An adaptive elitist differential evolution for optimization of truss structures with discrete design variables V. Ho-Huu, T. Nguyen-Thoi ⇑ , T. Vo-Duy, T. Nguyen-Trang Division of Computational Mathematics and Engineering, Institute for Computational Science, Ton Duc Thang University, Vietnam Faculty of Civil Engineering, Ton Duc Thang University, Vietnam article info Article history: Received 12 October 2015 Accepted 29 November 2015 Keywords: Differential evolution (DE) Adaptive elitist differential evolution (aeDE) Optimization of truss structures Optimization with discrete design variables abstract This paper proposes an adaptive elitist differential evolution (aeDE) for optimization of truss structures with discrete design variables. The aeDE algorithm is a newly improved version of the differential evolu- tion (DE) algorithm with three modifications. Firstly, in the mutation phase, an adaptive technique based on the deviation of objective function between the best individual and the whole population in the pre- vious generation is proposed to select a suitable mutation operator. This technique helps preserve the balance between global and local searching abilities in the DE. Secondly, in the selection phase, an elitist selection technique which helps choose the best individuals for the next generation is utilized to increase the convergence rate. Finally, a rounding technique is integrated into the aeDE for solving optimization problems with discrete design variables. The efficiency and reliability of the proposed method are demonstrated through six optimization problems of truss structures with discrete design variables. Numerical results reveal that in most of the test cases, the aeDE is more efficient than the DE and some other methods in the literature in terms of the quality of solution and convergence rate. Ó 2015 Elsevier Ltd. All rights reserved. 1. Introduction Structural optimization is one of the interesting subjects for many researchers. In this research direction, besides proposing the effective optimization algorithms which help reduce the cost and improve the performance of structures, the researchers have also tried to find the effective solutions for handling highly nonlin- ear constraints or discrete design variables. During the past dec- ades, many optimization techniques have been developed and successfully applied to a wide range of structural optimization problems. Some of them can be mentioned here such as sequential linear programming (SLP) [1–3], sequential quadratic program- ming (SQP) [4], optimality criterion (OC) [5,6], and forced method [7,8]. Most of them are applied to the continuous design variables. However, in many practical engineering problems, the design variables such as the cross-section areas or thickness are usually discrete values which can be selected from a set of available discrete values. For the optimization problems with discrete design variables, the round-off techniques based on continuous solutions were initially proposed. Nonetheless, the rounded-off solutions may be far from the optimum solution, or they may even be infea- sible, when the number of variables increases [9]. To overcome the computational shortcomings of conventional optimization methods, many direct searching methods based on the model of biological evolution and social interaction such as genetic algorithm (GA), differential evolution (DE), and particle swarm optimization (PSO) have been proposed. They were quite effective and robust for solving truss optimization problems with discrete design variables. These can be illustrated from the publica- tions by Rajeev and Krishnamoorthy [10], Wu and Chow [11] using GA and a modified GA, Lee et al. [9] using harmony search (HS), Li et al. [12] using heuristic particle swarm optimization (HPSO), Sadollah et al. [13] using mine blast algorithm (MBA), Sonmez [14] using artificial bee colony algorithm (ABC), Dede [15] using t eaching–learning-based-optimization (TBLO), and Kaveh and Mah- davi [16] using colliding bodies optimization (CBO). Although these direct search methods have occasionally over- come several restrictions of traditional optimization methods, structural design engineers are still concerned much in finding more simple, effective, and robust methods for structural opti- mization problems with discrete design variables. Among direct search optimization algorithms, the differential evolution (DE) http://dx.doi.org/10.1016/j.compstruc.2015.11.014 0045-7949/Ó 2015 Elsevier Ltd. All rights reserved. ⇑ Corresponding author at: Division of Computational Mathematics and Engi- neering, Institute for Computational Science, Ton Duc Thang University, Vietnam. Tel.: +84 933 666 226. E-mail addresses: [email protected], [email protected](V. Ho-Huu), [email protected], [email protected](T. Nguyen-Thoi), [email protected](T. Vo-Duy), [email protected]. vn (T. Nguyen-Trang). Computers and Structures 165 (2016) 59–75 Contents lists available at ScienceDirect Computers and Structures journal homepage: www.elsevier.com/locate/compstruc
17
Embed
An adaptive elitist differential evolution for optimization of truss structures with discrete design variables
This paper proposes an adaptive elitist differential evolution (aeDE) for optimization of truss structures with discrete design variables. The aeDE algorithm is a newly improved version of the differential evolution (DE) algorithm with three modifications. Firstly, in the mutation phase, an adaptive technique based on the deviation of objective function between the best individual and the whole population in the previous generation is proposed to select a suitable mutation operator . This technique helps preserve the balance between global and local searching abilities in the DE. Secondly, in the selection phase, an elitist selection technique which helps choose the best individuals for the next generation is utilized to increase the convergence rate. Finally, a rounding technique is integrated into the aeDE for solving optimization problems with discrete design variables. The efficiency and reliability of the proposed method are demonstrated through six optimization problems of truss structures with discrete design variables. Numerical results reveal that in most of the test cases, the aeDE is more efficient than the DE and some other methods in the literature in terms of the quality of solution and convergence rate.
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
An adaptive elitist differential evolution for optimization of trussstructures with discrete design variables
http://dx.doi.org/10.1016/j.compstruc.2015.11.0140045-7949/� 2015 Elsevier Ltd. All rights reserved.
⇑ Corresponding author at: Division of Computational Mathematics and Engi-neering, Institute for Computational Science, Ton Duc Thang University, Vietnam.Tel.: +84 933 666 226.
V. Ho-Huu, T. Nguyen-Thoi ⇑, T. Vo-Duy, T. Nguyen-TrangDivision of Computational Mathematics and Engineering, Institute for Computational Science, Ton Duc Thang University, VietnamFaculty of Civil Engineering, Ton Duc Thang University, Vietnam
a r t i c l e i n f o a b s t r a c t
Article history:Received 12 October 2015Accepted 29 November 2015
Keywords:Differential evolution (DE)Adaptive elitist differential evolution (aeDE)Optimization of truss structuresOptimization with discrete design variables
This paper proposes an adaptive elitist differential evolution (aeDE) for optimization of truss structureswith discrete design variables. The aeDE algorithm is a newly improved version of the differential evolu-tion (DE) algorithm with three modifications. Firstly, in the mutation phase, an adaptive technique basedon the deviation of objective function between the best individual and the whole population in the pre-vious generation is proposed to select a suitable mutation operator. This technique helps preserve thebalance between global and local searching abilities in the DE. Secondly, in the selection phase, an elitistselection technique which helps choose the best individuals for the next generation is utilized to increasethe convergence rate. Finally, a rounding technique is integrated into the aeDE for solving optimizationproblems with discrete design variables. The efficiency and reliability of the proposed method aredemonstrated through six optimization problems of truss structures with discrete design variables.Numerical results reveal that in most of the test cases, the aeDE is more efficient than the DE and someother methods in the literature in terms of the quality of solution and convergence rate.
� 2015 Elsevier Ltd. All rights reserved.
