following arguments:
stop Provides a way to stop the algorithm at the current iteration. stop
can have the following values:

9-60

false The algorithm continues to the next iteration.

true The algorithm terminates at the current iteration.

Simulated Annealing Options

options Options structure modified by the output function.

optchanged A boolean flag indicating changes were made to options.
This must be set to true if options are changed.

Display Options
Use the Display option to specify how much information is displayed at the
command line while the algorithm is running. The available options are
off No output is displayed. This is the default value for an options
structure created using saoptimset.
iter Information is displayed at each iteration.
diagnose Information is displayed at each iteration. In addition, the
diagnostic lists some problem information and the options that have been
changed from the defaults.
final The reason for stopping is displayed. This is the default.
Both iter and diagnose display the following information:
Iteration Iteration number
f-count Cumulative number of objective function evaluations
Best f(x) Best objective function value
Current f(x) Current objective function value
Mean Temperature Mean temperature function value

9-61

9-62

Options Reference

10
Function Reference
GlobalSearch (p. 10-2)

Create and solve GlobalSearch

problems

MultiStart (p. 10-2)

Create and solve MultiStart

problems

Genetic Algorithm (p. 10-2)

Use genetic algorithm and

Optimization Tool, and modify
genetic algorithm options

Direct Search (p. 10-3)

Use direct search and Optimization

Tool, and modify pattern search
options

Simulated Annealing (p. 10-3)

Use simulated annealing and

Optimization Tool, and modify
simulated annealing options

10

Function Reference

GlobalSearch
createOptimProblem

Create optimization problem

structure

run (GlobalSearch)

Find global minimum

createOptimProblem

Create optimization problem

structure

list (CustomStartPointSet)

List custom start points in set

list (RandomStartPointSet)

Generate start points

run (MultiStart)

Run local solver from multiple points

MultiStart

Genetic Algorithm

10-2

ga

Find minimum of function using

genetic algorithm

gamultiobj

Find minima of multiple functions

using genetic algorithm

gaoptimget

Obtain values of genetic algorithm

options structure

gaoptimset

Create genetic algorithm options

structure

Direct Search

Direct Search
patternsearch

Find minimum of function using

pattern search

psoptimget

Obtain values of pattern search

options structure

psoptimset

Create pattern search options

structure

Simulated Annealing
saoptimget

Values of simulated annealing

options structure

saoptimset

Create simulated annealing options

structure

simulannealbnd

Find unconstrained or
bound-constrained minimum of
function of several variables using
simulated annealing algorithm

10-3

10

10-4

Function Reference

11
Class Reference
CustomStartPointSet

User-supplied start points

GlobalOptimSolution

Optimization solution

GlobalSearch

Find global minimum

MultiStart

Find multiple local minima

RandomStartPointSet

Random start points

11

11-2

Class Reference

12
Functions Alphabetical
List

createOptimProblem

Purpose

Create optimization problem structure

Syntax

problem = createOptimProblem('solverName')
problem = createOptimProblem('solverName','ParameterName',
ParameterValue,...)

Description

problem = createOptimProblem('solverName') creates an empty

optimization problem structure for the solverName solver.
problem =
createOptimProblem('solverName','ParameterName',ParameterValue,...)

accepts one or more comma-separated parameter name/value pairs.

Specify ParameterName inside single quotes.

Input
Arguments

solverName

Name of the solver. For a GlobalSearch problem, use 'fmincon'.

For a MultiStart problem, use 'fmincon', 'fminunc',
'lsqcurvefit' or 'lsqnonlin'.

Parameter Name/Value Pairs

Aeq

Matrix for linear equality constraints. The constraints have the

form:
Aeq x = beq
Aineq

Matrix for linear inequality constraints. The constraints have

the form:
Aineq x bineq
beq

Vector for linear equality constraints. The constraints have the

form:
Aeq x = beq

12-2

createOptimProblem

bineq

Vector for linear inequality constraints. The constraints have

the form:
Aineq x bineq
lb

Vector of lower bounds.

nonlcon

Function handle to the nonlinear constraint function. The

constraint function must accept a vector x and return two vectors:
c, the nonlinear inequality constraints, and ceq, the nonlinear
equality constraints. If one of these constraint functions is empty,
nonlcon must return [] for that function.
If the GradConstr option is 'on', then in addition nonlcon must
return two additional outputs, gradc and gradceq. The gradc
parameter is a matrix with one column for the gradient of each
constraint, as is gradceq.
For more information, see Constraints on page 2-6.
objective

Function handle to the objective function. For all solvers except

lsqnonlin and lsqcurvefit, the objective function must accept
a vector x and return a scalar. If the GradObj option is 'on',
then the objective function must return a second output, a vector,
representing the gradient of the objective. For lsqnonlin, the
objective function must accept a vector x and return a vector.
lsqnonlin sums the squares of the objective function values. For
lsqcurvefit, the objective function must accept two inputs, x
and xdata, and return a vector.
For more information, see Computing Objective Functions on
page 2-2.
options

12-3

createOptimProblem

Options structure. Create this structure with optimset, or by

exporting from the Optimization Tool.
ub

Vector of upper bounds.

x0

A vector, a potential starting point for the optimization. Gives the

dimensionality of the problem.
xdata

Vector of data points for lsqcurvefit.

ydata

Vector of data points for lsqcurvefit.

Output
Arguments

problem

Examples

Create a problem structure using Rosenbrocks function as objective

(see Using a Hybrid Function on page 5-76), the interior-point
algorithm for fmincon, and bounds with absolute value 2:

Optimization problem structure.

anonrosen = @(x)(100*(x(2) - x(1)^2)^2 + (1-x(1))^2);

opts = optimset('Algorithm','interior-point');
problem = createOptimProblem('fmincon','x0',randn(2,1),...
'objective',anonrosen,'lb',[-2;-2],'ub',[2;2],...
'options',opts);

12-4

Alternatives

You can create a problem structure by exporting from the Optimization

Tool (optimtool), as described in Exporting from the Optimization
Tool on page 3-8.

See Also

optimtool | MultiStart | GlobalSearch

Tutorials

Create a Problem Structure on page 3-4

CustomStartPointSet class

Purpose

User-supplied start points

Description

An object wrapper of a matrix whose rows represent start points for

MultiStart.

Construction

tpoints = CustomStartPointSet(ptmatrix) generates a

CustomStartPointSet object from the ptmatrix matrix. Each row of
ptmatrix represents one start point.

Properties

DimStartPoints

Dimension of each start point, a read-only property.

DimStartPoints is the number of columns in ptmatrix.
DimStartPoints should be the same as the number of elements in
problem.x0, the problem structure you pass to run.
NumStartPoints

Number of start points, a read-only property. This is the number

of rows in ptmatrix.

Methods

list

List custom start points in set

Copy
Semantics

Value. To learn how value classes affect copy operations, see Copying
Objects in the MATLAB Programming Fundamentals documentation.

Examples

Create a CustomStartPointSet object with 40 three-dimensional rows.

Each row represents a normally distributed random variable with mean
[10,10,10] and variance diag([4,4,4]):
fortypts = 10*ones(40,3) + 4*randn(40,3); % a matrix
startpts = CustomStartPointSet(fortypts);
startpts is the fortypts matrix in an object wrapper.

12-5

CustomStartPointSet class

Get the fortypts matrix from the startpts object of the previous
example:
fortypts = list(startpts);

See Also

RandomStartPointSet | MultiStart | list

Tutorials

CustomStartPointSet Object for Start Points on page 3-17

How To

Class Attributes
Property Attributes

12-6

ga

Purpose

Find minimum of function using genetic algorithm

Syntax

x = ga(fitnessfcn,nvars)
x = ga(fitnessfcn,nvars,A,b)
x = ga(fitnessfcn,nvars,A,b,Aeq,beq)
x = ga(fitnessfcn,nvars,A,b,Aeq,beq,LB,UB)
x = ga(fitnessfcn,nvars,A,b,Aeq,beq,LB,UB,nonlcon)
x = ga(fitnessfcn,nvars,A,b,Aeq,beq,LB,UB,nonlcon,options)
x = ga(problem)
[x,fval] = ga(...)
[x,fval,exitflag] = ga(...)

Description

ga implements the genetic algorithm at the command line to minimize

an objective function.
x = ga(fitnessfcn,nvars) finds a local unconstrained minimum, x,
to the objective function, fitnessfcn. nvars is the dimension (number
of design variables) of fitnessfcn. The objective function, fitnessfcn,
accepts a vector x of size 1-by-nvars, and returns a scalar evaluated at x.

Note To write a function with additional parameters to the independent

variables that can be called by ga, see the section on Passing Extra
Parameters in the Optimization Toolbox documentation.
x = ga(fitnessfcn,nvars,A,b) finds a local minimum x to
fitnessfcn, subject to the linear inequalities A x b . fitnessfcn
accepts input x and returns a scalar function value evaluated at x.

If the problem has m linear inequality constraints and nvars variables,

then
A is a matrix of size m-by-nvars.
b is a vector of length m.

12-7

ga

Note that the linear constraints are not satisfied when the
PopulationType option is set to 'bitString' or 'custom'.
x = ga(fitnessfcn,nvars,A,b,Aeq,beq) finds a local minimum x to
fitnessfcn, subject to the linear equalities Aeq x = beq as well as

A x b . (Set A=[] and b=[] if no inequalities exist.)

If the problem has r linear equality constraints and nvars variables,
then
Aeq is a matrix of size r-by-nvars.
beq is a vector of length r.
Note that the linear constraints are not satisfied when the
PopulationType option is set to 'bitString' or 'custom'.
x = ga(fitnessfcn,nvars,A,b,Aeq,beq,LB,UB) defines a set of lower
and upper bounds on the design variables, x, so that a solution is found

in the range LB x UB . Use empty matrices for LB and UB if no

bounds exist. Set LB(i) = -Inf if x(i) is unbounded below; set UB(i) = Inf
if x(i) is unbounded above.
x = ga(fitnessfcn,nvars,A,b,Aeq,beq,LB,UB,nonlcon) subjects
the minimization to the constraints defined in nonlcon. The function
nonlcon accepts x and returns the vectors C and Ceq, representing the
nonlinear inequalities and equalities respectively. ga minimizes the
fitnessfcn such that C(x) 0 and Ceq(x)=0. (Set LB=[] and UB=[]
if no bounds exist.)

