See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228436815

Global Optimization for Some Engineering Problems: Difficulties and

Prospects

Article · January 2010

CITATIONS READS

0 193

2 authors:

Rimantas Belevičius Dmitrij Šešok

Vilnius Gediminas Technical University Vilnius Gediminas Technical University
46 PUBLICATIONS   298 CITATIONS    45 PUBLICATIONS   167 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Rimantas Belevičius on 01 June 2014.

The user has requested enhancement of the downloaded file.

 GLOBAL OPTIMIZATION FOR SOME ENGINEERING PROBLEMS:
DIFFICULTIES AND PROSPECTS
Rimantas Belevičius1, Dmitrij Šešok 2
1
Vilnius Gediminas Technical University, Saulėtekio ave. 11, LT-10223 Vilnius, Lithuania.
E-mail: [email protected]
2
Institute of Mathematics and Informatics, Akademijos str. 4, LT-08663 Vilnius, Lithuania.
E-mail:[email protected]

Abstract. On the basis of one engineering non-convex optimization problem – optimization of pile placement sche-
mes in grillage type foundations, several global optimization algorithms are compared. Comparison reveals clear ad-
vantages of stochastic algorithms, among which the simulated annealing is most promising. The prospects of using
simulated annealing on different computer platforms are shown; the largest solved problem contained 55 decision
variables.

Keywords: global optimization, pile placement schemes, simulated annealing, clustering, GRID computations.