1. Introduction
Structural optimization is one of the interesting subjects formany researchers. In this research direction, besides proposingthe effective optimization algorithms which help reduce the costand improve the performance of structures, the researchers havealso tried to find the effective solutions for handling highly nonlin-ear constraints or discrete design variables. During the past dec-ades, many optimization techniques have been developed andsuccessfully applied to a wide range of structural optimizationproblems. Some of them can be mentioned here such as sequentiallinear programming (SLP) [1–3], sequential quadratic program-ming (SQP) [4], optimality criterion (OC) [5,6], and forced method[7,8]. Most of them are applied to the continuous design variables.However, in many practical engineering problems, the designvariables such as the cross-section areas or thickness are usuallydiscrete values which can be selected from a set of available
discrete values. For the optimization problems with discrete designvariables, the round-off techniques based on continuous solutionswere initially proposed. Nonetheless, the rounded-off solutionsmay be far from the optimum solution, or they may even be infea-sible, when the number of variables increases [9].
To overcome the computational shortcomings of conventionaloptimization methods, many direct searching methods based onthe model of biological evolution and social interaction such asgenetic algorithm (GA), differential evolution (DE), and particleswarm optimization (PSO) have been proposed. They were quiteeffective and robust for solving truss optimization problems withdiscrete design variables. These can be illustrated from the publica-tions by Rajeev and Krishnamoorthy [10], Wu and Chow [11] usingGA and a modified GA, Lee et al. [9] using harmony search (HS), Liet al. [12] using heuristic particle swarm optimization (HPSO),Sadollah et al. [13] using mine blast algorithm (MBA), Sonmez[14] using artificial bee colony algorithm (ABC), Dede [15] using teaching–learning-based-optimization (TBLO), and Kaveh and Mah-davi [16] using colliding bodies optimization (CBO).
Although these direct search methods have occasionally over-come several restrictions of traditional optimization methods,structural design engineers are still concerned much in findingmore simple, effective, and robust methods for structural opti-mization problems with discrete design variables. Among directsearch optimization algorithms, the differential evolution (DE)
60 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
firstly introduced by Storn and Price 1997 [17] is one of the mostpopular algorithms. The DE is a population-based method anddemonstrated excellent performance in solving many differentengineering problems with continuous design variables [18–22].Nevertheless, similar to many methods in the evolution algorithmfamily, the DE is still costly in computational source to seek theglobal optimal solution. Especially, in many real-world problemswhere the evaluation of a candidate solution is a computationallyexpensive operation and consequently finding the global optimumor a good suboptimal solution with the original differential evolu-tion algorithm is too time-consuming, or even impossible withinthe time available [23]. Hence, the improvement of the DE withthe aim of accelerating the convergence rate, enhancing the qualityof solutions and integrating a technique for handling discretevariables is an attractive subject in the optimization field.
In the present study, the DE is employed for the first time tosolve truss optimization with discrete design variables. In orderto enhance the DE, three modifications are implemented to givea so-called the adaptive elitist differential evolution (aeDE). Firstly,an adaptive technique based on the deviation of objective functionbetween the best individual and the whole population in the pre-vious generation is proposed to choose a suitable mutation opera-tor. This technique helps preserve the balance between global andlocal searching abilities in the DE. Secondly, an elitist selectiontechnique which helps choose the best individuals for the nextgeneration is utilized to speed up the convergence. Thirdly, a tech-nique for handling discrete variables is integrated into the aeDE inwhich individuals in the population are rounded to the nearestvalue in set of discrete values. The detail of these improvementswill be presented in following sections of the paper. The robustnessand performance of the proposed aeDE are verified through sixnumerical examples. Numerical results indicate that in most ofthe cases, the aeDE is more efficient than the DE and many othermethods in the literature in terms of both the quality of solutionand convergence rate.
The remainder of the paper is organized as follows. Section 2presents mathematical model of structural optimization withdiscrete design variables. Section 3 provides a brief of basic DEalgorithm. The detail of the proposed aeDE is described in Section 4.Some numerical examples are considered in Section 5. Finally,some conclusions are drawn in Section 6.
2. Structural optimization with discrete design variables
A structural optimization problem with discrete design vari-ables is known as a nonlinear programming problem with multinonlinear constraints related to structural behavior. For the sizingoptimization of truss structures, the cross-section areas of the trussbars are often discrete design variables. All of them are selectedfrom a list of discrete cross-sections based on production standard.The optimization problem aims to minimize the weight of thestructure and to satisfy constraints about structural behavior andlimitations of design variables. The truss optimization problemwith discrete design variables can be expressed as follows:
Minimize weightðAÞ ¼Xe
i¼1
qiliAi; i ¼ 1;2; . . . ; e
subject to dmin 6 di 6 dmax; i ¼ 1;2; . . . ; nrmin 6 ri 6 rmax; i ¼ 1;2; . . . ; e
rbi 6 ri 6 0; i ¼ 1;2; . . . ; nc
A 2 S ¼ fA1;A2; . . . ;Adg
ð1Þ
where A is the design variable vectors containing the bar cross-sectional areas Ai; weightðAÞ is the weight of the whole truss struc-ture; qi and li are the material density and the length of the ith
member, respectively; e is the total number of bars in the truss; nis the number of nodes; nc is the number of elements subjectingto compression; di and ri are the nodal displacement and the stress,respectively; rb
i is the allowable buckling stress in the ith bar whenit is in compression and S is set of discrete value of areas.
To handle constraints for the problem (1) in the proposed algo-rithm, we use the penalty function method defined as follows [24]:
f penaltyðAÞ ¼ ð1þ e1 � vÞe2 �weightðAÞ; v ¼Xq
i¼1
maxf0; giðAÞg
ð2Þwhere v denotes the sum of the design violated constraints; q is thenumber of constraints in the problem; giðAÞ is the ith constraint ofthe optimization problem; e1 and e2 are constants presented theexploration and the exploitation rate of the search space. In thispaper, the value of the e1 is set 1, and e2 starts from 20 and thenlinearly increases to 40.
3. The differential evolution (DE) algorithm
The differential evolution (DE) algorithm firstly proposed byStorn and Price (1997) [17] was proven to be one of the mostpromising global search methods and widely used to solveoptimization problems in many fields such as communication[25], pattern recognition [26], and mechanical engineering[19–22]. The DE includes four main phases as follows.
3.1. Initialization
Initially, an initial population, includes NP individuals, isgenerated by means of randomly sampling from the search space.Each individual is a vector containing D design variablesxi ¼ ðx1; x2; . . . ; xDÞ and is created by
ð3Þwhere xlj and xuj are the lower and upper bound of xj, respectively;rand½0;1� is a uniformly distributed random number in [0,1]; NPis the population size; and D is the number of design variables.