Note that the nonlinear constraints are not satisfied when the
PopulationType option is set to 'bitString' or 'custom'.
x = ga(fitnessfcn,nvars,A,b,Aeq,beq,LB,UB,nonlcon,options)

minimizes with the default optimization parameters replaced by values

in the structure options, which can be created using the gaoptimset
function. See the gaoptimset reference page for details.
x = ga(problem) finds the minimum for problem, where problem is a
structure containing the following fields:

12-8

ga

fitnessfcn

Fitness function

nvars

Number of design variables

Aineq

A matrix for linear inequality constraints

Bineq

b vector for linear inequality constraints

Aeq

A matrix for linear equality constraints

Beq

b vector for linear equality constraints

lb

Lower bound on x

ub

Upper bound on x

nonlcon

Nonlinear constraint function

rngstate

Optional field to reset the state of the

random number generator

solver

'ga'

options

Options structure created using gaoptimset

or the Optimization Tool

Create the structure problem by exporting a problem from Optimization

Tool, as described in Importing and Exporting Your Work in the
Optimization Toolbox documentation.
[x,fval] = ga(...) returns fval, the value of the fitness function

at x.
[x,fval,exitflag] = ga(...) returns exitflag, an integer

identifying the reason the algorithm terminated. The following lists

the values of exitflag and the corresponding reasons the algorithm
terminated.
1 Average cumulative change in value of the fitness function over
options.StallGenLimit generations less than options.TolFun and
constraint violation less than options.TolCon.
2 Fitness limit reached and constraint violation less than
options.TolCon.

12-9

ga

3 The value of the fitness function did not change in

options.StallGenLimit generations and constraint violation less
than options.TolCon.
4 Magnitude of step smaller than machine precision and constraint
violation less than options.TolCon.
0 Maximum number of generations exceeded.
-1 Optimization terminated by the output or plot function.
-2 No feasible point found.
-4 Stall time limit exceeded.
-5 Time limit exceeded.
[x,fval,exitflag,output] = ga(...) returns output, a structure

that contains output from each generation and other information about
the performance of the algorithm. The output structure contains the
following fields:
rngstate The state of the MATLAB random number generator,
just before the algorithm started. You can use the values in rngstate
to reproduce the output of ga. See Reproducing Your Results on
page 5-44.
generations The number of generations computed.
funccount The number of evaluations of the fitness function
message The reason the algorithm terminated.
maxconstraint Maximum constraint violation, if any.
[x,fval,exitflag,output,population] = ga(...) returns the
matrix, population, whose rows are the final population.
[x,fval,exitflag,output,population,scores] = ga(...) returns
scores the scores of the final population.

12-10

ga

Note For problems that use the population type Double Vector (the
default), ga does not accept functions whose inputs are of type complex.
To solve problems involving complex data, write your functions so that
they accept real vectors, by separating the real and imaginary parts.

Example

Given the following inequality constraints and lower bounds

1 1
2
1 2 x1 2 ,

x
2 1 2 3
x1 0, x2 0,
the following code finds the minimum of the function, lincontest6,
that is provided your software:
A = [1 1; -1 2; 2 1];
b = [2; 2; 3];
lb = zeros(2,1);
[x,fval,exitflag] = ga(@lincontest6,...
2,A,b,[],[],lb)
Optimization terminated:
average change in the fitness value less than
options.TolFun.
x =
0.7794

1.2205

fval =
-8.03916
exitflag =
1

12-11

ga

References

[1] Goldberg, David E., Genetic Algorithms in Search, Optimization &

Machine Learning, Addison-Wesley, 1989.
[2] A. R. Conn, N. I. M. Gould, and Ph. L. Toint. A Globally Convergent
Augmented Lagrangian Algorithm for Optimization with General
Constraints and Simple Bounds, SIAM Journal on Numerical Analysis,
Volume 28, Number 2, pages 545572, 1991.
[3] A. R. Conn, N. I. M. Gould, and Ph. L. Toint. A Globally Convergent
Augmented Lagrangian Barrier Algorithm for Optimization with
General Inequality Constraints and Simple Bounds, Mathematics of
Computation, Volume 66, Number 217, pages 261288, 1997.

See Also

12-12

gaoptimget, gaoptimset, patternsearch, simulannealbnd

gamultiobj

Purpose

Find minima of multiple functions using genetic algorithm

Syntax

X = gamultiobj(FITNESSFCN,NVARS)
X = gamultiobj(FITNESSFCN,NVARS,A,b)
X = gamultiobj(FITNESSFCN,NVARS,A,b,Aeq,beq)
X = gamultiobj(FITNESSFCN,NVARS,A,b,Aeq,beq,LB,UB)
X = gamultiobj(FITNESSFCN,NVARS,A,b,Aeq,beq,LB,UB,options)
X = gamultiobj(problem)
[X,FVAL] = gamultiobj(FITNESSFCN,NVARS, ...)
[X,FVAL,EXITFLAG] = gamultiobj(FITNESSFCN,NVARS, ...)
[X,FVAL,EXITFLAG,OUTPUT] = gamultiobj(FITNESSFCN,NVARS, ...)
[X,FVAL,EXITFLAG,OUTPUT,POPULATION] = gamultiobj(FITNESSFCN,
...)
[X,FVAL,EXITFLAG,OUTPUT,POPULATION,
SCORE] = gamultiobj(FITNESSFCN, ...)

Description

gamultiobj implements the genetic algorithm at the command line to

minimize a multicomponent objective function.

X = gamultiobj(FITNESSFCN,NVARS) finds a local Pareto set X of the
objective functions defined in FITNESSFCN. NVARS is the dimension of
the optimization problem (number of decision variables). FITNESSFCN
accepts a vector X of size 1-by-NVARS and returns a vector of size
1-by-numberOfObjectives evaluated at a decision variable. X is a
matrix with NVARS columns. The number of rows in X is the same as the
number of Pareto solutions. All solutions in a Pareto set are equally
optimal; it is up to the designer to select a solution in the Pareto set
depending on the application.
X = gamultiobj(FITNESSFCN,NVARS,A,b) finds a local Pareto set X of
the objective functions defined in FITNESSFCN, subject to the linear

inequalities A x b . Linear constraints are supported only for the

default PopulationType option ('doubleVector'). Other population
types, e.g., 'bitString' and 'custom', are not supported.
X = gamultiobj(FITNESSFCN,NVARS,A,b,Aeq,beq) finds a local
Pareto set X of the objective functions defined in FITNESSFCN, subject

12-13

gamultiobj

to the linear equalities Aeq x = beq as well as the linear inequalities

A x b . (Set A=[] and b=[] if no inequalities exist.) Linear
constraints are supported only for the default PopulationType option
('doubleVector'). Other population types, e.g., 'bitString' and
'custom', are not supported.
X = gamultiobj(FITNESSFCN,NVARS,A,b,Aeq,beq,LB,UB) defines
a set of lower and upper bounds on the design variables X so that

a local Pareto set is found in the range LB x UB . Use empty

matrices for LB and UB if no bounds exist. Set LB(i) = -Inf if X(i)
is unbounded below; set UB(i) = Inf if X(i) is unbounded above.
Bound constraints are supported only for the default PopulationType
option ('doubleVector'). Other population types, e.g., 'bitString'
and 'custom', are not supported.
X = gamultiobj(FITNESSFCN,NVARS,A,b,Aeq,beq,LB,UB,options)
finds a Pareto set X with the default optimization parameters replaced
by values in the structure options. options can be created with the
gaoptimset function.
X = gamultiobj(problem) finds the Pareto set for problem, where
problem is a structure containing the following fields:

12-14

fitnessfcn

Fitness functions

nvars

Number of design variables

Aineq

A matrix for linear inequality constraints

bineq

b vector for linear inequality constraints

Aeq

A matrix for linear equality constraints

beq

b vector for linear equality constraints

lb

Lower bound on x

ub

Upper bound on x

gamultiobj

rngstate

Optional field to reset the state of the

random number generator

options

Options structure created using gaoptimset

Create the structure problem by exporting a problem from Optimization

Tool, as described in Importing and Exporting Your Work in the
Optimization Toolbox documentation.
[X,FVAL] = gamultiobj(FITNESSFCN,NVARS, ...) returns a matrix
FVAL, the value of all the objective functions defined in FITNESSFCN
at all the solutions in X. FVAL has numberOfObjectives columns and
same number of rows as does X.
[X,FVAL,EXITFLAG] = gamultiobj(FITNESSFCN,NVARS, ...) returns
EXITFLAG, which describes the exit condition of gamultiobj. Possible
values of EXITFLAG and the corresponding exit conditions are listed

in this table.
EXITFLAG Exit Condition
Value
1

Average change in value of the spread of Pareto set

over options.StallGenLimit generations less than
options.TolFun

Maximum number of generations exceeded

-1

Optimization terminated by the output or by the plot

function

-2

No feasible point found

-5

Time limit exceeded

[X,FVAL,EXITFLAG,OUTPUT] = gamultiobj(FITNESSFCN,NVARS,
...) returns a structure OUTPUT with the following fields:

12-15

gamultiobj

OUTPUT Field

Meaning

rngstate

State of the MATLAB random number generator,

just before the algorithm started. You can use the
values in rngstate to reproduce the output of ga.
See Reproducing Your Results on page 5-44.

generations

Total number of generations, excluding HybridFcn

iterations

funccount

Total number of function evaluations

maxconstraint

Maximum constraint violation, if any

message

gamultiobj termination message

[X,FVAL,EXITFLAG,OUTPUT,POPULATION] =
gamultiobj(FITNESSFCN, ...) returns the final POPULATION at

termination.
[X,FVAL,EXITFLAG,OUTPUT,POPULATION,SCORE] =
gamultiobj(FITNESSFCN, ...) returns the SCORE of the
final POPULATION.