Introduction The grillage-type foundations are ones of the most

popular and effective schemes of foundations, especially
Optimization is an inherent part of all engineering in case of weak grounds. The optimal scheme of grillage
practice. In the construction of buildings means that, all should possess the minimum possible number of piles.
parts of buildings from foundations to roofs should be Theoretically, reactive forces in all piles should approach
designed and built optimally and thrifty as much as the the limit magnitudes of reactions for those piles. This
conditions of safety and comfort allow. Mathematically, a goal can be achieved by choosing appropriate pile posi-
number of such optimization problems (e.g., topology tions. However, some problem constraints such a mini-
optimization of roof-supporting truss systems, load carry- mum allowable distance between adjacent piles (due to
ing structures of bridges, optimization of pile placement the capabilities of pile driver), or so-called immovable
schemes under grillage-type foundations or under pile piles which are introduced into certain positions of pile
caps, etc.) can be reduced to global minimization of mul- scheme by the analyst and do not participate in the opti-
timodal functions. Such problems are difficult in the mization process, may hinder to achieve the global solu-
sense of algorithmic complexity, and global optimization tion. In all numerical examples of the present paper the
algorithms are computationally very expensive. Aim of immovable piles are not dealt with.
this paper is to review the capabilities and application Whereas the beam optimization problems in form of
possibilities of different global optimization algorithms as optimal sizing of beams in grillage structures under given
well as techniques in engineering practice. Here the boundary and loading conditions (see, e.g. (Chamoret et
optimization of pile placement schemes serves as a really al. 2008) and references therein) or optimal layout of
good benchmark since the problem is non-convex, highly grillages (e.g. (Rozvany 1997)) attracted lot of attention,
sensitive to the design parameters (i.e., positions of piles only a few papers deal with the optimization of pile pla-
under the connecting beams), and the global solution (or cements schemes. In (Kim et al. 2005) an optimal pile
at least the lower bound of solution) is always known in placement scheme under the raft is sought that minimizes
advance. Several deterministic and stochastic global the differential settlements of the raft using genetic algo-
optimization algorithms without/and including some rithms (GA). In (Chan et al. 2009) the pile groups under
heuristic information on the problem are compared. The the solid pile cap together with the cap dimensions are
prospects of an industrial-scale optimization of optimized using hybrid GA. The local search algorithms
foundation schemes on a single PC, clusters of were employed for optimal placing of piles under a sepa-
computers, and distributed computing on a GRID are rate beam of grillage (Belevičius and Valentinavičius
shown. 2001) and under whole grillage using iterative algorithm
860
on the basis of mentioned work (Belevičius et al. 2002). where xi is a design parameter defining the position of i-
Experience shows that the objective function for practical th pile. L is the total length of all beams in the grillage. If
grillage optimization problems possesses many local the minimal possible distance δ between adjacent piles
minima points. Due to this the local search obviously is is specified, there are additional constraints
not a proper choice, and global optimization algorithms
are the necessity. The deterministic global optimization xi − x j ≥ δ, i ≠ j (4)
algorithms proved to require non-realistic computer re-
sources for even small-scale grillage optimization prob- where xi are two-dimensional coordinates of piles and
lems (Čiegis et al. 2006). Promising results for larger-
xi − x j denotes the distance between piles. To cope
scale grillages (up to tens of design variables) were
achieved with stochastic algorithms: GA (Belevičius and with this constraint a penalty is included in the objective
Šešok 2008), simulated annealing (SA) algorithm (Šešok function.
et al. 2010a; Šešok et al. 2010b). A finite element program is used as a “black-box”
The paper is organized as follows. In the next chap- routine to the optimization program for solution of direct
ter the mathematical formulation of the optimization pro- problem to find reactive forces in the grillage. In the di-
blem is presented. Chapter 3 provides comparison of rect problem the connecting beams in the grillage are
different optimization algorithms solving 10 real-life idealized as the beam elements, while the piles are treated
grillages. In the 4th chapter attention is given to the com- as supports, i.e. finite element mesh nodes with given
puting on different computer platforms. Finally, some elastic boundary conditions. Since time of optimization
general conclusions on the prospects of application of crucially depends on the solution time of the direct prob-
global optimization algorithms to the optimization of lem, fast problem-oriented original FORTRAN programs
grillages are drawn. with a special mesh pre-processor have been developed.
The finite element and optimization programs interact in
Problem formulation the following way: optimization program produces a
guess for pile positions in one-dimensional construct; the
The optimization problem is formulated as follows: finite element program transforms it to the two-
dimensional grillage, automatically subdivides the gril-
f * = min f (x ) (1)
x∈D
lage into finite element mesh, solves the statical problem
and obtains the objective function – maximum reactive
where f(x) is a nonlinear objective function of continuous force among all piles. On basis of this value the optimiza-
variables f : ℜ n → ℜ , n is the number of design parame- tion algorithm produces a new guess, and the loop is re-
ters x defining positions of piles, D ⊂ ℜ n is a feasible
peated.
The beam elements have 2 nodes with 6 degrees of
region of design parameters. Besides of the global mini-
freedom each (3 displacements along the coordinate axes
mum f * one or all global minimizers x * : f (x * ) = f * and 3 rotations about these axes). The stiffness matrix for
should be found. No assumptions on unimodality are element can be found in many textbooks, e.g. by (Zien-
included into formulation of the problem – many local kiewicz and Taylor 2005).
minima may exist. The main statics equation is
In this paper the maximal vertical reactive force at a
{u} = {F }
a a a
pile is considered as the objective function: [K ] (5)

f (x ) = max Ri (x ) (2) where a stands for the ensemble of elements (not shown
i =1,..., N a
in eq. below), {u} are the nodal displacements, and {F}
where N a is the number of piles, Ri (x ) is the reactive are the active forces. The reactive forces at piles are avai-
lable after obtaining the nodal displacements:
force at i-th pile.
Since a supporting pile may reside only under con-
necting beams, there are evident restrictions on the posi-
Ri = ∑Kj
[ u
ij ] j (6)

tions of piles: during the optimization process the piles The finite element program requires the following
can move only along the connecting beams. Therefore, a initial data:
two-dimensional beam structure of the grillage is “un- – The geometrical scheme of connecting beams;
folded” to a one-dimensional construct, and the piles are – Cross-section data of all beams (area, moments
allowed to range through this space freely. of inertia);
In such a formulation one design parameter corre- – Material data of all beams (material in one
sponds to a position of one pile in the one-dimensional beam is treated as isotropic);
construct (n=Na). The backward transformation restores – Positions of immovable piles (if any);
the positions of piles in the two-dimensional beam struc- – Maximum allowable reactive force at any pile
ture of the grillage. The constraints for the design pa- Rallw;
rameters are as follows: – Minimum possible distance between adjacent
0 ≤ x i ≤ L, i = 1,..., N a (3) piles;
– Stiffnesses of pile (vertical, rotational);