3.2. Mutation
Secondly, each individual called the target vector xi in thepopulation is used to generate a mutant vector vi via mutationoperators. Some popular mutation operators are usually used inthe DE as follows
� rand=1: vi ¼ xr1 þ F � ðxr2 � xr3 Þ ð4Þ� rand=2: vi ¼ xr1 þ F � ðxr2 � xr3 Þ þ F � ðxr4 � xr5 Þ ð5Þ� best=1: vi ¼ xbest þ F � ðxr1 � xr2 Þ ð6Þ� best=2: vi ¼ xbest þ F � ðxr1 � xr2 Þ þ F � ðxr3 � xr4 Þ ð7Þ� current� to� best=1: vi ¼ xi þ F � ðxbest � xiÞ þ F � ðxr1 � xr2 Þwhere integers r1; r2; r3; r4; r5 are randomly selected fromf1;2; . . . ;NPg such that r1 – r2 – r3 – r4 – r5 – i; the scale factor Fis randomly chosen within [0,1]; and xbest is the best individual inthe current population.
After this phase, the jth components v ij of mutant vector vi arereflected back to allowable region if the boundary constraints areviolated. This procedure is conducted as follows:
v ij ¼2xlj � v ij if v ij < xlj2xuj � v ij if v ij > xujv ij otherwise
8><>: ð8Þ
Fig. 1. A 10-bar planar truss structure.
0.4 0.6 0.8 1.0 [0.4,1.0]
5500
5600
5700
5800
5900
Mea
n w
eigh
t (lb
)
F
0.4 0.6 0.8 1.0 [0.4,1.0]1000
2000
3000
4000
5000
Num
ber o
f ana
lyse
s
F
Fig. 2. Influence of the mutant factor F on the mean optimal solution.
Table 2Influence of the crossover control parameter CR on the optimal solution.
Crossover control parameterCR
0.7 0.8 0.9 [0.7,1.0]
Best weight (lb) 5490.738 5490.738 5490.738 5490.738
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 61
3.3. Crossover
Thirdly, some elements of the target vector xi are replaced bysome elements of the mutant vector vi to create a trial vector usingbinomial crossover operation
uij ¼v ij if rand½0;1� 6 CR or j ¼ jrandxij otherwise
�ð9Þ
where i 2 f1;2; . . . ;NPg; j 2 f1;2; . . . ;Dg; jrand is an integer selectedfrom 1 to D; CR is the crossover control parameter.
Finally, based on the value of objective function, the trial vectorui is compared to the target vector xi. The better one having lowerobjective function value will survive to the next generation
xi ¼ui if f ðuiÞ 6 f ðxiÞxi otherwise
�ð10Þ
4. The adaptive elitist differential evolution algorithm (aeDE)
In the DE, the parameters such as mutant factor F and crossovercontrol parameter CR and trial vector generation strategies havesignificant influence on its performance [27]. Commonly, toacquire the most satisfactory optimization performance for a par-ticular problem, they are often defined using a trial-and-errorsearch. Obviously, it may lead to a huge amount of the computa-tional cost. In addition, a rigorous selection mechanism in theselection phase may make the DE computationally expensive.Therefore, to overcome these limitations, the paper introducestwomodifications, the first one for the mutation phase and the sec-ond one for the selection phase in order to enhance the searchcapability as well as the convergence speed of the DE algorithm.Their details are presented in the next subsections.
Table 1Influence of the mutant factor F on the optimal solution.
As pointed out by Das et al. [28], the balance between globalexploration and local exploitation significantly influence on thesuccess of the population-based search methods. In the DE, themutation scheme plays a pivotal role in its searching ability andconvergence rate. There are at least five mutation operators havebeen proposed for the DE with different purposes. For example,for the mutation operator ‘‘rand/1”, the DE is good at global search,but bad at local search and hence slowly converges to the globaloptimal solution [29]. In contrast, for the mutation operator ‘‘current-to-best/1”, the DE is good at local search, but bad at globalsearch and easily trapped at the local optimal solutions [30]. Thus,to equalize the global and local search capabilities of the DE,Mohamed & Sabry [31] introduced a new mutation operator whichbased on the weighted difference vector between the best and theworst individuals at a particular generation. This operator is thencombinedwith the ‘‘rand/1” to generate trial vectors with the equalprobability of 0.5. Compared with other versions of the DE, the
Fig. 3. Influence of the crossover control parameter CR on the mean optimalsolution.
0.1 0.01 0.001 0.0001 0.000015500
5550
5600
5650
Mea
n w
eigh
t (lb
)
Threshold
0.1 0.01 0.001 0.0001 0.000011000
1500
2000
2500
3000
Num
ber o
f ana
lyse
s
Threshold
Fig. 4. Influence of the threshold on the mean optimal solution.
62 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
modified method gives good solutions; however, it is only limitedon benchmark test functions. Wang et al. [27] investigated the DEwith composite trial vector generation strategies and controlparameters for unconstraint optimization problems of benchmarktest functions. In particular, for each target vector, three trial vec-tors are created by randomly combining between three mutationoperators ‘‘rand/1”, ‘‘rand/2” and ‘‘current-to-rand/1” with threecontrol parameter settings [F = 1.0, CR = 0.1], [F = 1.0, CR = 0.9],and [F = 0.8, CR = 0.2]. Although the improved method outperformsother variants of the DE, its computational cost is relatively highdue to increasing of the population in each generation. Followingthe trend, in this work, a new adaptive mutation scheme for themutation phase of the DE is proposed. In this scheme, we use twomutation operators, the first one is ‘‘rand/1” which aims to ensurediversity of the population and prohibits the population from get-ting stuck in a local optimum, and the second one is ‘‘current-to-best/1” which aims to accelerate convergence speed of the popula-tion by mean of guiding the population toward the best individual.These two mutation operators are adaptively chosen based on theabsolute value of deviation of objective function between the bestindividual and the entire population in the previous generation(denoted as delta). More specifically, the value of delta is defined by
delta ¼ jfmean=f best � 1j ð11Þwhere f best is the objective function value of the best individual andf mean is the mean objective function value of the whole population.The new mutation scheme is described as follows
if ðdelta > thresholdÞ ð12Þvi ¼ xr1 þ F � ðxr2 � xr3 Þ %rand=1 ð13Þ
elsevi ¼ xi þ F � ðxbest � xiÞ þ F � ðxr2 � xr3 Þ
%current � to� best=1 ð14Þend
Table 3Influence of the threshold on the optimal solution.
where F is the mutant factor which is randomly created in the inter-val [0.4,1]; threshold is a criterion value which is chosen based onthe stopping criterion of the algorithm. This definition will be dis-cussed in more detail in Section 4.3.
In the proposed adaptive mechanism, for each individual, onlyone of two mutation operators is utilized for producing the currenttrial vector. If the value of delta is bigger than the threshold, themutation operator ‘‘rand/1” is used. Otherwise, the mutation oper-ator ‘‘current-to-best/1” is employed. It should be noted that thevalue of delta can self-adjust during the searching process for vari-ous optimization problems, which leads to the change in the imple-mentation of the mutation operators. For example, for unimodalproblems, the value of delta usually reduces gradually during thesearching process. As a result, in the first generations, the value ofdelta is bigger than threshold. This leads to the mutation operator‘‘rand/1” is applied. However, after many generations, the diversityof the population is gradually stable, and then once the value ofdelta is smaller than threshold, the mutation operator ‘‘current-to-best/1” will be utilized. More generally, for multimodal problems,the value of delta is jumpy and becomes unpredictable during thesearching process. Consequently, the implementation of the muta-tion operators is also unpredictable. In addition, the variety in glo-bal and local search abilities is also enriched by randomlygenerating the mutant factor F within the range [0.4, 1.0] which isrecommended as a good choice from many previous studies [27].Summarily, with the proposed adaptive mutation mechanism, theglobal and local search capabilities as well as the convergence rateof the DE may be enhanced significantly.