Example

This example optimizes two objectives defined by Schaffers second

function: a vector-valued function of two components and one input
argument. The Pareto front is disconnected. Define this function in a
file:
function y = schaffer2(x) % y has two columns
% Initialize y for two objectives and for all x
y = zeros(length(x),2);
% Evaluate first objective.
% This objective is piecewise continuous.
for i = 1:length(x)
if x(i) <= 1
y(i,1) = -x(i);
elseif x(i) <=3

12-16

gamultiobj

y(i,1) = x(i) -2;

elseif x(i) <=4
y(i,1) = 4 - x(i);
else
y(i,1) = x(i) - 4;
end
end
% Evaluate second objective
y(:,2) = (x -5).^2;

First, plot the two objectives:

x = -1:0.1:8;
y = schaffer2(x);
plot(x,y(:,1),'.r'); hold on
plot(x,y(:,2),'.b');

The two component functions compete in the range [1, 3] and [4, 5]. But
the Pareto-optimal front consists of only two disconnected regions: [1, 2]
and [4, 5]. This is because the region [2, 3] is inferior to [1, 2].
Next, impose a bound constraint on x, 5 x 10 setting
lb = -5;
ub = 10;

The best way to view the results of the genetic algorithm is to visualize
the Pareto front directly using the @gaplotpareto option. To optimize
Schaffers function, a larger population size than the default (15) is
needed, because of the disconnected front. This example uses 60. Set
the optimization options as:
options = gaoptimset('PopulationSize',60,'PlotFcns',...
@gaplotpareto);

12-17

gamultiobj

Now call gamultiobj, specifying one independent variable and only the
bound constraints:
[x,f,exitflag] = gamultiobj(@schaffer2,1,[],[],[],[],...
lb,ub,options);
Optimization terminated: average change in the spread of
Pareto solutions less than options.TolFun.
exitflag
exitflag = 1

The vectors x, f(:,1), and f(:,2) respectively contain the Pareto set
and both objectives evaluated on the Pareto set.

Demos
The gamultiobjfitness demo solves a simple problem with one
decision variable and two objectives.
The gamultiobjoptionsdemo demo shows how to set options for
multiobjective optimization.

Algorithm

gamultiobj uses a controlled elitist genetic algorithm (a variant of

NSGA-II [1]). An elitist GA always favors individuals with better fitness
value (rank). A controlled elitist GA also favors individuals that can
help increase the diversity of the population even if they have a lower
fitness value. It is important to maintain the diversity of population
for convergence to an optimal Pareto front. Diversity is maintained
by controlling the elite members of the population as the algorithm
progresses. Two options, ParetoFraction and DistanceFcn, control the
elitism. ParetoFraction limits the number of individuals on the Pareto
front (elite members). The distance function, selected by DistanceFcn,
helps to maintain diversity on a front by favoring individuals that are
relatively far away on the front.

References

[1] Deb, Kalyanmoy. Multi-Objective Optimization Using Evolutionary

Algorithms. John Wiley & Sons, 2001.

12-18

gamultiobj

See Also

ga, gaoptimget, gaoptimset, patternsearch, @ (Special

Characters), rand, randn

12-19

gaoptimget

Purpose

Obtain values of genetic algorithm options structure

Syntax

val = gaoptimget(options, 'name')

val = gaoptimget(options, 'name', default)

Description

val = gaoptimget(options, 'name') returns the value of the

parameter name from the genetic algorithm options structure options.
gaoptimget(options, 'name') returns an empty matrix [] if the
value of name is not specified in options. It is only necessary to type
enough leading characters of name to uniquely identify it. gaoptimget

ignores case in parameter names.

val = gaoptimget(options, 'name', default) returns the 'name'

parameter, but will return the default value if the name parameter is
not specified (or is []) in options.

See Also

For more about these options, see Genetic Algorithm Options on page
9-29.
ga, gamultiobj, gaoptimset

12-20

gaoptimset

Purpose

Create genetic algorithm options structure

Syntax

gaoptimset
options = gaoptimset
options = gaoptimset(@ga)
options = gaoptimset(@gamultiobj)
options = gaoptimset('param1',value1,'param2',value2,...)
options = gaoptimset(oldopts,'param1',value1,...)
options = gaoptimset(oldopts,newopts)

Description

gaoptimset with no input or output arguments displays a complete list

of parameters with their valid values.
options = gaoptimset (with no input arguments) creates a structure
called options that contains the options, or parameters, for the genetic
algorithm and sets parameters to [], indicating default values will
be used.
options = gaoptimset(@ga) creates a structure called options that
contains the default options for the genetic algorithm.
options = gaoptimset(@gamultiobj) creates a structure called
options that contains the default options for gamultiobj.
options = gaoptimset('param1',value1,'param2',value2,...)
creates a structure called options and sets the value of 'param1' to
value1, 'param2' to value2, and so on. Any unspecified parameters are

set to their default values. It is sufficient to type only enough leading

characters to define the parameter name uniquely. Case is ignored
for parameter names.
options = gaoptimset(oldopts,'param1',value1,...) creates a
copy of oldopts, modifying the specified parameters with the specified
values.
options = gaoptimset(oldopts,newopts) combines an existing
options structure, oldopts, with a new options structure, newopts.
Any parameters in newopts with nonempty values overwrite the
corresponding old parameters in oldopts.

12-21

gaoptimset

Options

12-22

The following table lists the options you can set with gaoptimset. See
Genetic Algorithm Options on page 9-29 for a complete description of
these options and their values. Values in {} denote the default value.
You can also view the optimization parameters and defaults by typing
gaoptimset at the command line.

Option

Description

Values

CreationFcn

Handle to the function

that creates the initial
population

@gacreationuniform |
@gacreationlinearfeasible

CrossoverFcn

Handle to the function that

the algorithm uses to create
crossover children

@crossoverheuristic |
{@crossoverscattered} |
@crossoverintermediate
| @crossoversinglepoint
| @crossovertwopoint |
@crossoverarithmetic

CrossoverFraction

The fraction of the

population at the next
generation, not including
elite children, that is created
by the crossover function

Positive scalar | {0.8}

Display

Level of display

'off' | 'iter' | 'diagnose' |

{'final'}

DistanceMeasureFcn

Handle to the function that

computes distance measure
of individuals, computed in
decision variable or design
space (genotype) or in
function space (phenotype)

{@distancecrowding,'phenotype'}

gaoptimset

Option

Description

Values

EliteCount

Positive integer specifying

how many individuals in
the current generation are
guaranteed to survive to the
next generation. Not used
in gamultiobj.

Positive integer | {2}

FitnessLimit

Scalar. If the fitness

function attains the value
of FitnessLimit, the
algorithm halts.

Scalar | {-Inf}

FitnessScalingFcn

Handle to the function that

scales the values of the
fitness function

@fitscalingshiftlinear
| @fitscalingprop
| @fitscalingtop |
{@fitscalingrank}

Generations

Positive integer specifying

the maximum number
of iterations before the
algorithm halts

Positive integer |{100}

HybridFcn

Handle to a function that

continues the optimization
after ga terminates

Function handle | @fminsearch

| @patternsearch | @fminunc |
@fmincon | {[]}

or

or

Cell array specifying the

hybrid function and its
options structure

1-by-2 cell array | {@solver,

hybridoptions}, where solver
= fminsearch, patternsearch,
fminunc, or fmincon {[]}

Initial value of penalty

parameter

Positive scalar | {10}

InitialPenalty

12-23

gaoptimset

12-24

Option

Description

Values

InitialPopulation

Initial population used to

seed the genetic algorithm;
can be partial

Matrix | {[]}

InitialScores

Initial scores used to

determine fitness; can be
partial

Column vector | {[]}

MigrationDirection

Direction of migration

'both' | {'forward'}

MigrationFraction

Scalar between 0 and 1

specifying the fraction
of individuals in each
subpopulation that migrates
to a different subpopulation

Scalar | {0.2}

MigrationInterval

Positive integer specifying

the number of generations
that take place between
migrations of individuals
between subpopulations

Positive integer | {20}

MutationFcn

Handle to the function that

produces mutation children

@mutationuniform |
@mutationadaptfeasible |
{@mutationgaussian}

OutputFcns

Functions that ga calls at

each iteration

@gaoutputgen | {[]}

ParetoFraction

Scalar between 0 and 1

specifying the fraction of
individuals to keep on the
first Pareto front while the
solver selects individuals
from higher fronts

Scalar | {0.35}

PenaltyFactor

Penalty update parameter

Positive scalar | {100}

gaoptimset

Option

Description

Values

PlotFcns

Array of handles to
functions that plot data
computed by the algorithm

@gaplotbestf |
@gaplotbestindiv |
@gaplotdistance |
@gaplotexpectation
| @gaplotgeneology
| @gaplotselection
| @gaplotrange |
@gaplotscorediversity
| @gaplotscores |
@gaplotstopping | {[]}

PlotInterval

Positive integer specifying

the number of generations
between consecutive calls to
the plot functions

Positive integer | {1}

PopInitRange

Matrix or vector specifying

the range of the individuals
in the initial population

Matrix or vector | [0;1]

PopulationSize

Size of the population

Positive integer | {20}

PopulationType

String describing the data

type of the population

'bitstring' | 'custom' |
{'doubleVector'}

Note that linear and nonlinear

constraints are not satisfied
when PopulationType is set to
'bitString' or 'custom'.
SelectionFcn

Handle to the function that

selects parents of crossover
and mutation children

@selectionremainder |
@selectionuniform |
{@selectionstochunif}
| @selectionroulette |
@selectiontournament

12-25

gaoptimset

12-26

Option

Description

Values

StallGenLimit

Positive integer. The

algorithm stops if there
is no improvement in
the objective function for
StallGenLimit consecutive
generations.

Positive integer | {50}

StallTimeLimit

Positive scalar. The

algorithm stops if there
is no improvement in
the objective function for
StallTimeLimit seconds.

Positive scalar | {Inf}

TimeLimit

Positive scalar. The

algorithm stops after
running for TimeLimit
seconds.

Positive scalar | {Inf}

TolCon

Positive scalar. TolCon

is used to determine the
feasibility with respect to
nonlinear constraints.

Positive scalar | {1e-6}

TolFun

Positive scalar. The

algorithm runs until the
cumulative change in the
fitness function value over
StallGenLimit is less than
TolFun.

Positive scalar | {1e-6}

UseParallel

Compute fitness functions

of a population in parallel.

'always' | {'never'}

Vectorized

String specifying whether

the computation of the
fitness function is vectorized

'on' | {'off'}

gaoptimset

See Also

For more about these options, see Genetic Algorithm Options on page
9-29.
ga, gamultiobj, gaoptimget

12-27

GlobalOptimSolution class

Purpose

Optimization solution

Description

Information on a local minimum, including location, objective function

value, and start point or points that lead to the minimum.
GlobalSearch and MultiStart generate a vector of
GlobalOptimSolution objects. The vector is ordered by

objective function value, from lowest (best) to highest (worst).