861
 – Loading data. Active forces can be applied in
the form of concentrated loads and moments at
any point on beam, or in the form of distributed
trapezoidal loadings at any segment of beam.

Comparison of optimization algorithms

The algorithms were compared using the results of

optimization of pile placement schemes of 10 practical
grillages. All these grillages are of small-to-medium
scale, requiring from 17 to 55 piles. Data for these prob-
lems are obtained from several Dutch design bureaus
(courtesy of Consultancy W. F. O. B.V., Paauw B. V.
Aannemingsbedrijf, and others) which use the analysis
and design package MatrixFrame. MatrixFrame has Fig 1. Grillage No 1 (according to the Table 1)
some capabilities for optimization of pile placement
schemes employing local search methods (Belevičius and Table 1. Characteristics of grillages
Valentinavičius 2001). One trait is common to all these Problem
Na L Rallw Rideal
problems: the current optimization routine of Matrix- No
Frame was not capable to yield even a rational scheme of 1 25 172.9 325 307.47
pile placement. 2 18 52.9 110 104.12
The typical grillage is shown in Fig 1. All the char- 3 31 84.1 105 101.85
acteristics of all 10 grillages are summarized in Table 1.
4 31 84.9 105 101.24
There are no immovable piles, therefore the numbers of
piles Na and design variables n coincide. Rallw denotes the 5 30 63.9 100 97.51
allowable reactive force at the piles, it is equal for all 6 37 80.1 100 97.53
piles in all problems. The number of piles is obtained 7 23 129.1 300 287.35
dividing the total sum of loadings by the allowable reac- 8 34 137.9 250 236.28
tive force of piles – the number of piles cannot be less 9 17 97.6 250 244.71
than this number. Rideal denotes the ideal theoretical valu- 10 55 315.61 350 349.05
e; the ideal solution is when all reactive forces at the piles
are equal. In all these problems the proportion between MRS is a random search algorithm, where the deci-
the total loading and the allowable reaction is such that sion variables are generated randomly with uniform dis-
the engineering solution requires achieving almost the tribution, but with one additional constraint: the differ-
ideal solution. This is the main reason, why these prob- ence between two decision variables should be larger
lems are difficult to solve. than some given distance. If the constraint is violated, the
Generally, all global optimization methods can be structure is considered to be non-feasible. Such a heuris-
categorized into non-gradients methods which require tic modification is motivated by the fact that due to the
information from the objective function only, and gradi- usual distribution of loading over the grillage beams, the
ent methods which beside the information about objective piles also should be spread over the whole space of gril-
function use in addition the gradient or higher derivative lage. Reactive forces at the piles are found for N feasible
information. Accurate calculation of function gradients is random structures and one with the smallest maximal
usually computationally expensive and problematic. Also, reactive force at a pile is considered to be the best solu-
the algorithms may be divided into deterministic algo- tion found.
rithms which guarantee the same solution on every run of SA is also a stochastic algorithm which was in-
algorithm, and stochastic ones. The last algorithms pro- vented for simulation of metal annealing process (Kirkpa-
duce different results on every run therefore the decisions trick et al. 1983). Here the current solution is replaced by
about achieved solutions can be made only on the basis of a random solution with a probability that depends on the
several tens of runs. difference of the function value and the so-called tem-
Algorithms involved in comparison cover all this perature parameter. In the beginning the temperature
nomenclature. To make a fair comparison, the total num- parameter is large allowing non-improving changes.
ber of objective function evaluations for each algorithm is Gradually temperature is decreased and the search be-
the same – N = 5000. The following algorithms were comes descent. In our implementation the initial solution
used: is obtained using MRS taking the best feasible solution
– Modified random search (MRS) out of N_init=200. Then SA algorithm is employed.
– Simulated annealing (SA) Stochastic GAs (Goldberg 1989; Šešok et al. 2008)
– Genetic algorithm (GA) simulate evolution (selection, mutation, crossover) in
– Simplex method (SM) which a population of solutions evolves improving values
– Variable metric method (VM) of the objective function. In our implementation, the ini-
– NEWUOA algorithm tial population of popsize individuals is generated using