4.2. Modification of the selection phase
The selection mechanism in Section 3.4 shows that each trialvector ui created after crossover phase will be compared with thetarget vector xi to choose a better individual for the next genera-tion. This may lead to a disadvantage that some good information
Average of number of analyses 1820 2550 3406 4014 5100
15 20 25 30 355450
5500
5550
5600
Mea
n w
eigh
t (lb
)
NP
15 20 25 30 350
2000
4000
6000
Num
ber
of a
naly
ses
NP
Fig. 5. Influence of the population size on the mean optimal solution.
0 2000 4000 6000
5500
6500
7500
8500
Wei
ght (
lb)
Number of analyses
DE: Mean valueDE: Best valueaeDE: Mean valueaeDE: Best value
Fig. 6. Comparison of convergence of the DE and aeDE for the 10-bar planar trussstructure.
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 63
of unselected individuals can be neglected. Although an individualis not good compared to its target individual in the pair, it can bestill better than other individuals in the entire population. Thus,to save good information for the next generation, we use the elitistselection technique introduced by Padhye et al. [32] for the selec-tion progress instead of the basic selection in the DE. This newmechanism is performed as follows: firstly, the children populationC consisting of trial vectors is combined with parent population Pof target vectors to create a combined population Q. Then, NP bestindividuals are chosen from the Q to construct the population forthe next generation. In this way, the best individual of the whole
Table 5Comparison of optimized designs for the 10-bar planar truss structure.
Design variable (area in.2) Rajeev et al. [10] Li et al. [1
population are always stored for the next generation. This helpsthe algorithm obtain better convergence rate. The elitist operatoris depicted as in Algorithm 1.
Algorithm 1: Elitist selection operator
1: Input: Children population C and parent population P2: Assign Q = C [ P3: Select NP best individuals from Q and assign to P4: Output: P
64 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
4.3. The proposed aeDE algorithm
By integrating two above-mentioned improvements into theDE, a so-called adaptive elitist differential evolution (aeDE) algo-rithm is proposed. This method is summarily shown as in Algo-rithm 2 below:
Algorithm 2: The proposed algorithm (aeDE)
1: Generate the initial population2: Evaluate the fitness for each individual in the population3: while delta > tolerance or MaxIter is not reached do4: for i = 1 to NP do5: F = rand[0.4,1]6: CR = rand[0.7,1]7: jrand = randint(1,D)8: for j = 1 to D do9: if rand[0,1] < CR or j == jrand then10: if delta > threshold then11: Select randomly r1 – r2 – r3 – i;
8i 2 f1; . . . ;NPg15: ui;j ¼ xi;j þ F � ðxbest;j � xi;jÞ þ F � ðxr2 ;j � xr3 ;jÞ16: end if17: else18: ui;j ¼ xi;j19: end if20: end for21: Evaluate the trial vector ui
22: end for23: Do selection phase based on Algorithm 124: Define f best; f mean
25: delta ¼ jfmean=f best � 1j26: end while
where tolerance is the allowed error; MaxIter is the maximum num-ber of iterations; and randint(1,D) is the function which returns auniformly distributed random integer number between 1 and D.
According to the Algorithm 2, the aeDE will finish the searchingprogress either when the absolute value of deviation of the objec-tive function of the best individual and the whole population(delta) is less than or equal to the previously assigned value of
(a)
0 2 4 6 8 10 12
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
Dis
plac
men
t (in
.)
Number of displacement
Allowable limitCase 1Case 2
Fig. 7. Constraint boundaries evaluated a
the tolerance or when the maximum number of iterations (MaxIter)is achieved. It can be seen that the threshold is chosen based on thetolerance and obviously, it must be bigger than the tolerance. More-over, it is important to know that choosing the value of thresholdwill directly impact on the global and local search capabilities ofthe aeDE. If the value of threshold is too big compared to the valueof tolerance, the aeDE will have a priority in global searching;otherwise, if it is too small, the aeDE will have a priority in localsearching. Therefore, it can be set to an adequate value based onthe characteristic of a particular problem. For example, for highlynonlinear and complex problems, the value of threshold shouldbe small (10�4 or 10�5 for threshold and 10�6 for tolerance). Onthe contrary, for small and simple problems, the value of thresholdshould be large (10�2 or 10�3 for threshold and 10�6 for tolerance).
4.4. Handling discrete variables for the aeDE
A simple method with a rounding function which permits tochange the continuous value of a result to the nearest discretevalue is utilized. The method is described as follows [16]
xdiscretei ¼ fixðxcontinuous
i Þ ð15Þwhere fix(x) is a function which rounds each element of x to thenearest permissible discrete value. In the same way, in the aeDE(Algorithm 2), before evaluating the fitness function, each individ-ual in the population will be rounded to the nearest discrete valuein the set of available discrete values by using the Eq. (15).
5. Numerical examples
In this section, the proposed aeDE method in Algorithm 2 isused to solve six truss optimization problems with discrete designvariables. The first three examples consist of the planar trusseshaving number of bars of 10, 52 and 200, respectively and the nextthree examples include space trusses having the number of bars of25, 72, and 160, respectively. The influence of the mutant factor F,crossover control parameter CR, population size NP and the thresh-old of the aeDE algorithm on the optimal solution is investigatedfor all examples; however, to avoid wordiness these investigationsis presented only for the first example of the 10 bar-truss problem.Based on obtained results of the investigations, some adequateparameters are recommended. In details, for the 10, 25, 52 and72 bar truss problems, the NP and threshold are set 20, and 10�3,respectively and for the 200 and 160 bar truss problems, the NPand threshold are respectively set 25, and 10�4. Besides, the values
(b)
0 2 4 6 8 10
-25
-15
-5
5
15
25
Str
ess
(ksi
)
Element number
Allowable limitCase 1Case 2
t the optimized design by the aeDE.
Fig. 8. A 52-bar planar truss structure.
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 65
of F and CR are set within the range [0.4,1.0] and [0.7,1.0], respec-tively, for all examples. The aeDE algorithm will be stopped eitherwhen the value of delta is greater than the value of tolerance orwhen the maximum number of iterations (MaxIter) is reached.Here, the values of tolerance and MaxIter are respectively set 10-6
and 2000 for all examples. The behavior of trusses is analyzed byfinite element method using two-node linear elements. All theproblems are implemented in Matlab and are run with 20 indepen-dent times. The gained results are verified in comparison to those
Table 6The available cross-section areas of the ASIC code.
obtained by other researches in the literature. The Matlab code ofall the problems can be obtained by sending an email to the firstauthor.