Construction

When you run them, GlobalSearch and MultiStart create

GlobalOptimSolution objects as output.

Properties

Exitflag

An integer describing the result of the local solver run.

For the meaning of the exit flag, see the description in the
appropriate local solver function reference page:
fmincon
fminunc
lsqcurvefit
lsqnonlin
Fval

Objective function value at the solution.

Output

Output structure returned by the local solver.

X

Solution point, with the same dimensions as the initial point.

X0

Cell array of start points that led to the solution.

12-28

GlobalOptimSolution class

Copy
Semantics

Value. To learn how value classes affect copy operations, see Copying
Objects in the MATLAB Programming Fundamentals documentation.

Examples

Use MultiStart to create a vector of GlobalOptimSolution objects:

ms = MultiStart;
sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...
+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3]);
[xmin,fmin,flag,outpt,allmins] = run(ms,problem,30);
allmins is the vector of GlobalOptimSolution objects:
allmins
allmins =
1x30 GlobalOptimSolution
Properties:
X
Fval
Exitflag
Output
X0

See Also

MultiStart | MultiStart.run | GlobalSearch | GlobalSearch.run

Tutorials

Multiple Solutions on page 3-24

Example: Visualizing the Basins of Attraction on page 3-32

How To

Class Attributes
Property Attributes

12-29

GlobalSearch class

Purpose

Find global minimum

Description

A GlobalSearch object contains properties (options) that affect how

the run method searches for a global minimum, or generates a
GlobalOptimSolution object.

Construction

gs = GlobalSearch constructs a new global search optimization solver

with its properties set to the defaults.
gs = GlobalSearch('PropertyName',PropertyValue,...)

constructs the object using options, specified as property name and

value pairs.
gs = GlobalSearch(oldgs,'PropertyName',PropertyValue,...)
constructs a copy of the GlobalSearch solver oldgs. The gs object has

the named properties altered with the specified values.

gs = GlobalSearch(ms) constructs gs, a GlobalSearch solver, with
common parameter values from the ms MultiStart solver.

Properties

BasinRadiusFactor

A basin radius decreases after MaxWaitCycle consecutive start

points are within the basin. The basin radius decreases by a
factor of 1 BasinRadiusFactor.
Set BasinRadiusFactor to 0 to disable updates of the basin
radius.
Default: 0.2
Display

Detail level of iterative display. Possible values:

'final' Report summary results after run finishes.
'iter' Report results after the initial fmincon run, after
Stage 1, after every 200 start points, and after every run of
fmincon, in addition to the final summary.

12-30

GlobalSearch class

'off' No display.
Default: 'final'
DistanceThresholdFactor

A multiplier for determining whether a trial point is in an existing

basin of attraction. For details, see Examine Stage 2 Trial Point
to See if fmincon Runs on page 3-39.
Default: 0.75
MaxTime

Time in seconds for a run. GlobalSearch halts when it sees

MaxTime seconds have passed since the beginning of the run.
Default: Inf
MaxWaitCycle

A positive integer tolerance appearing in several points in the

algorithm:
If the observed penalty function of MaxWaitCycle consecutive
trial points is at least the penalty threshold, then raise the
penalty threshold (see PenaltyThresholdFactor).
If MaxWaitCycle consecutive trial points are in a basin, then
update that basins radius (see BasinRadiusFactor).
Default: 20
NumStageOnePoints

Number of start points in Stage 1. For details, see Obtain Stage

1 Start Point, Run on page 3-38.
Default: 200

12-31

GlobalSearch class

NumTrialPoints

Number of potential start points to examine in addition to x0 from

the problem structure. GlobalSearch runs only those potential
start points that pass several tests. For more information, see
GlobalSearch Algorithm on page 3-37.
Default: 1000
PenaltyThresholdFactor

Determines increase in the penalty threshold. For details, see

React to Large Counter Values.
Default: 0.2
StartPointsToRun

Directs the solver to exclude certain start points from being run:
all Accept all start points.
bounds Reject start points that do not satisfy bounds.
bounds-ineqs Reject start points that do not satisfy bounds
or inequality constraints.
GlobalSearch checks the StartPointsToRun property only
during Stage 2 of the GlobalSearch algorithm (the main loop).

For more information, see GlobalSearch Algorithm on page 3-37.

Default: 'all'
TolFun

Describes how close two objective function values must be for

solvers to consider them identical for creating the vector of local
solutions. Solvers consider two solutions identical if they are
within TolX distance of each other and have objective function
values within TolFun of each other. If both conditions are not

12-32

GlobalSearch class

met, solvers report the solutions as distinct. Set TolFun to 0 to

obtain the results of every local solver run. Set TolFun to a larger
value to have fewer results.
Default: 1e-6
TolX

Describes how close two points must be for solvers to consider

them identical for creating the vector of local solutions. Solvers
compute the distance between a pair of points with norm, the
Euclidean distance. Solvers consider two solutions identical if
they are within TolX distance of each other and have objective
function values within TolFun of each other. If both conditions
are not met, solvers report the solutions as distinct. Set TolX to
0 to obtain the results of every local solver run. Set TolX to a
larger value to have fewer results.
Default: 1e-6

Methods

run

Find global minimum

Copy
Semantics

Value. To learn how value classes affect copy operations, see Copying
Objects in the MATLAB Programming Fundamentals documentation.

Examples

Solve a problem using a default GlobalSearch object:

opts = optimset('Algorithm','interior-point');
problem = createOptimProblem('fmincon','objective',...
@(x) x.^2 + 4*sin(5*x),'x0',3,'lb',-5,'ub',5,'options',opts);
gs = GlobalSearch;
[x,f] = run(gs,problem)

Algorithm

A detailed description of the algorithm appears in GlobalSearch

Algorithm on page 3-37. Ugray et al. [1] describes both the algorithm
and the scatter-search method of generating trial points.

12-33

GlobalSearch class

References

[1] Ugray, Zsolt, Leon Lasdon, John Plummer, Fred Glover, James
Kelly, and Rafael Mart. Scatter Search and Local NLP Solvers: A
Multistart Framework for Global Optimization. INFORMS Journal on
Computing, Vol. 19, No. 3, 2007, pp. 328340.

See Also

MultiStart | createOptimProblem | GlobalOptimSolution

Tutorials

Chapter 3, Using GlobalSearch and MultiStart

GlobalSearch and MultiStart Examples on page 3-63

How To

Class Attributes
Property Attributes

12-34

CustomStartPointSet.list

Purpose

List custom start points in set

Syntax

tpts = list(CS)

Description

tpts = list(CS) returns the matrix of start points in the CS

CustomStartPoints object.

Input
Arguments

CS

Output
Arguments

tpts

Examples

Create a CustomStartPointSet containing 40 seven-dimensional

normally distributed points, then use list to get the matrix of points
from the object:

A CustomStartPointSet object.

Matrix of start points. The rows of tpts represent the start points.

startpts = randn(40,7) % 40 seven-dimensional start points

cs = CustomStartPointSet(startpts); % cs is an object
startpts2 = list(cs) % startpts2 = startpts

See Also

CustomStartPointSet | createOptimProblem

Tutorials

CustomStartPointSet Object for Start Points on page 3-17

12-35

RandomStartPointSet.list

Purpose

Generate start points

Syntax

points = list(RandSet,problem)

Description

points = list(RandSet,problem) generates pseudorandom start

points using the parameters in the RandSet RandomStartPointSet
object, and information from the problem problem structure.

Input
Arguments

RandSet

A RandomStartPointSet object. This contains parameters for

generating the points: number of points, random stream, and
artificial bounds.
problem

An optimization problem structure. list generates points

uniformly within the bounds of the problem structure. If a
component is unbounded, list uses the artificial bounds from
RandSet. list takes the dimension of the points from the x0 field
in problem.

Output
Arguments

points

Examples

Create a matrix representing 40 seven-dimensional start points:

A k-by-n matrix. The number of rows k is the number of start

points that RandSet specifies. The number of columns n is the
dimension of the start points. n is equal to the number of elements
in the x0 field in problem. The MultiStart algorithm uses each
row of points as an initial point in an optimization.

rs = RandomStartPointSet('NumStartPoints',40); % 40 points
problem = createOptimProblem('fminunc','x0',ones(7,1),...
'objective',@rosenbrock);
ptmatrix = list(rs,problem); % matrix values between
% -1000 and 1000 since those are the default bounds
% for unconstrained dimensions

12-36

RandomStartPointSet.list

Algorithm

The list method generates a pseudorandom random matrix. Each row

of the matrix represents a start point to run. list generates points that
satisfy the bounds in problem. If lb is the vector of lower bounds, ub is
the vector of upper bounds, there are n dimensions in a point, and there
are k rows in the matrix, the random matrix is
lb + (ub - lb).*rand(k,n)

If a component has no bounds, RandomStartPointSet uses

a lower bound of -ArtificialBound, and an upper bound of
ArtificialBound.
If a component has a lower bound lb, but no upper
bound, RandomStartPointSet uses an upper bound of
lb + 2*ArtificialBound.
Similarly, if a component has an upper bound ub, but no
lower bound, RandomStartPointSet uses a lower bound of
ub - 2*ArtificialBound.
The default value of ArtificialBound is 1000.

See Also

RandomStartPointSet | createOptimProblem | MultiStart

Tutorials

RandomStartPointSet Object for Start Points on page 3-17

12-37

MultiStart class

Purpose

Find multiple local minima

Description

A MultiStart object contains properties (options) that affect how

the run method repeatedly runs a local solver, or generates a
GlobalOptimSolution object.

Construction

MS = MultiStart constructs MS, a MultiStart solver with its properties

set to the defaults.
MS = MultiStart('PropertyName',PropertyValue,...) constructs
MS using options, specified as property name and value pairs.
MS = MultiStart(oldMS,'PropertyName',PropertyValue,...)
creates a copy of the oldMS MultiStart solver, with the named

properties changed to the specified values.

MS = MultiStart(GS) constructs MS, a MultiStart solver, with
common parameter values from the GS GlobalSearch solver.