862
MRS. The new generation is obtained from the previous sover and mutation for GA are tuned to the problems
one using selection, crossover and mutation operations. under consideration (Šešok et al. 2010a, 2010b). All al-
Besides the population size and generation number, re- gorithms are initialized using MRS. This heuristic modi-
sults of algorithm depend on the probabilities of cross- fication significantly improves the results and enables
over and mutation, and the selection strategy. Here all optimal design of real grillages. If a pure random search
values of these parameters are chosen numerically, on the is used for initialization, the results are much worse.
basis of numerous experiments.
The SM is a general method for optimization of Table 2. Parameters for algorithms
nonlinear multidimensional function requiring only func- Algorithm Parameters
tion evaluations, but not derivatives. A simplex is the MRS N = 5000
geometrical figure consisting of n + 1 vertices in n di-
SA N_init = 200, N = 5000, t1 = 5, t2 = 2
mensions. The simplex method takes a series of steps,
popsize = 20, p_cross = 0.9, p_mut = 0.3,
moving the vertex of the simplex where the objective GA
G = N/popsize = 5000/20 = 250
function is worst. We use implementation of the method
SM N_init = 300, N = 5000
form (Press 1992).
VM is based on the Newton’s method to find the sta- VM N_init = 300, N = 5000
tionary point of objective function where its gradient is 0. NEWUOA N_init = 300, N = 5000
Here the BFGS form of method (Press 1992) is used.
Since variable metric method requires gradient informa- The algorithms were tested on a personal computer
tion, the results of sensitivity analysis are used. The start- with Intel(R) Xeon(R) CPU E5420 @ 2.50GHz,
ing point is the best solution obtained using MRS in the 3069 MB RAM, 32-bit Operating System. 28 independ-
given number of iterations N_init. Then the local search ent runs were performed for each algorithm.
is performed. Initialization and local search is repeated Table 3 shows the average results of all 28 runs. In
until the overall number of objective function evaluations Table 4 the best obtained results are rendered. The aver-
reaches N. age times of one run are given in Table 5. In all tables the
The NEWUOA algorithm is iterative algorithm for bold type highlights the best achieved results. The last
nonlinear optimization. A quadratic model is used in a column in Tables 3 and 4 presents discrepancies between
trust region procedure for adjusting the variables (Powell the best obtained results and the corresponding ideal
2006). The quadratic model interpolates the function at value.
2n+1 points, only one interpolation point is altered on Thus, an ideal solution was not obtained for any
each iteration. The starting point is the best solution ob- problem in any of 28 independent runs; in each run the
tained using MRS in the given number of iterations objective function was evaluated 5000 times. The best
N_init. The NEWUOA algorithm is stopped when the objective values found differ from ideal ones from 2.2 %
number of objective function evaluations reaches N. (Problem 2, 18 piles) to 46.8 % (Problem 8, 34 piles),
More thorough description of algorithms is given in while the average results – from 4.8 % to 63.4 % for the
(Šešok et al. 2010a, 2010b). Parameters for all algorithms same problems. As it can be expected, generally discrep-
are summarized in Table 2. Other parameters, for exam- ancy is larger for problems with the larger number of
ple the initial temperature for SA or probabilities of cros- piles.

