An Overview of
Particle Swarm Optimization
Jagdish Chand Bansal
Mathematics Group
Birla Institute of Technology and Science, Pilani
Email:
[email protected], bits-pilani.ac.in
Overview
¾ Introduction
¾ An Example
¾ Some Developments
¾ Research Issues
2
Optimization Methods
¾ Deterministic and
¾ Probabilistic
3
Deterministic Method
Merits
¾ Give exact solutions
¾ Do not use any stochastic technique
¾ Rely on the thorough search of the feasible domain.
Demerits
¾ Not Robust- can only be applied to restricted class of
problems.
¾ Often too time consuming or sometimes unable to
solve real world problems.
4
Probabilistic Method
Merits
• Applicable to wider set of problems i.e. function need not be
convex, continuous or explicitly defined
• Use the stochastic or probabilistic approach i.e. random
approach
Demerits
¾ Converges to the global optima probabilistically
¾ Some times get stuck at local optima.
5
Some Existing Probabilistic Methods
¾ Simulated Annealing (SA)
¾ Random Search Technique (RST)
¾ Genetic Algorithm (GA)
¾ Memetic Algorithm (MA)
¾ Ant Colony Optimization (ACO)
¾ Differential Evolution (DE)
¾ Particle Swarm Optimization (PSO)
6
Why PSO for Optimization ?
¾ Continuous Optimization Problem
Non-differentiable,
Non-Convex
Highly nonlinear
Many local-optima
¾ Discrete Optimization Problem
NP-Complete problems: Nobody has found so far any
good algorithm for any problem in this class
¾ Search speed
7
Particle Swarm Optimization Inspiration
Artificial Life
The term artificial life is used to describe research into
human made systems that possess some of the
essential properties of life. A-life includes two folded
research:
¾ A-life studies how computational techniques
can help studying biological phenomena
¾ A-life studies how biological techniques can
help out with computational problem
8
Inspiration cont..
Based on bird flocking or fish schooling and
swarming theory of A-Life.
About fish schooling: “In theory at least,
individual members of the school can
profit from the discoveries and previous
experience of all other members of the
school during the search for food “.
(a sociobiologist E. O. Wilson)
This is the basic concept behind PSO.
9
Inventors
Developed in 1995 by
¾ Prof. James Kennedy (Right)
¾ Prof. Russel Eberhart (Left)
10
¾ PSO uses a population of individuals, to search
feasible region of the function space. In this
context, the population is called swarm and the
individuals are called particles.
¾ Though the PSO algorithm has been shown to
perform well, researchers have not been able to
explain fully how it works yet.
11
Update Equations
Each particle tries to modify its current position and
velocity according to the distance between its current
position and pbest, and the distance between its current
position and gbest.
12
Updated Velocity
rand (0,1), to stop the swarm converging too quickly
Current Velocity Acceleration factors, can be used to
change the weighting between
personal and population experience
v = v + c1r1 ( pbest − current) + c2 r2 ( gbest − current)
This is the cognitive component This is the social component where
which draws individuals back to individuals compare themselves to
their previous best situations. others in their group.
Velocity Update Equation
(Rate of Change in Particle’s Position)
current = current + v Position Update Equation
13
PSO Parameters
1. The number of particles :
20 – 40 particles. For most of the problems 10
particles are large enough to get good results.
2. Dimension of particles :
It is determined by the problem to be optimized.
3. Range of particles :
It is also determined by the problem to be optimized,
we can specify different ranges for different
dimension of particles.
14
4. Vmax :
This is done to help keep the swarm under control.
we set the range of the particle as the Vmax.
e.g. X belongs [-10, 10], then Vmax = 20.
One another approach is
Vmax= ⎣(UpBound – LoBound)/5⎦
5. Learning/Acceleration factors :
c1 and c2 usually equal to 2. However, other settings
were also used in different papers. But usually c1
equals to c2 and ranges from [0, 4].
6. The stopping criteria :
The maximum number of iterations the PSO execute
and the minimum error requirement.
Basic Flow of PSO
1. Initialize the swarm from the solution space
2. Evaluate fitness of individual particles
3. Modify gbest, pbest and velocity
4. Move each particle to a new position.
5. Go to step 2, and repeat until convergence or a
stopping condition is satisfied.
16
An Example
Understanding of Step by step
Procedure of PSO
17
Two Versions of PSO
¾ gbest PSO - global version is
faster but might converge to
local optimum for some
problems.
¾ lbest PSO - local version is a
little bit slower but not easy to
be trapped into local optimum.
One can use global version to get quick result and
use local version to refine the search
18
BINARY PSO
¾ This version has attracted much lesser attention as
compared to PSO
¾ Particle position is not a real value, but either 0 or 1
¾ Velocity represents the probability of a bit to take the
value 0 or 1 not the rate of change in particle’s position
as in PSO for continuous optimization
19
BINARY PSO
¾The particle’s position in a dimension is
randomly generated using sigmoid function
1
1 0.8
sigm( x) = 0.6
1 + exp(− x)
0.4
0.2
0
-6 -4 -2 0 2 4 6
20
Velocity and Position Update
vid = vid + c1r1 ( pid − xid ) + c 2 r2 ( p gd − xid )
⎧1 if rand () < sigm (vid )
xid = ⎨
⎩0 otherwise
21
No Free Lunch Theorem
• In a controversial paper in 1997 (available at
AUC library), Wolpert and Macready proved
that “averaged over all possible problems or
cost functions, the performance of all search
algorithms is exactly the same”
• No algorithm is better on average than blind
guessing
22
Important Developments
Almost all modifications vary in some
way the velocity update equation.