5.1. A 10-bar planar truss structure
In the first example, a simple 10-bar truss, as shown in Fig. 1 isconsidered. This work was previously investigated by manyresearchers such as Rajeev and Krishnamoorthy [10] using GeneticAlgorithm (GA), Ringertz [33] using Branch and Bound (BB), Li et al.[12] using Heuristic Paritcal Swarm Optimization (HPSO), andSadollah et al. [13] using Mine Blast Algorithm (MBA). The materialdensity is 0.1 lb/in.3 and the modulus of elasticity is 104ksi. Themembers are subjected to stress limitations of ±25 ksi. All nodesare subjected to displacement limitations of ±2.0 in. in both xand y directions. The vertical load P is 105 lbs. There are 10 designvariables corresponding to 10 bars and selected from the setS = {1.62,1.80,1.99,2.13,2.38,2.62, 2.63,2.88,2.93, 3.09,3.13,3.38,3.47,3.55,3.63,3.84,3.87,3.88,4.18,4.22,4.49,4.59,4.80,4.97,5.12,5.74,7.22,7.97,11.50,13.50,13.90,14.20,15.50,16.00,16.90,18.80,19.90,22.00,22.90,26.50,30.00,33.50} (in.2).
5.1.1. Influence of the parameters F, CR, NP and threshold on theoptimal solution
To get satisfactory parameters of CR, F, NP and threshold of theaeDE for this problem, the influence of these factors on the optimalsolutions is evaluated.
Table 1 and Fig. 2 show the affection of the mutant factor F,Table 2 and Fig. 3 describe influence of the crossover controlparameter CR. It can be found that if the value of F and CR are ran-domly generated in the interval [0.4,1.0] and [0.7,1.0] respectively,a capability of equalizing between the solution quality and compu-tational cost of the aeDE are better than those obtained by othervalues.
The impact of threshold on the optimum results is presented inTable 3 and Fig. 4. It can be seen that for the large value of threshold(e.g. 10�1, 10�2), the algorithm has fast convergences rate with lessnumber of structural analyses, but the results are not stable. Incontrast, with a smaller value of threshold (e.g. 10�4, 10�5), thealgorithm convergences more slowly, but the results are morestable. From Table 3 and Fig. 4, it can be recognized that the thresh-old = 10-3 is a good choice for this problem.
The influence of the population size NP on the results of theproblem is provided in Table 4 and Fig. 5. It can be realizedthat with the cases of NP > 15, the deviation of the mean weightsis not significant; however, the computational cost increases
Average of number of analyses – – – – – – 11314 3402
0 2000 4000 6000 8000 100001000
3000
5000
7000
9000
Wei
ght (
kg)
Number of analyses
DE: Mean valueDE: Best valueaeDE: Mean valueaeDE: Best value
Fig. 9. Comparison of convergence of the DE and aeDE for the 52-bar planar trussstructure.
0 10 20 30 40 50
-180
-80
0
80
180
Element number
Stre
ss (M
Pa)
Allowable limit
Fig. 10. Stress constraint boundaries of the example 5.2 evaluated at the optimizeddesign by the aeDE.
Fig. 11. A 200-bar planar truss structure.
66 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
significantly when the value of NP increases. For this problem, itcan be found that NP = 20 is a proper selection which has an abilityto effectively balance the computational cost and the quality of thesolution.
Table 8Comparison of optimized designs for the 200-bar planar truss structure.
Element number Members in the group Toǧan andDaloǧlu [35]
DE: Mean valueDE: Best valueaeDE: Mean valueaeDE: Best value
Fig. 12. Comparison of convergence of the DE and aeDE for the 200-bar planar trussstructure.
0 50 100 150 200
-10
-5
0
5
10
Stre
ss (k
si)
Element number
Allowable limitLoad case 1Load case 2Load case 3
Fig. 13. Stress constraint boundaries of the example 5.3 evaluated at the optimizeddesign by the aeDE.
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 67
Fig. 14. A 25-bar space truss structure.
Table 9Load condition for the 25-bar space truss structure.
Nodes Loads
Px (kips) Py (kips) Pz (kips)
1 1 �10 �102 0 �10 �103 0.5 0 06 0.6 0 0
0 500 1000 1500 2000
500
550
600
650
700
Wei
ght (
kg)
Number of analyses
DE: Mean valueDE: Best valueaeDE: Mean valueaeDE: Best value
Fig. 15. Comparison of convergence of the DE and aeDE for the 25-bar space trussstructure.
68 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
5.1.2. Comparison with other methods in the literatureA comparison of the obtained results by the present work and
some other methods in the literature is given in Table 5. The resultsshow that the optimumweights of both the DE and aeDE are betterthan those of other methods (5490.738 lb for the DE and aeDE,5613.84 lb for GA, 5531.98 lb for HPSO and 5507.75 lb for MBA).The aeDE requires less number of structural analyses than the DEand MBA (2380 structural analyses for the aeDE, 6440 analysesfor the DE and 3600 analyses for MBA). However, the MBA is more
Table 10Comparison of optimized designs for the 25-bar space truss structure.
Design variable (area in.2) Wu and Chow [11] Lee et al. [9] Li et al. [12] Kaveh and Ghazaan [38] Sadollah et al. [13] This study
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 69
stable than the aeDE and DE with the smallest standard deviation(11.38 lb for MBA, 19.521 lb for the DE and 20.780 lb for the aeDE).As shown in Fig. 6, the convergence of the aeDE is considerably fas-ter than that of the DE. After about 2500 structural analyses, theaeDE reaches the optimal result, while to obtain the same optimalresult, the DE needs about 5000 structural analyses. Fig. 7 demon-strates that constraint boundaries evaluated at the optimumdesign by the aeDE are satisfied the requirements of the problem.
5.2. A 52-bar planar truss structure
The second example considers the optimization problem for a52-bar planar truss structure as shown in Fig. 8 This problemwas previously solved by Lee et al. [9], Li et al. [12], Sadollahet al. [13,34], Kaveh and Mahdavi [16], etc. The modulus of elastic-ity and the mass density are E = 207 GPa and q = 7860 kg/m3,respectively. All the truss members are subject to stress limitationsof 180 MPa in both tension and compression. The vertical loadswere set equal to Px = 100 kN and Py = 200 kN. The members of thisstructure are divided into 12 groups: (1) A1–A4, (2) A5–A10, (3) A11–A13, (4) A14–A17, (5) A18–A23, (6) A24–A26, (7) A27–A30, (8) A31–A36,(9) A37–A39, (10) A40–A43, (11) A44–A49, and (12) A50–A52 corre-sponding to 12 design variables. The design variables are selectedfrom set of discrete value listed in Table 6.
The comparison of optimal results for this example is providedin Table 7. It can be found that the optimal weight gained by theaeDE agrees well with those of the DE and other methods. How-ever, the aeDE requires the least number of structural analyseswith only 3720, while the number of structural analyses obtainedby using the DE, HS, HPSO, MBA, CBO, WCA and IMBA are 13240,60000, 100000, 5450, 3840, 7100 and 4750, respectively. Fig. 9compares the convergence of the DE and aeDE. It again indicatesthat the aeDE converges much faster than the DE. Constraintboundaries of the problem evaluated at the optimum design areshown in Fig. 10. It exhibited that all constraints are satisfied.