Properties

Display

Detail level of the output to the Command Window:

'final' Report summary results after run finishes.
'iter' Report results after each local solver run, in addition
to the final summary.
'off' No display.
Default: final
MaxTime

Tolerance on the time MultiStart runs. MultiStart and its

local solvers halt when they see MaxTime seconds have passed
since the beginning of the run. Time means wall clock as opposed
to processor cycles.
Default: Inf

12-38

MultiStart class

StartPointsToRun

Directs the solver to exclude certain start points from being run:
all Accept all start points.
bounds Reject start points that do not satisfy bounds.
bounds-ineqs Reject start points that do not satisfy bounds
or inequality constraints.
Default: all
TolFun

Describes how close two objective function values must be for

solvers to consider them identical for creating the vector of local
solutions. Solvers consider two solutions identical if they are
within TolX distance of each other and have objective function
values within TolFun of each other. If both conditions are not
met, solvers report the solutions as distinct. Set TolFun to 0 to
obtain the results of every local solver run. Set TolFun to a larger
value to have fewer results.
Default: 1e-6
TolX

Describes how close two points must be for solvers to consider

them identical for creating the vector of local solutions. Solvers
compute the distance between a pair of points with norm, the
Euclidean distance. Solvers consider two solutions identical if
they are within TolX distance of each other and have objective
function values within TolFun of each other. If both conditions
are not met, solvers report the solutions as distinct. Set TolX to
0 to obtain the results of every local solver run. Set TolX to a
larger value to have fewer results.
Default: 1e-6

12-39

MultiStart class

UseParallel

Distribute local solver calls to multiple processors:

always Distribute the local solver calls to multiple
processors.
never Cannot not run in parallel.
Default: never

Methods

run

Run local solver from multiple

points

Copy
Semantics

Value. To learn how value classes affect copy operations, see Copying
Objects in the MATLAB Programming Fundamentals documentation.

Examples

Run MultiStart on 20 instances of a problem using the fmincon sqp

algorithm:
opts = optimset('Algorithm','sqp');
problem = createOptimProblem('fmincon','objective',...
@(x) x.^2 + 4*sin(5*x),'x0',3,'lb',-5,'ub',5,'options',opts);
ms = MultiStart;
[x,f] = run(ms,problem,20)

Algorithm

A detailed description of the algorithm appears in MultiStart

Algorithm on page 3-41.

See Also

GlobalSearch | createOptimProblem | RandomStartPointSet |

CustomStartPointSet | GlobalOptimSolution

Tutorials

Chapter 3, Using GlobalSearch and MultiStart

GlobalSearch and MultiStart Examples on page 3-63
Chapter 8, Parallel Processing

12-40

MultiStart class

How To

Class Attributes
Property Attributes

12-41

patternsearch

Purpose

Find minimum of function using pattern search

Syntax

x = patternsearch(@fun,x0)
x = patternsearch(@fun,x0,A,b)
x = patternsearch(@fun,x0,A,b,Aeq,beq)
x = patternsearch(@fun,x0,A,b,Aeq,beq,LB,UB)
x = patternsearch(@fun,x0,A,b,Aeq,beq,LB,UB,nonlcon)
x = patternsearch(@fun,x0,A,b,Aeq,beq,LB,UB,nonlcon,options)
x = patternsearch(problem)
[x,fval] = patternsearch(@fun,x0, ...)
[x,fval,exitflag] = patternsearch(@fun,x0, ...)
[x,fval,exitflag,output] = patternsearch(@fun,x0, ...)

Description

patternsearch finds the minimum of a function using a pattern search.

x = patternsearch(@fun,x0) finds the local minimum, x, to the
MATLAB function, fun, that computes the values of the objective
function f(x), and x0 is an initial point for the pattern search algorithm.
The function patternsearch accepts the objective function as a function
handle of the form @fun. The function fun accepts a vector input and

returns a scalar function value.

Note To write a function with additional parameters to the
independent variables that can be called by patternsearch, see the
section on Passing Extra Parameters in the Optimization Toolbox
documentation.
x = patternsearch(@fun,x0,A,b) finds a local minimum x to the
function fun, subject to the linear inequality constraints represented

in matrix form by Ax b .
If the problem has m linear inequality constraints and n variables, then
A is a matrix of size m-by-n.
b is a vector of length m.

12-42

patternsearch

x = patternsearch(@fun,x0,A,b,Aeq,beq) finds a local minimum x

to the function fun, starting at x0, and subject to the constraints

Ax b
Aeq x = beq
where Aeq x = beq represents the linear equality constraints in
matrix form. If the problem has r linear equality constraints and n
variables, then
Aeq is a matrix of size r-by-n.
beq is a vector of length r.
If there are no inequality constraints, pass empty matrices, [], for
A and b.
x = patternsearch(@fun,x0,A,b,Aeq,beq,LB,UB) defines a set of
lower and upper bounds on the design variables, x, so that a solution is

found in the range LB < x < UB . If the problem has n variables, LB

and UB are vectors of length n. If LB or UB is empty (or not provided),
it is automatically expanded to -Inf or Inf, respectively. If there are
no inequality or equality constraints, pass empty matrices for A, b, Aeq
and beq.
x = patternsearch(@fun,x0,A,b,Aeq,beq,LB,UB,nonlcon) subjects
the minimization to the constraints defined in nonlcon, a nonlinear
constraint function. The function nonlcon accepts x and returns

the vectors C and Ceq, representing the nonlinear inequalities and

equalities respectively. fmincon minimizes fun such that C(x) 0 and
Ceq(x) = 0. (Set LB = [] and UB = [] if no bounds exist.)
x =
patternsearch(@fun,x0,A,b,Aeq,beq,LB,UB,nonlcon,options)
minimizes fun with the default optimization parameters replaced
by values in options. The structure options can be created
using psoptimset.
x = patternsearch(problem) finds the minimum for problem, where
problem is a structure containing the following fields:

12-43

patternsearch

objective Objective function

X0 Starting point
Aineq Matrix for linear inequality constraints
bineq Vector for linear inequality constraints
Aeq Matrix for linear equality constraints
beq Vector for linear equality constraints
lb Lower bound for x
ub Upper bound for x
nonlcon Nonlinear constraint function
Solver 'patternsearch'
options Options structure created with psoptimset
rngstate Optional field to reset the state of the random number
generator
Create the structure problem by exporting a problem from the
Optimization Tool, as described in Importing and Exporting Your
Work in the Optimization Toolbox documentation.
Note problem must have all the fields as specified above.
[x,fval] = patternsearch(@fun,x0, ...) returns the value of the

objective function fun at the solution x.

[x,fval,exitflag] = patternsearch(@fun,x0, ...) returns
exitflag, which describes the exit condition of patternsearch.
Possible values of exitflag and the corresponding conditions are

12-44

patternsearch

Magnitude of mesh size is less than specified tolerance and

constraint violation less than options.TolCon.

Change in x less than the specified tolerance and constraint

violation less than options.TolCon.

Magnitude of step smaller than machine precision and

constraint violation less than options.TolCon.

Maximum number of function evaluations or iterations

reached.

-1

Optimization terminated by the output or plot function.

-2

No feasible point found.

[x,fval,exitflag,output] = patternsearch(@fun,x0, ...)

returns a structure output containing information about the search.

The output structure contains the following fields:

function Objective function
problemtype Type of problem: unconstrained, bound constrained
or linear constrained
pollmethod Polling technique
searchmethod Search technique used, if any
iterations Total number of iterations
funccount Total number of function evaluations
meshsize Mesh size at x
maxconstraint Maximum constraint violation, if any
message Reason why the algorithm terminated

12-45

patternsearch

Note patternsearch does not accepts functions whose inputs are of

type complex. To solve problems involving complex data, write your
functions so that they accept real vectors, by separating the real and
imaginary parts.

Example

Given the following constraints

1 1
2
1 2 x1 2 ,

x
2 1 2 3
x1 0, x2 0,
the following code finds the minimum of the function, lincontest6,
that is provided with your software:
A = [1 1; -1 2; 2 1];
b = [2; 2; 3];
lb = zeros(2,1);
[x,fval,exitflag] = patternsearch(@lincontest6,[0 0],...
A,b,[],[],lb)
Optimization terminated: mesh size less than
options.TolMesh.
x =
0.6667
fval =
-8.2222
exitflag =
1

12-46

1.3333

patternsearch

References

[1] Audet, Charles and J. E. Dennis Jr. Analysis of Generalized

Pattern Searches. SIAM Journal on Optimization, Volume 13, Number
3, 2003, pp. 889903.
[2] Conn, A. R., N. I. M. Gould, and Ph. L. Toint. A Globally
Convergent Augmented Lagrangian Barrier Algorithm for Optimization
with General Inequality Constraints and Simple Bounds. Mathematics
of Computation, Volume 66, Number 217, 1997, pp. 261288.
[3] Abramson, Mark A. Pattern Search Filter Algorithms for Mixed
Variable General Constrained Optimization Problems. Ph.D. Thesis,
Department of Computational and Applied Mathematics, Rice
University, August 2002.
[4] Abramson, Mark A., Charles Audet, J. E. Dennis, Jr., and Sebastien
Le Digabel. ORTHOMADS: A deterministic MADS instance with
orthogonal directions. SIAM Journal on Optimization, Volume 20,
Number 2, 2009, pp. 948966.
[5] Kolda, Tamara G., Robert Michael Lewis, and Virginia Torczon.
Optimization by direct search: new perspectives on some classical and
modern methods. SIAM Review, Volume 45, Issue 3, 2003, pp. 385482.
[6] Kolda, Tamara G., Robert Michael Lewis, and Virginia Torczon.
A generating set direct search augmented Lagrangian algorithm for
optimization with a combination of general and linear constraints.
Technical Report SAND2006-5315, Sandia National Laboratories,
August 2006.
[7] Lewis, Robert Michael, Anne Shepherd, and Virginia Torczon.
Implementing generating set search methods for linearly constrained
minimization. SIAM Journal on Scientific Computing, Volume 29,
Issue 6, 2007, pp. 25072530.