23
A Brief Review
PSO-W : With Inertia Weight
PSO-C : With Constriction Factor
FIPSO : Fully Informed PSO
HPSOM : Hybrid PSO with Mutation
MeanPSO : Mean PSO
qPSO : Quadratic approximation PSO
24
Inertia Weight
Shi and Eberhart introduced the inertia weight w in the
algorithm (PSO-W).
Then the iterative expression becomes:
v = w * v + c1r1 ( pbest − current) + c2 r2 ( gbest − current)
current = current + v
w represents the inertia weight,
which enhances the exploration ability of particles
25
Why Inertia Weight
¾ When using PSO, it is possible for the magnitude of
the velocities to become very large.
¾ Performance can suffer if Vmax is inappropriately set.
¾ For controlling the growth of velocities a dynamically
adjusted or constant inertia weight were introduced.
¾ Larger w - greater global search ability
¾ Smaller w - greater local search ability.
26
Constriction Factor
Clerc and Kennedy proposed that the constriction factor is
effective for the algorithm to converge (PSO-C)
v = χ * (v + c1r1 ( pbest − current) + c2 r2 ( gbest − current))
current = current + v
χ=
2
φ = c1 + c 2 > 4
2 − φ − φ (φ − 4 )
27
Fully Informed PSO
A particle is attracted by every other particle in its
neighborhood.
⎡ ⎤
v = χ * ⎢v + ∑ ci ri {p (i ) − current (i )}⎥
⎣ i∈N i ⎦
28
Stagnation
v = w * v + c1r1 ( pbest − current ) + c2 r2 ( gbest − current )
v = χ * v + c1r1 ( pbest − current) + c2 r2 ( gbest − current)
¾PSO algorithm performs well in the early stage, but easily
becomes premature in the local optima area.
¾The velocity is only related with inertia weight and
constriction factor
¾If the current position of a particle is identical with the global
best position and if the current velocity is a small value, the
velocity in next iteration will be smaller. Then the particle will
be trapped in this area which leads to premature convergence.
¾This phenomenon is known as stagnation
29
Hybrid Particle Swarm Optimizer with
Mutation (HPSOM).
HPSOM has the potential to escape from the local
optimum and search in a new position. The mutation
scheme randomly chooses a particle and then move to
a different position in search area. The operation
shows as follows:
mut ( x id ) = x id + ∆ x , rand () < 0 . 5
mut ( x id ) = x id − ∆ x , rand () > 0 . 5
∆x is randomly obtained from
[ 0 , 0 . 1 × (max range ( d ) − min range ( d ))]
This mutation operation is governed by a constant called probability of
mutation
30
MeanPSO
MeanPSO
gbest − pbest gbest + pbest
1.47 1 . 47
current pbest gbest Solution
0
PSO
31
32
qPSO:Quadratic Approximation (QA)
R1 Æ Particle with best fitness value
R2 and R3 Æ Randomly chosen distinct particles
⎛
R = 0.5 * ⎜⎜
* ( R2
2
− R3
2
) f ( R1 ) (
+ R3
2
− R1
2
)f ( R2 ) (
+ R1
2
− R 2
)
2 f (R3 )
⎞
⎟⎟
⎝ ( R2 − R3 ) f ( R1 ) + (R3 − R1 ) f (R2 ) + (R1 − R2 ) f (R3 ) ⎠
Where f(Ri) Æ is the objective function value at Ri , for i=1, 2, and 3.
The calculations are to be done component wise to obtain R* ..
33
The Process of Hybridization
Figure 4.1: Transition from ith iteration to i+1th iteration
s1 s'1
s2 s'2
- -
- -
- PSO - PSO
sp s'p
Particle
Index
qPSO qPSO
sp+1 s'p+1
sp+2 s'p+2
- -
- QA - QA
sm s'm
ith iteration i+1th iteration
The percentage of swarm which is to be updated by QA is called Coefficient
of Hybridization (CH)
34
Start
Random Swarm Flowchart
ITER =0 of qPSO Process
Evaluate Objective Function Value of all
Particles and Determine GBEST
Yes
Stopping Criterion
Satisfied?
No
ITER =ITER + 1
Split Swarm S into subswarms S1 and S2
For S1 For S2
Determine pbest
and gbest (=
GBEST) Is it possible to No
Determine R1, R2 and
R3 such that atleast two
of them are distinct?
Velocity Update
Yes
Determine R1 (= GBEST), R2 and
R3
Position Update
using PSO Position Update using QA
Evaluate Objective Function Value
of all Particles and Determine
GBEST
Report Best Particle
35
End
Research Issues
¾Hybridization
¾Parallel Implementation
¾New Variants
¾modification in Velocity Update Equation
¾Introduce some new operators in PSO
¾Discrete Particle Swarm Optimization
¾Interaction with biological intelligence
¾Convergence Analysis 36
Some Unsolved Issues
• Convergence analysis.
• Dealing with discrete variables.
• Combination of various PSO techniques
to deal with complex problems.
• Interaction with biological intelligence.
• Cryptanalysis.
37