5.3. A 200-bar planar truss structure
The third problem is a 200-bar truss as shown in Fig. 11. Thematerial density and the elasticity of modulus are 30000 ksi and0.283 lb/in.3, respectively. The 200 bars of the truss are categorizedinto 29 groups corresponding to 29 design variables. Stress limita-tions of ±10 ksi is adopted for the truss members. The truss is sup-ported to three loading conditions: (1) 1 kip acting in the positive xdirection at nodes 1, 6, 15, 20, 29, 34, 43, 48, 57, 62, and 71; (2) 10kips acting in the negative y direction at nodes 1, 2, 3, 4, 5, 6, 8, 10,12, 14, 15, 16, 17, 18, 19, 20, 22, 24, 26, 28, 29, 30, 31, 32, 33, 34, 36,38, 40, 42, 43, 44, 45, 46, 47, 48, 50, 52, 54, 56, 57, 58, 59, 60, 61, 62,64, 66, 68, 70, 71, 72, 73, 74, and 75; and (3) cases 1 and 2 are com-bined together. The available set of area is S = {0.100, 0.347, 0.440,0.539, 0.954, 1.081, 1.174, 1.333, 1.488, 1.764, 2.142, 2.697, 2.800,3.131, 3.565, 3.813, 4.805, 5.952, 6.572, 7.192, 8.525, 9.300, 10.850,13.330, 14.290, 17.170, 19.180, 23.680, 28.080, 33.700 in.2}. This
Table 12Comparison of optimized designs for the 72-bar space truss structure.
DE: Mean valueDE: Best valueaeDE: Mean valueaeDE: Best value
Fig. 18. Comparison of convergence of the DE and aeDE for the 72-bar space trussstructure.
(a)
0 10 20 30 40 50 60
-0.25
-0.15
-0.05
0.05
0.15
0.25
Dis
plac
men
t (in
.)
Number of displacements
Allowable limitLoad case 1Load case 2
Fig. 19. Constraint boundaries evaluated at the optimized design by
70 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
example was previously solved using an improved genetic algo-rithm (IGA) by Toǧan and Daloǧlu [35], a hybridized ant colony–harmony search-genetic algorithm called HACOHS-T by Talebpouret al. [36], and an elitist self-adaptive step-size search (ESASS) byAzad and Hasançebi [37].
Table 8 shows a comparison between the optimal designobtained by the present work and those obtained by some othermeta-heuristic algorithms in the literature. It can be observed thatthe aeDE outperforms most of the considered methods in both thequality of solution and number of structural analyses except for theESASS by Azad and Hasançebi [37]. The aeDE takes the best resultwith 27858.5 lb after 12325 analyses, while the IGA is28544.014 lb with 51360 analyses, the DE is 27901.583 lb with41475 analyses, the ARCGA is 28347.594 lb with 25000 analyses,the MABC is 28366.365 lb with 40000 analyses and the ESASS is28075.488 lb with 11156 analyses. Fig. 12 compares the conver-gence of the DE and aeDE. It indicates that the aeDE convergesmuch faster than the DE. Fig. 13 displays the stress constraintboundary of the problem evaluated at the optimum design. It exhi-bits that all constraints are satisfied.
(b)
0 10 20 30 40 50 60 70
-25
-15
-5
5
15
25
Stre
ss (k
si)
Element number
Allowable limitLoad case 1Load case 2
the aeDE. (a) Displacement constraints. (b) Stress constraints.
Fig. 20. A 160-bar space truss structure.
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 71
5.4. A 25-bar space truss structure
The fourth example considers the optimization problem for a25-bar space truss structure as shown in Fig. 14. This problemwas previously studied by Wu and Chow [11], Lee et al. [9], Liet al. [12], Sadollah et al. [13], Kaveh and Ghazaan [38], etc. Themodulus of elasticity is 104 ksi and the material density is 0.1 lb/in.3. All nodes are subjected to displacement limitations of ±0.35in. in three directions x, y, z. The stress limitations of the membersare ±40000 psi. The structure includes 25 members which aredivided into 8 groups as follows: (1) A1, (2) A2–A5, (3) A6–A9, (4)A10–A11, (5) A12–A13, (6) A14–A17, (7) A18–A21 and (8) A22–A25. Thedesign variables are selected from the set S = {0.1, 0.2, 0.3, 0.4,0.5, 0.6, 0.7, 0.8, 0.9, 1.0, 1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9,2.0, 2.1, 2.2, 2.3, 2.4, 2.6, 2.8, 3.0, 3.2, 3.4} (in.2). The loads areshown in Table 9.
Table 13Load cases for the 160-bar space truss structure.
Table 10 compares the optimal results of the aeDE and othermethods. It can be seen that all methods give the same value ofthe minimum weight with 484.85 lb except SGA with 486.29 lb.The number of structural analyses of the aeDE is the smallestamong all considered methods. It takes the optimal weight after1440 analyses, while the SGA, HS, HSPO, MBA, ECBO and DE take40000, 13523, 3750, 2150, 7050 and 3500, respectively. The con-vergence of the aeDE and DE and constraint boundaries evaluatedat the optimum design by the aeDE are shown in Fig. 15 and Fig. 16,respectively. It can be seen that the aeDE converges faster than theDE and no constraints are violated.
5.5. A 72-bar space truss structure
The five example executes the optimization problem for a 72-bar space truss structure as shown in Fig. 17. This problem waspreviously examined by Wu and Chow [11], Lee et al. [9], Li et al.[12], Sadollah et al. [13,34], Kaveh and Mahdavi [16], Kaveh andTalatahari [39], Kaveh and Ghazaan [38], etc. The material densityis 0.1 lb/in.3 and the modulus of elasticity is 104 ksi. The stress lim-itations of the members are ±25,000 psi. All nodal displacementsmust be smaller than ±0.25 in. There are 72 truss elements whichare divided into 16 groups: (1) A1–A4, (2) A5–A12, (3) A13–A16, (4)A17–A18, (5) A19–A22, (6) A23–A30, (7) A31–A34, (8) A35–A36, (9) A37–A40, (10) A41–A48, (11) A49–A52, (12) A53–A54, (13) A55–A58, (14)A59–A66 (15), A67–A70, and (16) A71–A72. The design variables areselected from Table 6. This structure was designed for two separateload conditions given in Table 11.
The results gained by this work in comparison with previousstudies is presented in Table 12. It can be seen that the aeDEacquires the optimal weight which agrees well with those acquiredby different methods and has the least number of structural anal-yses. Fig. 18 demonstrates that the aeDE reaches the optimumsolutions much faster than the DE. Fig. 19 shows that all con-straints of the problem evaluated at the optimum results by theaeDE are satisfied.
5.6. A 160-bar space truss structure
The last example is a 160-bar truss shown in Fig. 20. This prob-lem has been solved using selective dynamic rounding (SDR) byGroenwold and Stander [40], the regional genetic algorithm(RGA) by Groenwold et al. [41], and rank-based ant system (RBAS)by Capriles et al. [42]. The material density is 0.00785 kg/cm3 andthe modulus of elasticity is 2.047 � 106 kgf/cm2. The 160 membersof the truss are linked to 38 independent design variables. The
DE: Mean valueDE: Best valueaeDE: Mean valueaeDE: Best value
Fig. 21. Comparison of convergence of the DE and aeDE for the 160-bar space trussstructure.