See Also

optimtool, psoptimget, psoptimset, ga, simulannealbnd

12-47

psoptimget

Purpose

Obtain values of pattern search options structure

Syntax

val = psoptimget(options,'name')
val = psoptimget(options,'name',default)

Description

val = psoptimget(options,'name') returns the value of the

parameter name from the pattern search options structure options.
psoptimget(options, 'name') returns an empty matrix [] if the
value of name is not specified in options. It is only necessary to type
enough leading characters of name to uniquely identify it. psoptimget

ignores case in parameter names.

val = psoptimget(options,'name',default) returns the value of the
parameter name from the pattern search options structure options, but
returns default if the parameter is not specified (as in []) in options.

Example

val = psoptimget(opts,'TolX',1e-4);

returns val = 1e-4 if the TolX property is not specified in opts.

See Also

For more about these options, see Pattern Search Options on page 9-7.
psoptimset, patternsearch

12-48

psoptimset

Purpose

Create pattern search options structure

Syntax

psoptimset
options = psoptimset
options = psoptimset('param1',value1,'param2',value2,...)
options = psoptimset(oldopts,'param1',value1,...)
options = psoptimset(oldopts,newopts)

Description

psoptimset with no input or output arguments displays a complete list

of parameters with their valid values.
options = psoptimset (with no input arguments) creates a structure
called options that contains the options, or parameters, for the pattern
search and sets parameters to their default values.
options = psoptimset('param1',value1,'param2',value2,...)
creates a structure options and sets the value of 'param1' to value1,
'param2' to value2, and so on. Any unspecified parameters are set

to their default values. It is sufficient to type only enough leading

characters to define the parameter name uniquely. Case is ignored
for parameter names.
options = psoptimset(oldopts,'param1',value1,...) creates a
copy of oldopts, modifying the specified parameters with the specified
values.
options = psoptimset(oldopts,newopts) combines an existing
options structure, oldopts, with a new options structure, newopts.
Any parameters in newopts with nonempty values overwrite the
corresponding old parameters in oldopts.

Options

The following table lists the options you can set with psoptimset. See
Pattern Search Options on page 9-7 for a complete description of
the options and their values. Values in {} denote the default value.
You can also view the optimization parameters and defaults by typing
psoptimset at the command line.

12-49

psoptimset

12-50

Option

Description

Values

Cache

With Cache set to 'on',

patternsearch keeps
a history of the mesh
points it polls and does
not poll points close to
them again at subsequent
iterations. Use this option
if patternsearch runs
slowly because it is taking
a long time to compute
the objective function. If
the objective function is
stochastic, it is advised not
to use this option.

'on' | {'off'}

CacheSize

Size of the history

Positive scalar | {1e4}

CacheTol

Positive scalar specifying

how close the current mesh
point must be to a point
in the history in order for
patternsearch to avoid
polling it. Use if 'Cache'
option is set to 'on'.

Positive scalar | {eps}

CompletePoll

Complete poll around

current iterate

'on' | {'off'}

CompleteSearch

Complete poll around

current iterate

'on' | {'off'}

Display

Level of display

'off' | 'iter' | 'diagnose' |

{'final'}

InitialMeshSize

Initial mesh size for

pattern algorithm

Positive scalar | {1.0}

InitialPenalty

Initial value of the penalty

parameter

Positive scalar | {10}

psoptimset

Option

Description

Values

MaxFunEvals

Maximum number
of objective function
evaluations

Positive integer |

Maximum number of
iterations

Positive integer |

MaxMeshSize

Maximum mesh size used

in a poll/search step

Positive scalar | {Inf}

MeshAccelerator

Accelerate convergence
near a minimum

'on' | {'off'}

MeshContraction

Mesh contraction factor,

used when iteration is
unsuccessful

Positive scalar | {0.5}

MeshExpansion

Mesh expansion factor,

expands mesh when
iteration is successful

Positive scalar | {2.0}

MeshRotate

Rotate the pattern before

declaring a point to be
optimum

'off' | {'on'}

OutputFcn

Specifies a user-defined
function that an
optimization function
calls at each iteration

@psoutputhistory | {[]}

PenaltyFactor

Penalty update parameter

Positive scalar | {100}

PlotFcn

Specifies plots of output

from the pattern search

@psplotbestf | @psplotmeshsize |
@psplotfuncount | @psplotbestx |
{[]}

PlotInterval

Specifies that plot

functions will be called
at every interval

{1}

MaxIter

{2000*numberOfVariables}

{100*numberOfVariables}

12-51

psoptimset

12-52

Option

Description

Values

PollingOrder

Order of poll directions in

pattern search

'Random' | 'Success' |
{'Consecutive'}

PollMethod

Polling strategy used in

pattern search

{'GPSPositiveBasis2N'} |
'GPSPositiveBasisNp1'|
'GSSPositiveBasis2N'|
'GSSPositiveBasisNp1'|
'MADSPositiveBasis2N'|
'MADSPositiveBasisNp1'

ScaleMesh

Automatic scaling of
variables

{'on'} | 'off'

SearchMethod

Type of search used in

pattern search

@GPSPositiveBasis2N |
@GPSPositiveBasisNp1 |
@GSSPositiveBasis2N |
@GSSPositiveBasisNp1 |
@MADSPositiveBasis2N |
@MADSPositiveBasisNp1 |
@searchga | @searchlhs |
@searchneldermead | {[]}

TimeLimit

Total time (in seconds)

allowed for optimization

Positive scalar | {Inf}

TolBind

Binding tolerance

Positive scalar | {1e-3}

TolCon

Tolerance on constraints

Positive scalar | {1e-6}

TolFun

Tolerance on function

Positive scalar | {1e-6}

TolMesh

Tolerance on mesh size

Positive scalar | {1e-6}

TolX

Tolerance on variable

Positive scalar | {1e-6}

UseParallel

Compute objective
functions of a poll or
search in parallel.

'always' | {'never'}

Vectorized

Specifies whether
functions are vectorized

'on' | {'off'}

psoptimset

See Also

patternsearch, psoptimget

12-53

GlobalSearch.run

Purpose

Find global minimum

Syntax

x = run(gs,problem)
[x,fval] = run(gs,problem)
[x,fval,exitflag] = run(gs,problem)
[x,fval,exitflag,output] = run(gs,problem)
[x,fval,exitflag,output,solutions] = run(gs,problem)

Description

x = run(gs,problem) finds a point x that solves the optimization

problem described in the problem structure.
[x,fval] = run(gs,problem) returns the value of the objective
function in problem at the point x.
[x,fval,exitflag] = run(gs,problem) returns an exit flag
describing the results of the multiple local searches.
[x,fval,exitflag,output] = run(gs,problem) returns an output

structure describing the iterations of the run.

[x,fval,exitflag,output,solutions] = run(gs,problem) returns
a vector of solutions containing the distinct local minima found during
the run.

Input
Arguments

gs

A GlobalSearch object.
problem

Problem structure. Create problem with createOptimProblem or

by exporting a problem structure from the Optimization Tool.
problem must contain at least the following fields:
solver
objective
x0

12-54

GlobalSearch.run

options Both createOptimProblem and the Optimization

Tool always include an options field in the problem structure.

Output
Arguments

Minimizing point of the objective function.

fval

Objective function value at the minimizer x.

exitflag

Describes the results of the multiple local searches. Values are:

2

At least one local minimum found. Some runs of

the local solver converged.

At least one local minimum found. All runs of

the local solver converged.

No local minimum found. Local solver called at

least once, and at least one local solver exceeded
the MaxIter or MaxFunEvals tolerances.

-1

One or more local solver runs stopped by the

local solver output or plot function.

-2

No feasible local minimum found.

-5

MaxTime limit exceeded.

-8

No solution found. All runs had local solver exit

flag -2 or smaller, not all equal -2.

-10

Failures encountered in user-provided functions.

output

A structure describing the iterations of the run. Fields in the

structure:

12-55

GlobalSearch.run

funcCount

Number of function evaluations.

localSolverIncomplete

Number of local solver runs with 0 exit flag.

localSolverNoSolution

Number of local solver runs with negative exit flag.

localSolverSuccess

Number of local solver runs with positive exit flag.

localSolverTotal

Total number of local solver runs.

message

Exit message.
solutions

A vector of GlobalOptimSolution objects containing the distinct

local solutions found during the run. The vector is sorted by
objective function value; the first element is best (smallest value).
The object contains:

Examples

Solution point returned by the local solver.

Fval

Objective function value at the solution.

Exitflag

Integer describing the result of the local solver

run.

Output

Output structure returned by the local solver.

X0

Cell array of start points that led to the solution.

Use a default GlobalSearch object to solve the six-hump camel back

problem (see Run the Solver on page 3-19):
gs = GlobalSearch;
sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...
+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3]);
[xmin,fmin,flag,outpt,allmins] = run(gs,problem);

12-56

GlobalSearch.run

Algorithm

A detailed description of the algorithm appears in GlobalSearch

Algorithm on page 3-37. Ugray et al. [1] describes both the algorithm
and the scatter-search method of generating trial points.

References

[1] Ugray, Zsolt, Leon Lasdon, John Plummer, Fred Glover, James
Kelly, and Rafael Mart. Scatter Search and Local NLP Solvers: A
Multistart Framework for Global Optimization. INFORMS Journal on
Computing, Vol. 19, No. 3, 2007, pp. 328340.

See Also

GlobalSearch | createOptimProblem | GlobalOptimSolution

Tutorials

Run the Solver on page 3-19

How to Optimize with GlobalSearch and MultiStart on page 3-2
GlobalSearch and MultiStart Examples on page 3-63

12-57

MultiStart.run

Purpose

Run local solver from multiple points

Syntax

x = run(ms,problem,k)
x = run(ms,problem,startpts)
[x,fval] = run(...)
[x,fval,exitflag] = run(...)
[x,fval,exitflag,output] = run(...)
[x,fval,exitflag,output,solutions] = run(...)

Description

x = run(ms,problem,k) runs the ms MultiStart object on the

optimization problem specified in problem for a total of k runs. x is the
point where the lowest function value was found. For the lsqcurvefit
and lsqnonlin solvers, MultiStart minimizes the sum of squares at x,

also known as the residual.

x = run(ms,problem,startpts) runs the ms MultiStart object on
the optimization problem specified in problem using the start points
described in startpts.
[x,fval] = run(...) returns the lowest objective function value fval
at x. For the lsqcurvefit and lsqnonlin solvers, fval contains the

residual.
[x,fval,exitflag] = run(...) returns an exit flag describing the

return condition.
[x,fval,exitflag,output] = run(...) returns an output structure

containing statistics of the run.