Table 3. Average of the best values found in 28 runs

Problem Discrepancy, %
Rideal RS MRS SA GA SM VM NEWUOA
No
1 307.47 593.33 470.90 371.55 405.36 486.50 454.01 394.92 20.8
2 104.12 153.07 125.71 109.10 112.49 131.75 127.64 113.35 4.8
3 101.85 258.45 144.46 119.06 124.48 153.82 143.60 119.80 16.9
4 101.24 265.41 141.18 117.10 123.72 147.46 139.82 116.81 15.4
5 97.51 318.16 126.08 106.25 112.23 133.28 128.13 108.68 9.0
6 97.53 460.31 160.18 132.22 144.16 172.38 160.85 132.07 35.4
7 287.35 472.74 379.80 314.11 332.11 402.36 382.45 330.16 9.3
8 236.28 695.60 494.83 413.57 444.88 520.98 491.25 385.99 63.4
9 244.71 402.17 343.91 281.12 294.62 369.06 334.90 292.28 14.9
10 349.05 1321.48 702.53 562.79 636.87 759.33 705.57 559.88 60.4

863
Table 4. The best values found in 28 runs
Problem Discrepancy,
Rideal RS MRS SA GA SM VM NEWUOA
No %
1 307.47 503.93 430.67 339.30 360.73 436.68 429.34 370.79 10.4
2 104.12 136.23 119.16 106.36 106.52 125.83 123.18 107.18 2.2
3 101.85 181.75 138.37 107.25 115.04 138.58 135.54 109.00 5.3
4 101.24 175.33 131.15 106.80 112.10 132.10 133.65 108.16 5.5
5 97.51 170.61 116.88 102.00 104.58 120.42 120.74 101.05 3.6
6 97.53 364.21 148.24 117.26 119.65 151.95 146.01 125.37 20.2
7 287.35 409.73 351.71 298.11 310.24 355.82 358.73 306.31 3.7
8 236.28 538.04 440.42 357.67 363.15 426.74 430.11 346.94 46.8
9 244.71 354.98 318.38 253.00 274.55 339.68 313.40 270.68 3.4
10 349.05 1026.48 670.72 463.34 519.07 698.15 648.84 486.46 32.7

Table 5. Average solution times of one experiment, sec

Problem
Na RS MRS SA GA SM VM NEWUOA
No
1 25 899 943 960 844 821 855 878
2 18 92 96 91 82 90 91 91
3 31 835 876 822 732 811 801 753
4 31 1015 1045 1046 935 910 914 916
5 30 319 319 302 276 313 318 315
6 37 811 823 742 718 674 761 750
7 23 649 695 683 635 622 660 656
8 34 3009 2982 2956 2668 2467 2455 2489
9 17 424 417 410 359 413 419 423
10 55 2198 2136 2125 1934 1826 1811 1850