72 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
members’ linkage and the nodal coordinates of this truss is given inAppendix A. The structure is designed to subject to the eight inde-pendent load cases given in Table 13. Buckling stress constraintsfor compression members are considered. For a member undercompression, the buckling stress is calculated as: rb = 1300 –(kl/r)2/24 if kl/r 6 120, and rb = 107/(kl/r)2 if kl/r > 120, where l isthe length of the member, r is the radius of gyration, and k is theeffective length factor. For this problem k is supposed to be 1.0.The cross-section areas and the corresponding radii of gyrationfor the 42 prescribed discrete sections are S = {1.84, 2.26, 2.66,3.07, 3.47, 3.88, 4.79, 5.27, 5.75, 6.25, 6.84, 7.44, 8.06, 8.66, 9.40,10.47, 11.38, 12.21, 13.79, 15.39, 17.03, 19.03, 21.12, 23.20,25.12, 27.50, 29.88, 32.76, 33.90, 34.77, 39.16, 43.00, 45.65,46.94, 51.00, 52.10, 61.82, 61.90, 68.30, 76.38, 90.60, 94.13 cm2},and r = {0.47, 0.57, 0.67, 0.77, 0.87, 0.97, 0.97, 1.06, 1.16, 1.26,1.15, 1.26, 1.36, 1.46, 1.35, 1.36, 1.45, 1.55, 1.75, 1.95, 1.74, 1.94,2.16, 2.36, 2.57, 2.35, 2.56, 2.14, 2.33, 2.97, 2.54, 2.93, 2.94, 2.94,2.92, 3.54, 3.96, 3.52, 3.51, 3.93, 3.92, 3.92 cm}.
The best solution vector for the thirty-eight design variablesacquired by this work and some methods in the literature are
(a) (b)
0 50 100 150
0
0.2
0.4
0.6
0.8
1
Stre
ss ra
tio
Element number
Load case 1Load case 2Load case 3Load case 4
0 50 60 150
0
0.2
0.4
0.6
0.8
1
Stre
ss ra
tio
Element number
Load case 5Load case 6Load case 7Load case 8
Fig. 22. Constraint boundaries evaluated at the optimized design by the aeDE. (a) Load case 1–4. (b) Load case 5–8.
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 73
presented in Table 14. The results show that the aeDE is the mosteffective algorithm in all considered methods. The optimizationweight of the aeDE is 1336.634 kg. It is much lighter than thosegained by the DE with 1448.306 kg, SDR with 1359.781 kg, RGAwith 1337.442 kg and RBAS with 1348.905 kg. Moreover, the aeDEis much efficient than the DE in term of computational cost. TheaeDE only needs 23925 analyses to reach the best solution, whilethe DE requires 50025 analyses to reach the best solution. Fig. 21shows comparison of the convergence of the DE and aeDE.Fig. 22 describes the existing buckling stress ratios of the problemevaluated at the optimum design by the aeDE. It shows that noconstraints are violated.
6. Conclusion
In this work, three modifications of the DE are proposed to givea so-called the adaptive elitist differential evolution (aeDE) forsolving truss optimization problems with discrete design variables.
Table 15Coordinate data for the 160-bar space truss structure.
Firstly, an adaptive technique based on the absolute deviation ofobjective function between the best individual and the entire pop-ulation in the previous generation is proposed to select a suitablemutation operator. This technique helps preserve the balancebetween global and local searching abilities. Secondly, an elitistselection technique is used in the selection phase to enhance theconvergence rate of the algorithm. Thirdly, a rounding techniqueis integrated into the aeDE to deal with optimization problemswith discrete design variables.
The aeDE algorithm is applied to solve six optimization prob-lems of truss structures with discrete design variables. The numer-ical results demonstrate that in most of the cases the aeDE caneffectively attain the optimum solutions with less iterations thanthe DE and some other methods in the literature. Particularly,the number of structural analysis obtained by the aeDE is approx-imately a third of that obtained by the DE. Especially, for the largeproblems of 200 and 160 bar trusses, the optimal solutionsobtained by the aeDE are better than those found by the DE and
74 V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75
some methods in the literature. These illustrate that with properimprovements, the aeDE can offer a robust, effective and reliableoptimization method for solving optimizations of truss structureswith discrete design variables.
In addition, the aeDE is quite similar to the standardDE. Thus, it issimple tounderstandand implement.Hence, it canbe easy to extendfor various engineering optimization problems such as optimizationof frame, composite plates/shell structures, and stiffened plates/shell structures. Furthermore, with fast convergence rate, the aeDEcan be also applied for the reliability-based design optimizationproblems where the computational cost is always a concern.
Acknowledgements
This research is funded by Vietnam National Foundation forScience and Technology Development (NAFOSTED) under grantnumber 107.99-2014.11.
Appendix A
See Tables 15 and 16.
References
[1] Lamberti L, Pappalettere C. Comparison of the numerical efficiency of differentsequential linear programming based algorithms for structural optimisationproblems. Comput Struct 2000;76:713–28. http://dx.doi.org/10.1016/S0045-7949(99)00185-6.
[2] Lamberti L, Pappalettere C. Move limits definition in structural optimizationwith sequential linear programming. Part II: Numerical examples. ComputStruct 2003;81:215–38. http://dx.doi.org/10.1016/S0045-7949(02)00443-1.
[3] Lamberti L, Pappalettere C. Improved sequential linear programmingformulation for structural weight minimization. Comput Methods Appl MechEng 2004;193:3493–521. http://dx.doi.org/10.1016/j.cma.2003.12.040.
[4] Sedaghati R. Benchmark case studies in structural design optimization usingthe force method. Int J Solids Struct 2005;42:5848–71. http://dx.doi.org/10.1016/j.ijsolstr.2005.03.030.
[5] NS Khot LB, Venkayya VB. Comparison of optimality criteria algorithms forminimum weight design of structures. AIAA J 1979;17:182–90. http://dx.doi.org/10.2514/3.61093.
[6] Ko F-T, Wang BP. An improved method of optimality criteria for structuraloptimization. Comput Struct 1991;41:629–36. http://dx.doi.org/10.1016/0045-7949(91)90175-L.
[7] Kaveh A, Kalatjari V. Topology optimization of trusses using genetic algorithm,force method and graph theory. Int J Numer Methods Eng 2003;58:771–91.http://dx.doi.org/10.1002/nme.800.
[8] Farshi B, Alinia-Ziazi A. Sizing optimization of truss structures by method ofcenters and force formulation. Int J Solids Struct 2010;47:2508–24. http://dx.doi.org/10.1016/j.ijsolstr.2010.05.009.
[9] Lee KS, Geem ZW, Lee S, Bae K. The harmony search heuristic algorithm fordiscrete structural optimization. Eng Optim 2005;37:663–84. http://dx.doi.org/10.1080/03052150500211895.
[10] Rajeev S, Krishnamoorthy CS. Discrete optimization of structures using geneticalgorithms. J Struct Eng 1992;118:1233–50.
[12] Li LJ, Huang ZB, Liu F. A heuristic particle swarm optimization method for trussstructures with discrete variables. Comput Struct 2009;87:435–43. http://dx.doi.org/10.1016/j.compstruc.2009.01.004.