[x,fval,exitflag,output,solutions] = run(...) returns a vector

of solutions containing the distinct local minima found during the run.

Input
Arguments

ms
MultiStart object.
problem

12-58

MultiStart.run

Problem structure. Create problem with createOptimProblem or

by exporting a problem structure from the Optimization Tool.
problem must contain the following fields:
solver
objective
x0
options Both createOptimProblem and the Optimization
Tool always include an options field in the problem structure.
k

Number of start points to run. MultiStart generates

k - 1 start points using the same algorithm as list for a
RandomStartPointSet object. MultiStart also uses the x0 point
from the problem structure.
startpts

A CustomStartPointSet or RandomStartPointSet object.

startpts can also be a cell array of these objects.

Output
Arguments

Point at which the objective function attained the lowest value

during the run. For lsqcurvefit and lsqnonlin, the objective
function is the sum of squares, also known as the residual.
fval

Smallest objective function value attained during the run. For

lsqcurvefit and lsqnonlin, the objective function is the sum of
squares, also known as the residual.
exitflag

Integer describing the return condition:

12-59

MultiStart.run

At least one local minimum found. Some runs of

the local solver converged.

At least one local minimum found. All runs of

the local solver converged.

No local minimum found. Local solver called at

least once, and at least one local solver exceeded
the MaxIter or MaxFunEvals tolerances.

-1

One or more local solver runs stopped by the

local solver output or plot function.

-2

No feasible local minimum found.

-5

MaxTime limit exceeded.

-8

No solution found. All runs had local solver exit

flag -2 or smaller, not all equal -2.

-10

Failures encountered in user-provided functions.

output

Structure containing statistics of the optimization. Fields in the

structure:
funcCount

Number of function evaluations.

localSolverIncomplete

Number of local solver runs with 0 exit flag.

localSolverNoSolution

Number of local solver runs with negative exit flag.

localSolverSuccess

Number of local solver runs with positive exit flag.

localSolverTotal

Total number of local solver runs.

message

Exit message.
solutions

A vector of GlobalOptimSolution objects containing the distinct

local solutions found during the run. The vector is sorted by

12-60

MultiStart.run

objective function value; the first element is best (smallest value).

The object contains:

Examples

Solution point returned by the local solver.

Fval

Objective function value at the solution.

Exitflag

Integer describing the result of the local solver

run.

Output

Output structure returned by the local solver.

X0

Cell array of start points that led to the solution.

Use a default MultiStart object to solve the six-hump camel back

problem (see Run the Solver on page 3-19):
ms = MultiStart;
sixmin = @(x)(4*x(1)^2 - 2.1*x(1)^4 + x(1)^6/3 ...
+ x(1)*x(2) - 4*x(2)^2 + 4*x(2)^4);
problem = createOptimProblem('fmincon','x0',[-1,2],...
'objective',sixmin,'lb',[-3,-3],'ub',[3,3]);
[xmin,fmin,flag,outpt,allmins] = run(ms,problem,30);

Algorithm

A detailed description of the algorithm appears in MultiStart

Algorithm on page 3-41.

See Also

MultiStart | createOptimProblem | CustomStartPointSet |

RandomStartPointSet | GlobalOptimSolution

Tutorials

Run the Solver on page 3-19

How to Optimize with GlobalSearch and MultiStart on page 3-2
GlobalSearch and MultiStart Examples on page 3-63

12-61

RandomStartPointSet class

Purpose

Random start points

Description

Describes how to generate a set of pseudorandom points for use with

MultiStart. A RandomStartPointSet object does not contain points. It
contains parameters used in generating the points when MultiStart
runs, or by the list method.

Construction

RS = RandomStartPointSet constructs a default RandomStartPointSet

object.
RS = RandomStartPointSet('PropertyName',PropertyValue,...)

constructs the object using options, specified as property name and

value pairs.
RS =
RandomStartPointSet(OLDRS,'PropertyName',PropertyValue,...)
creates a copy of the OLDRS RandomStartPointSet object, with the

named properties altered with the specified values.

Properties

ArtificialBound

Absolute value of default bounds to use for unbounded problems

(positive scalar).
If a component has no bounds, list uses a lower bound of
-ArtificialBound, and an upper bound of ArtificialBound.
If a component has a lower bound lb, but no upper bound, list
uses an upper bound of lb + 2*ArtificialBound. Similarly, if a
component has an upper bound ub, but no lower bound, list uses
a lower bound of ub - 2*ArtificialBound.
Default: 1000
NumStartPoints

Number of start points to generate (positive integer)

Default: 10

12-62

RandomStartPointSet class

Methods

list

Generate start points

Copy
Semantics

Value. To learn how value classes affect copy operations, see Copying
Objects in the MATLAB Programming Fundamentals documentation.

Examples

Create a RandomStartPointSet object for 40 points, and use list to

generate a point matrix for a seven-dimensional problem:
rs = RandomStartPointSet('NumStartPoints',40); % 40 points
problem = createOptimProblem('fminunc','x0',ones(7,1),...
'objective',@rosenbrock);
ptmatrix = list(rs,problem); % 'list' generates the matrix

See Also

MultiStart | CustomStartPointSet | list

Tutorials

RandomStartPointSet Object for Start Points on page 3-17

How To

Class Attributes
Property Attributes

12-63

saoptimget

Purpose

Values of simulated annealing options structure

Syntax

val = saoptimget(options, 'name')

val = saoptimget(options, 'name', default)

Description

val = saoptimget(options, 'name') returns the value of the

parameter name from the simulated annealing options structure
options. saoptimget(options, 'name') returns an empty matrix []
if the value of name is not specified in options. It is only necessary
to type enough leading characters of name to uniquely identify the
parameter. saoptimget ignores case in parameter names.
val = saoptimget(options, 'name', default) returns the 'name'
parameter, but returns the default value if the 'name' parameter is
not specified (or is []) in options.

Example

opts = saoptimset('TolFun',1e-4);
val = saoptimget(opts,'TolFun');

returns val = 1e-4 for TolFun.

See Also

For more about these options, see Simulated Annealing Options on

page 9-53.
saoptimset, simulannealbnd

12-64

saoptimset

Purpose

Create simulated annealing options structure

Syntax

saoptimset
options = saoptimset
options = saoptimset('param1',value1,'param2',value2,...)
options = saoptimset(oldopts,'param1',value1,...)
options = saoptimset(oldopts,newopts)
options = saoptimset('simulannealbnd')

Description

saoptimset with no input or output arguments displays a complete list

of parameters with their valid values.
options = saoptimset (with no input arguments) creates a structure
called options that contains the options, or parameters, for the
simulated annealing algorithm, with all parameters set to [].
options = saoptimset('param1',value1,'param2',value2,...)
creates a structure options and sets the value of 'param1' to value1,
'param2' to value2, and so on. Any unspecified parameters are set to
[]. It is sufficient to type only enough leading characters to define the

parameter name uniquely. Case is ignored for parameter names. Note

that for string values, correct case and the complete string are required.
options = saoptimset(oldopts,'param1',value1,...) creates a
copy of oldopts, modifying the specified parameters with the specified
values.
options = saoptimset(oldopts,newopts) combines an existing
options structure, oldopts, with a new options structure, newopts.
Any parameters in newopts with nonempty values overwrite the
corresponding old parameters in oldopts.
options = saoptimset('simulannealbnd') creates an options
structure with all the parameter names and default values relevant to
'simulannealbnd'. For example,
saoptimset('simulannealbnd')
ans =

12-65

saoptimset

AnnealingFcn:
TemperatureFcn:
AcceptanceFcn:
TolFun:
StallIterLimit:
MaxFunEvals:
TimeLimit:
MaxIter:
ObjectiveLimit:
Display:
DisplayInterval:
HybridFcn:
HybridInterval:
PlotFcns:
PlotInterval:
OutputFcns:
InitialTemperature:
ReannealInterval:
DataType:

Options

The following table lists the options you can set with saoptimset. See
Chapter 9, Options Reference for a complete description of these
options and their values. Values in {} denote the default value. You
can also view the options parameters by typing saoptimset at the
command line.

Option

Description

Values

AcceptanceFcn

Handle to the function the

algorithm uses to determine
if a new point is accepted

Function handle
|{@acceptancesa}

AnnealingFcn

Handle to the function the

algorithm uses to generate
new points

Function handle |
@annealingboltz |

Type of decision variable

'custom' | {'double'}

DataType

12-66

@annealingfast
@temperatureexp
@acceptancesa
1.0000e-006
'500*numberofvariables'
'3000*numberofvariables'
Inf
Inf
-Inf
'final'
10
[]
'end'
[]
1
[]
100
100
'double'

{@annealingfast}

saoptimset

Option

Description

Values

Display

Level of display

'off' | 'iter' | 'diagnose' |

{'final'}

DisplayInterval

Interval for iterative display

Positive integer | {10}

HybridFcn

Automatically run HybridFcn

(another optimization
function) during or at the end
of iterations of the solver

Function handle | @fminsearch |

@patternsearch | @fminunc |
@fmincon | {[]}

or
1-by-2 cell array | {@solver,
hybridoptions}, where solver
= fminsearch, patternsearch,
fminunc, or fmincon {[]}

Interval (if not 'end' or

'never') at which HybridFcn
is called

Positive integer | 'never' |

InitialTemperature

Initial value of temperature

Positive scalar |{100}

MaxFunEvals

Maximum number of
objective function evaluations
allowed

Positive integer |

MaxIter

Maximum number of
iterations allowed

Positive integer | {Inf}

ObjectiveLimit

Minimum objective function

value desired

Scalar | {-Inf}

OutputFcns

Function(s) get(s) iterative

data and can change options
at run time

Function handle or cell array of

function handles | {[]}

PlotFcns

Plot function(s) called during

iterations

Function handle or cell

array of function handles |
@saplotbestf | @saplotbestx |
@saplotf | @saplotstopping |
@saplottemperature | {[]}

HybridInterval

{'end'}

{3000*numberOfVariables}

12-67

saoptimset

Option

Description

Values

PlotInterval

Plot functions are called at

every interval

Positive integer |{1}

ReannealInterval

Reannealing interval

Positive integer | {100}

StallIterLimit

Number of iterations over

which average change in
fitness function value at
current point is less than

Positive integer |
{500*numberOfVariables}

options.TolFun

Function used to update

temperature schedule

Function handle |

TimeLimit

The algorithm stops after

running for TimeLimit
seconds

Positive scalar | {Inf}

TolFun

Termination tolerance on
function value

Positive scalar | {1e-6}

TemperatureFcn

See Also

For more about these options, see Simulated Annealing Options on

page 9-53.
saoptimget, simulannealbnd

12-68

@temperatureboltz |
@temperaturefast |
{@temperatureexp}

simulannealbnd

Purpose

Find unconstrained or bound-constrained minimum of function of

several variables using simulated annealing algorithm

Syntax

x = simulannealbnd(fun,x0)
x = simulannealbnd(fun,x0,lb,ub)
x = simulannealbnd(fun,x0,lb,ub,options)
x = simulannealbnd(problem)
[x,fval] = simulannealbnd(...)
[x,fval,exitflag] = simulannealbnd(...)
[x,fval,exitflag,output] = simulannealbnd(fun,...)