The comparison of algorithms clearly shows the ad- the SA algorithm and the largest optimization problem –
vantages of SA with tuned parameters both for best ob- grillage No 10 with 55 piles, for which the discrepancy
tained as well as for average results of 28 runs. Compar- between average and ideal solutions reaches ~60 %, were
ing the average results, only the NEWUOA significantly chosen for further experiments with massive computation
outruns the SA for Problem 8 (34 piles). It shows slightly systems.
better results also for Problems 4 (31) and 6 (37). As to
the best results, the NEWUOA obtains the best solutions Comparison of solutions on different computer
for two problems. Nearly in all cases the GA shows the platforms
third best solutions; it never outruns the SA or NEWUOA
algorithms. The results of SM and VM are sometimes According to (Mockus 1967), the convergence rate
even worse than that of MRS what means that it is not of stochastic algorithm is asymptotic and slow, not more
worth to spend time to find local minima for these prob- than log(N), where N is the number of iterations. Theo-
lems, but rather to search wider globally since local retically, in order to decrease the discrepancy between
searches do not improve the value of the objective func- obtained optimization result and the global solution
tion significantly. twice, we have to perform about 100 times more itera-
One numerical experiment takes from approximately tions. Thus, for the 55-pile grillage the number of itera-
1.5 (algorithm GA, Problem 2, 18 piles) to 41 minutes tions of ~107 should guarantee the discrepancy between
(VM, Problem 8, 34 piles). Comparing the timings of all obtained and ideal solutions of around 10 %. Since one
algorithms, the clear winners here are the GA and SM. run of SA with 5000 iterations for 55-pile grillage takes
However, the other algorithms are not far behind, until 2125 seconds on PC (Belevičius et al. 2010) and at least
~18 % in ultimate cases. Thus, the timings can not be several tens of independent runs are needed, such a num-
treated as the decisive factor for comparison of these ber of iterations is inconceivable for PC platform.
algorithms. SA and GA algorithms ideally suit for parallel com-
All these results were obtained evaluating the objec- putations, because one run of an algorithm can be directly
tive function for 5000 times and repeating the experiment allotted for one processor. Any changes are not needed
28 times. It can be expected, that longer computations for an algorithm, if the intrinsic parallelization of the task
may render the relevant solutions. On the basis of this, (e.g. in the solution of a system of linear equations) is not
864
involved; data exchange between processors is not nee-
ded in this case. Two computer platforms were used for
the solution: GRID and PC cluster.

GRID computing. Solution scheme in GRID is

shown in Fig 2. Each processor obtains its task, which is
completely independent from the others. Each process
reads its data files and produces its result files in ASCII
format. The produced files are not large therefore it is not
necessary to store them in storage elements. At the end of
computations result files are automatically processed in
order to sort numerical solutions, considering the com-
puted values of objective function, and to store the best
results. It is quite difficult to handle large amount of tasks
and result files, therefore good software including graphi-
cal user interface is required for automatic job submis-
sion, monitoring and result processing.

Job_1 Job_2 Job_N

Proc_1 Proc_2 Proc_N

Res_1 Res_2 Res_N

Fig 3. Web form for grillage optimization application

Best results
GRID computations have some limitations: all the
numerical tests on the nodes of GRID must be accom-
Fig 2. Solution scheme in GRID plished in a real time (usually, in a few hours). Since
some computers in the grid may be incapable, it is not
Grillage optimization software has been deployed in possible to explore long lifespan of each run, however, a
BalticGrid infrastructure by using Gridcom (Šešok et al. large number of the nodes can be taken. Here we render
2010b). It is a simple web interface for launching com- results (Šešok et al. 2010b) of two contrary strategies of
plex applications on grid. Gridcom splits input data into grillage optimization on the GRIDs that exploit all capa-
intervals; generates and submits as many jobs as needed. bilities of our GRID: first, large number of independent
It can scatter parametric jobs into simple jobs or resubmit runs of SA with a short lifespan of one run, and second,
aborted jobs. Finally, Gridcom collects, merges and visu- less number of independent runs of SA with a maximum
alizes the computed results. All these functions are per- possible lifespan.
formed automatically, including transparent upload of For the first strategy, as in case of solution on a sin-
large files to storage elements. Gridcom launches spe- gle PC, we choose the total number of iterations N =
cially developed Gridcom applications. Once created, an 5000, but 1000 independent runs. For the second strategy,
application can be launched many times with different 100 independent runs were performed with 100,000 itera-
input data. Generally it takes some hours to adapt grid tions each. Other parameters used in the solution are gi-
application for Gridcom. Web form for grillage optimiza- ven in (Šešok et al. 2010a, 2010b).
tion application is shown in Fig 3. Obtained results show, the 35 times larger number of
numerical tests as compared to the calculations on a sin-
gle PC allows improving solution from 463.34 to the
value 449.93, i.e. for only ~4%. The second strategy gave
865
the best solution of 433.42, or improving of an error by
~8 %.