[13] Sadollah A, Bahreininejad A, Eskandar H, Hamdi M. Mine blast algorithm foroptimization of truss structures with discrete variables. Comput Struct2012;102–103:49–63. http://dx.doi.org/10.1016/j.compstruc.2012.03.013.
V. Ho-Huu et al. / Computers and Structures 165 (2016) 59–75 75
[14] Sonmez M. Discrete optimum design of truss structures using artificial beecolony algorithm. Struct Multidiscip Optim 2011;43:85–97. http://dx.doi.org/10.1007/s00158-010-0551-5.
[15] Dede T. Application of teaching-learning-based-optimization algorithm for thediscrete optimization of truss structures. KSCE J Civ Eng 2014;00:1–9. http://dx.doi.org/10.1007/s12205-014-0553-8.
[16] Kaveh A, Mahdavi VR. Colliding bodies optimization method for optimumdiscrete design of truss structures. Comput Struct 2014;139:43–53. http://dx.doi.org/10.1016/j.advengsoft.2014.01.002.
[17] Storn R, Price K. Differential evolution – a simple and efficient heuristic forglobal optimization over continuous spaces. J Glob Optim 1997;11:341–59.http://dx.doi.org/10.1023/A:1008202821328.
[18] Wang ZWZ, Tang HTH, Li PLP. Optimum design of truss structures based ondifferential evolution strategy. In: 2009 int conf inf eng comput sci 2009. p. 0–4. http://dx.doi.org/10.1109/ICIECS.2009.5365996.
[20] Le-Anh L, Nguyen-Thoi T, Ho-Huu V, Dang-Trung H, Bui-Xuan T. Static andfrequency optimization of folded laminated composite plates using anadjusted differential evolution algorithm and a smoothed triangular plateelement. Compos Struct 2015;127:382–94. http://dx.doi.org/10.1016/j.compstruct.2015.02.069.
[21] Ho-Huu V, Nguyen-Thoi T, Nguyen-Thoi MH, Le-Anh L. An improvedconstrained differential evolution using discrete variables (D-ICDE) forlayout optimization of truss structures. Expert Syst Appl 2015;42:7057–69.http://dx.doi.org/10.1016/j.eswa.2015.04.072.
[22] Rogalsky T, Derksen RW, Kocabiyik S. Differential evolution in aerodynamicoptimization. In: Proc. 46th annu. conf. can. aeronaut. sp. inst., 1999. p. 29–36.
[23] Fan H-Y, Lampinen J. A trigonometric mutation operation to differentialevolution. J Glob Optim 2003;27:105–29. http://dx.doi.org/10.1023/A:1024653025686.
[24] Kaveh A, Ilchi Ghazaan M. Hybridized optimization algorithms for design oftrusses with multiple natural frequency constraints. Adv Eng Softw2015;79:137–47. http://dx.doi.org/10.1016/j.advengsoft.2014.10.001.
[25] Storn R. On the usage of differential evolution for function optimization. In:Fuzzy inf process soc 1996 NAFIPS, 1996 Bienn conf North Am 1996. p. 519–23. http://dx.doi.org/10.1109/NAFIPS.1996.534789.
[26] Ilonen J, Kamarainen J-K, Lampinen J. Differential evolution training algorithmfor feed-forward neural networks. Neural Process Lett 2003;17:93–105. http://dx.doi.org/10.1023/A:1022995128597.
[27] Wang Y, Cai Z, Zhang Q. Differential evolution with composite trial vectorgeneration strategies and control parameters. IEEE Trans Evol Comput2011;15:55–66. http://dx.doi.org/10.1109/TEVC.2010.2087271.
[28] Das S, Abraham A, Chakraborty UK, Konar A. Differential evolution using aneighborhood-based mutation operator. IEEE Trans Evol Comput2009;13:526–53. http://dx.doi.org/10.1109/TEVC.2008.2009457.
[29] Qin AK, Huang VL, Suganthan PN. Differential evolution algorithm withstrategy adaptation for global numerical optimization. IEEE Trans Evol Comput2009;13:398–417. http://dx.doi.org/10.1109/TEVC.2008.927706.
[30] Jia G, Wang Y, Cai Z, Jin Y. An improved (muy + lamda)-constrained differentialevolution for constrained optimization. Inf Sci (Ny) 2013;222:302–22. http://dx.doi.org/10.1016/j.ins.2012.01.017.
[31] Mohamed AW, Sabry HZ. Constrained optimization based on modifieddifferential evolution algorithm. Inf Sci (Ny) 2012;194:171–208. http://dx.doi.org/10.1016/j.ins.2012.01.008.
[32] Padhye N, Bhardawaj P, Deb K. Improving differential evolution through aunified approach. J Glob Optim 2013;55:771–99. http://dx.doi.org/10.1007/s10898-012-9897-0.
[33] Ringertz UT. On methods for discrete structural optimization. Eng Optim1988;13:47–64. http://dx.doi.org/10.1080/03052158808940946.
[34] Sadollah A, Eskandar H, Bahreininejad A, Kim JH. Water cycle, mine blast andimproved mine blast algorithms for discrete sizing optimization of trussstructures. Comput Struct 2015;149:1–16. http://dx.doi.org/10.1016/j.compstruc.2014.12.003.
[35] Toǧan V, Daloǧlu AT. An improved genetic algorithm with initial populationstrategy and self-adaptive member grouping. Comput Struct2008;86:1204–18. http://dx.doi.org/10.1016/j.compstruc.2007.11.006.
[36] Talebpour MH, Kaveh A, Kalatjari VR. Optimization of skeletal structures usinga hybridized ant colony–harmony search-genetic algorithm. Iran J Sci TechnolTrans Civ Eng 2014;38:1–20.
[37] Azad SK, Hasançebi O. An elitist self-adaptive step-size search for structuraldesign optimization. Appl Soft Comput 2014;19:226–35. http://dx.doi.org/10.1016/j.asoc.2014.02.017.
[38] Kaveh A, Ilchi Ghazaan M. A comparative study of CBO and ECBO for optimaldesign of skeletal structures. Comput Struct 2015;153:137–47. http://dx.doi.org/10.1016/j.compstruc.2015.02.028.
[39] Kaveh A, Talatahari S. A particle swarm ant colony optimization for trussstructures with discrete variables. J Constr Steel Res 2009;65:1558–68. http://dx.doi.org/10.1016/j.jcsr.2009.04.021.
[40] Groenwold AA, Stander N. Optimal discrete sizing of truss structures subject tobuckling constraints. Struct Optim 1997;14:71–80. http://dx.doi.org/10.1007/BF01812508.
[41] Groenwold AA, Stander N, Snyman JA. A regional genetic algorithm for thediscrete optimal design of truss structures. Int J Numer Methods Eng1999;44:749–66. http://dx.doi.org/10.1002/(SICI)1097-0207(19990228)44:6<749::AID-NME523>3.0.CO;2-F.
[42] Capriles PVSZ, Fonseca LG, Barbosa HJC, Lemonge ACC. Rank-based ant colonyalgorithms for truss weight minimization with discrete variables. CommunNumer Methods Eng 2007;23:553–75. http://dx.doi.org/10.1002/cnm.912.