Description

x = simulannealbnd(fun,x0) starts at x0 and finds a local minimum

x to the objective function specified by the function handle fun. The
objective function accepts input x and returns a scalar function value
evaluated at x. x0 may be a scalar or a vector.
x = simulannealbnd(fun,x0,lb,ub) defines a set of lower and upper
bounds on the design variables, x, so that a solution is found in the
range lb x ub. Use empty matrices for lb and ub if no bounds exist.
Set lb(i) to -Inf if x(i) is unbounded below; set ub(i) to Inf if x(i)

is unbounded above.
x = simulannealbnd(fun,x0,lb,ub,options) minimizes with the
default optimization parameters replaced by values in the structure
options, which can be created using the saoptimset function. See the
saoptimset reference page for details.
x = simulannealbnd(problem) finds the minimum for problem, where
problem is a structure containing the following fields:
objective

Objective function

x0

Initial point of the search

lb

Lower bound on x

ub

Upper bound on x

rngstate

Optional field to reset the state of the

random number generator

12-69

simulannealbnd

solver

'simulannealbnd'

options

Options structure created using saoptimset

Create the structure problem by exporting a problem from Optimization

Tool, as described in Importing and Exporting Your Work in the
Optimization Toolbox documentation.
[x,fval] = simulannealbnd(...) returns fval, the value of the

objective function at x.
[x,fval,exitflag] = simulannealbnd(...) returns exitflag,

an integer identifying the reason the algorithm terminated. The

following lists the values of exitflag and the corresponding reasons
the algorithm terminated:
1 Average change in the value of the objective function over
options.StallIterLimit iterations is less than options.TolFun.
5 options.ObjectiveLimit limit reached.
0 Maximum number of function evaluations or iterations exceeded.
-1 Optimization terminated by the output or plot function.
-2 No feasible point found.
-5 Time limit exceeded.
[x,fval,exitflag,output] = simulannealbnd(fun,...) returns
output, a structure that contains information about the problem and
the performance of the algorithm. The output structure contains the

following fields:
problemtype Type of problem: unconstrained or bound
constrained.
iterations The number of iterations computed.
funccount The number of evaluations of the objective function.
message The reason the algorithm terminated.

12-70

simulannealbnd

temperature Temperature when the solver terminated.

totaltime Total time for the solver to run.
rngstate State of the MATLAB random number generator, just
before the algorithm started. You can use the values in rngstate to
reproduce the output of simulannealbnd. See Reproducing Your
Results on page 6-17.

Examples

Minimization of De Jongs fifth function, a two-dimensional function

with many local minima. Enter the command dejong5fcn to generate
the following plot.

x0 = [0 0];
[x,fval] = simulannealbnd(@dejong5fcn,x0)

12-71

simulannealbnd

Optimization terminated: change in best function value

less than options.TolFun.
x =
0.0392

-31.9700

fval =
2.9821

Minimization of De Jongs fifth function subject to lower and upper

bounds:
x0 = [0 0];
lb = [-64 -64];
ub = [64 64];
[x,fval] = simulannealbnd(@dejong5fcn,x0,lb,ub)
Optimization terminated: change in best function value
less than options.TolFun.
x =
-31.9652

-32.0286

fval =
0.9980

The objective can also be an anonymous function:

fun = @(x) 3*sin(x(1))+exp(x(2));
x = simulannealbnd(fun,[1;1],[0 0])
Optimization terminated: change in best function value
less than options.TolFun.
x =
457.1045

12-72

simulannealbnd

0.0000

Minimization of De Jongs fifth function while displaying plots:

x0 = [0 0];
options = saoptimset('PlotFcns',{@saplotbestx,...
@saplotbestf,@saplotx,@saplotf});
simulannealbnd(@dejong5fcn,x0,[],[],options)
Optimization terminated: change in best function value
less than options.TolFun.
ans =
0.0230

-31.9806

The plots displayed are shown below.

12-73

simulannealbnd

See Also

12-74

ga, patternsearch, saoptimget, saoptimset

A
Examples
Use this list to find examples in the documentation.

Examples

GlobalSearch and MultiStart

Example: Creating a Problem Structure with createOptimProblem on
page 3-6
Example: Creating a Problem Structure with the Optimization Tool on
page 3-11
Example: Creating a Nondefault GlobalSearch Object on page 3-15
Example: Creating a Nondefault MultiStart Object on page 3-15
CustomStartPointSet Object for Start Points on page 3-17
Cell Array of Objects for Start Points on page 3-18
Example of Run with GlobalSearch on page 3-19
Example of Run with MultiStart on page 3-20
Example: Changing the Definition of Distinct Solutions on page 3-26
Example: Types of Iterative Display on page 3-28
Example: Visualizing the Basins of Attraction on page 3-32
Example: Searching for a Better Solution on page 3-52
Examples of Updating Problem Options on page 3-57
Changing Global Options on page 3-57
Example: Reproducing a GlobalSearch or MultiStart Result on page 3-59
Example: Finding Global or Multiple Local Minima on page 3-63
Example: Using Only Feasible Start Points on page 3-70
Example: Parallel MultiStart on page 3-74
Example: Isolated Global Minimum on page 3-77

Pattern Search
Example: Finding the Minimum of a Function Using the GPS Algorithm
on page 4-7
Example: A Linearly Constrained Problem on page 4-26
Example: Working with a Custom Plot Function on page 4-30
Example: Using a Complete Poll in a Generalized Pattern Search on
page 4-44
Example: Comparing the Efficiency of Poll Options on page 4-48
Using a Search Method on page 4-55
Constrained Minimization Using patternsearch on page 4-69
Example of Vectorized Objective and Constraints on page 4-76

A-2

Genetic Algorithm

Genetic Algorithm
Example: Rastrigins Function on page 5-8
Example: Creating a Custom Plot Function on page 5-30
Example: Resuming the Genetic Algorithm from the Final Population
on page 5-34
Example: Setting the Initial Range on page 5-50
Example: Linearly Constrained Population and Custom Plot Function
on page 5-53
Example: Global vs. Local Minima with GA on page 5-72
Using a Hybrid Function on page 5-76
Constrained Minimization Using ga on page 5-82
Example: Multiobjective Optimization on page 7-7

Simulated Annealing
Example: Minimizing De Jongs Fifth Function on page 6-7

A-3

A-4

Examples

Index
A
Index

accelerator
mesh 4-64
algorithm
genetic 5-20
pattern search 4-14
simulated annealing 6-12
annealing 6-10
annealing schedule 6-10

C
cache 4-65
children
crossover 5-22
elite 5-22
in genetic algorithms 5-19
mutation 5-22
constraint function
vectorizing 4-72
creation function 9-34
linear feasible 5-53
range example 5-50
crossover 5-63
children 5-22
fraction 5-66

D
direct search 4-2
diversity 5-18

E
elite children 5-22
expansion
mesh 4-59

F
fitness function 5-17

vectorizing 5-81
fitness scaling 5-59
Function files
writing 2-2

G
ga function 12-7
gamultiobj function 12-13
gaoptimget function 12-20
gaoptimset function 12-21

generations 5-18
genetic algorithm
description 5-20
nonlinear constraint algorithm, ALGA 5-27
options 9-29
overview 5-2
setting options at command line 5-40
stopping criteria 5-24
using from command line 5-39
global and local minima 5-72
GPS 4-2

H
hybrid function 5-76

I
individuals
applying the fitness function 5-17
initial population 5-21

M
MADS 4-2
maximizing functions 2-5
mesh 4-12
accelerator 4-64
expansion and contraction 4-59
minima

Index-1

Index

global 5-72
local 5-72
minimizing functions 2-5
multiobjective optimization 7-2
mutation 5-63
options 5-64

N
noninferior solution 7-3
nonlinear constraint
pattern search 4-37
nonlinear constraint algorithms
ALGA 5-27
ALPS 4-24

patternsearch function 12-42

Plot function
custom 5-53
plots
genetic algorithm 5-29
pattern search 4-29
poll 4-13
complete 4-44
method 4-42
population 5-18
initial 5-21
initial range 5-50
options 5-49
size 5-58
psoptimget function 12-48
psoptimset function 12-49

O
objective function 6-10
vectorizing 4-72
Optimization Tool
displaying genetic algorithm plots 5-29
displaying pattern search plots 4-29
genetic algorithm 5-29
pattern search 4-26
options
genetic algorithm 9-29
simulated annealing algorithm 9-53

P
parents in genetic algorithms 5-19
Pareto optimum 7-4
pattern search
description 4-14
nonlinear constraint algorithm, ALPS 4-24
options 9-7
overview 4-2
setting options at command line 4-38
terminology 4-11
using from command line 4-36

Index-2

R
Rastrigins function 5-8
reannealing 6-10
reproduction options 5-63

S
saoptimget function 12-64
saoptimset function 12-65

scaling
fitness 5-59
search method 4-55
selection function 5-62
setting options
genetic algorithm 5-49
pattern search 4-42
simulannealbnd function 12-69
simulated annealing
description 6-12
overview 6-2
simulated annealing algorithm
options 9-53

Index

setting options at command line 6-15

stopping criteria 6-12
using from command line 6-14
stopping criteria
pattern search 4-21

V
vectorizing fitness functions 5-81
vectorizing objective and constraint
functions 4-72

T
temperature 6-10

Index-3