Cluster computing. We used the PC cluster VILKAS

(Rocks Cluster Distribution v 5.0, CentOS release 5,
x86_64) at Vilnius Gediminas Technical University. The
cluster consisted of 18 PCs connected by Gigabit Ether-
net (D-Link DGS 1224T Gigabit Smart Switch, 24-Ports
10/100/1000Mbps Base-T Module). Hardware character-
istics of the PC are listed below: Intel® Core2Quad
Q6600 2.40GHz CPU (2x4MB L2 cache and bus fre-
quency equal 1067 MHz), 2x2GB DDR2 800 RAM,
300GB HDD (SATA II Extensions and 16 MB cache),
Fig 4. Best decision (55 piles, 106 iterations)
Gigabit Ethernet NIC. 340 Gflop/s performance was
measured running HPL benchmark. 40 VILKAS’ proces- Conclusions
sors were employed in computations; each of them solved
the task once. The total CPU time was about 50 hours. From several global optimization algorithms belong-
The average results of 40 independent samples for each ing to the stochastic algorithms (genetic algorithms,
of the cluster computers are in Table 6 (Šešok et al. simulated annealing, modified random search, and also
2010a). the simplex since it is launched from a random initial
start) and to the gradient ones (BFGS, NEWUOA), the
Table 6. Optimization of 55-pile grillage: cluster, 1,000,000 simulated annealing gives the most promising results in
iterations
optimization of grillages.
Sample Value Sample Value Computer hardware that is common to a typical civil
1 426.88 21 467.01 engineering design bureaus and a reasonable computation
2 388.64 22 423.25 time for engineering practice do not allow solution of the
3 421.98 23 463.05 problem of pile placement to the very end with any global
4 455.17 24 468.59 optimization algorithm applied. However, first, the fact
that ideal solution usually is not required for the engineer-
5 429.52 25 480.05
ing purposes, and second, the increasing accessibility and
6 464.72 26 418.57
popularity of the distributed computing in contemporary
7 408.67 27 423.29 engineering practice makes it possible solving global
8 425.45 28 462.71 optimization problems possessing to 30–40 design pa-
9 440.74 29 420.65 rameters.
10 461.86 30 441.54 The larger optimization problem with 55 decision
11 417.97 31 455.12 variables was solved only to a discrepancy between ob-
12 488.54 32 417.27 tained result and ideal one of 9% using a cluster of PC
13 458.99 33 403.18
and 50 hours of computations.
14 456.22 34 416.97 Acknowledgements
15 458.31 35 457.97
16 457.55 36 385.15 The computational resources provided by the Bal-
17 462.85 37 430.81 ticGrid-II infrastructure (FP7-INFRA-2007-1.2.3:
18 417.69 38 460.69 e-Science Grid infrastructures contract No 223807) are
19 397.27 39 460.87 gratefully acknowledged. This work is supported by pro-
ject GridTechno (Lithuanian State Science and Studies
20 379.35 40 461.78
Foundation, B-07043).
This work is also supported by The Lithuanian State
The best result (379.35) differs from global mini- Science and Studies Foundation within the project on
mum (349.05) by 9 %. Thus, this error is better than the S-20/2009 “Use of stochastic global optimization meth-
theoretical estimation of error for the number of iterations ods in engineering”.
performed. From the engineering point of view, this is an
acceptable error. Fig 4 shows the graph of this decision. References
The average results of 40 samples are 438.37.
Belevičius, R.; Valentinavičius, S. 2001. Optimisation of gril-
lage-type foundations, in Proc. of the 2nd European EC-
COMAS and IACM Conference “Solids, Structures and
Coupled Problems Problems in Engineering”. Cracow,
Poland 26–29 June, 2001. CD-ROM.

866
 Belevičius, R.; Valentinavičius, S.; Michnevič, E. 2002. Multi- teenth International Offshore and Polar Engineering
level optimization of grillages, Journal of Civil Engineer- Conference, Seoul, Korea, June 19-24, 2005. 730–735.
ing and Management 8(2): 98–103. Kirkpatrick, S.; Gelatt, C. D. J.; Vecchi, M. P. 1983. Optimiza-
Belevičius, R.; Šešok, D. 2008. E. Global optimization of gril- tion by simulated annealing, Science 220(4598): 671–680.
lages using genetic algorithms, Mechanika 74(6): 38–44. doi:10.1126/science.220.4598.671
Belevičius, R.; Ivanikovas, S.; Šešok, D.; Valentinavičius, S.; Mockus, J. 1967. Multimodal problems in engineering design.
Žilinskas, J. 2010. Optimal placement of piles in real gril- Moscow: Nauka.
lages: experimental comparison of optimization algo- Powell, M. J. D. 2006, The NEWUOA software for uncon-
rithms, Engineering Optimization (Under Review). strained optimization without derivatives. Di Pillo, G.;
Chamoret, D.; Qiu, K.; Labed, N.; Domaszewski, M. 2008. Roma, M. (Eds.) Large-Scale Nonlinear Optimization.
Optimization of Truss and Grillage Structures by a Non- Vol. 83 of Nonconvex Optimization and its Applications,
Deterministic Method. In B.H.V. Topping, M. Papadra- Springer, 255–296.
kakis (Editors), in “Proceedings of the Ninth Interna- Press, W. H. 1992. Numerical Recipes in Fortran 77: The Art of
tional Conference on Computational Structures Technol- Scientific Computing. Second Edition. Cambridge Univer-
ogy”, Civil-Comp Press, Stirlingshire, UK, Paper 51, sity Press.
2008. doi:10.4203/ccp.88.51
Rozvany, G. I. N. 1997. Exact optimal layout of grillages for
Chan, C. M.; Zhang, L. M.; Jenny, T. M. 2009. Optimization of partially upward and partially downward loading, Struc-
pile groups using hybrid genetic algorithms, Journal of tural and Multidisciplinary Optimization 13(4): 267–270.
Geotechnical and Geoenvironmental Engineering 135(4):
497–505. Šešok, D.; Belevičius, R 2008. Global optimization of truss
systems using a modified genetic algorithm. Journal of
doi:10.1061/(ASCE)1090-0241(2009)135:4(497) Civil Engineering and Management, 14(3): 147–154.
Čiegis, R.; Baravykaitė, M.; Belevičius, R. 2006. Parallel global doi:10.3846/1392-3730.2008.14.10
optimization of foundation schemes in civil engineering. Šešok, D.; Mockus, J.; Belevičius, R.; Kačeniauskas, A. 2010a.
Applied Parallel Computing: 7th International Confer- Global optimization of grillages using simulated anneal-
ence, PARA 2004, Lyngby, Denmark 20-23 June, 2004: ing and high performance computing, Journal of Civil
revised selected papers. Lecture Notes in Computer Sci- Engineering and Management 16(1): 95–101.
ence, 3732: 305–313. Springer. doi:10.3846/jcem.2010.09
Goldberg, D. E. 1989. Genetic Algorithms in Search, Optimiza- Šešok, D.; Mockus, J.; Belevičius, R.; Kačeniauskas, A. 2010b.
tion and Machine Learning. New York: Addison-Wesley. Application of GRID computing for optimization of gril-
412 p. lages, Mechanika (accepted).
Kim, Y.; Gotoh, K.; Kim, K. S. Toyosada, M. 2005. Optimum Zienkiewicz, O. C., Taylor, R. L. 2005. The Finite Element
Grillage Structure Design under a Worst Point Load using Method for Solid and Structural Mechanics. London: El-
Real-coded Micro-Genetic Algorithm, in Proc. of the Fif- sevier Butterworth-Heinemann, 1–632 p.

867

View publication stats