Information Sciences: Yao-Chen Chuang, Chyi-Tsong Chen, Chyi Hwang

INS 11379 No.
of Pages 29, Model 3G

14 February 2015
Information Sciences xxx (2015) xxx–xxx
1
Contents lists available at ScienceDirect
Information Sciences
journal homepage: www.elsevier.com/locate/ins
5
6
3 A real-coded genetic algorithm with a direction-based

4 crossover operator
7 Yao-Chen Chuang a, Chyi-Tsong Chen a,b,⇑, Chyi Hwang b
8 a
Department of Chemical Engineering, Feng Chia University, No. 100 Wenhwa Rd., Seatwen, Taichung 407, Taiwan
9 b
Center of General Education, National Quemoy University, No. 1 University Rd., Jinning Township, Kinmen 892, Taiwan
10
a r t i c l e i n f o a b s t r a c t
1
2 2
7
13 Article history: In this paper, we develop a parallel-structured real-coded genetic algorithm (RCGA), 28
14 Received 5 October 2012 named the RGA-RDD, for numerical optimization. Technically, the proposed RGA-RDD 29
15 Received in revised form 14 November 2014 integrates three specially designed evolutionary operators – the Ranking Selection (RS), 30
16 Accepted 30 January 2015
Direction-Based Crossover (DBX), and the Dynamic Random Mutation (DRM) – as a whole 31
17 Available online xxxx
to mimic a specific evolutionary process. Unlike the conventional RCGAs that perform evo- 32
lutionary operators in a series framework, the RGA-RDD embeds a coordinator in the inner 33
18 Keywords:
parallel loop to organize the operations of the DBX and DRM so that a higher possibility of 34
19 Evolutionary algorithm
20 Real-coded genetic algorithm
locating the global optimum is ensured. Besides, based on the results of a systematic para- 35
21 Parallel structure metric analysis, we provide a parameter selection guideline for the settings of the proposed 36
22 Real-parameter optimization RGA-RDD. Furthermore, a data-driven optimization scheme, which incorporates the uni- 37
23 Data-driven optimization scheme form design for design of experiments and a shape-tunable neural network for auxiliary 38
24 Design of experiments decision support, is applied to search for an optimal set of the algorithm parameters. 39
25 Parameter tuning The effectiveness and applicability of the proposed RGA-RDD are demonstrated through 40
26
a variety of benchmarked optimization problems, followed by comprehensive comparisons 41
with some existing state-of-the-art evolutionary algorithms. Extensive simulation results 42
reveal that the performance of the proposed RGA-RDD is superior to comparative methods 43
in locating the global optimum for real-parameter optimization problems, especially for 44
unsolved multimodal and high-dimensional hybrid functions. 45
Ó 2015 Elsevier Inc. All rights reserved. 46
47
48
49
50 1. Introduction
51 1.1. A review of conventional real-coded genetic algorithms
52 In the past few decades, many different types of evolutionary algorithms (EAs) have been proposed to solve optimization
53 problems resulted from diversified fields of science, economics, engineering, and so on. According to the evolutionary mech-
54 anisms used, EAs are conventionally classified into the following categories: genetic algorithms (GAs) [7,34,36], evolutionary
55 strategies (ESs) [33,39,60], genetic programming (GP) [69,92], evolutionary programming (EP) [15,57], differential evolution
56 (DE) [19], and those inspired from biology and nature [25,26,44]. Among the developed EA schemes, the genetic algorithms
57 (GAs) are widely recognized as one of the most popular and commonly used optimization methods, and based on which
⇑ Corresponding author at: Department of Chemical Engineering, Feng Chia University, No. 100 Wenhwa Rd., Seatwen, Taichung 407, Taiwan. Tel.: +886
82 313971; fax: +886 82 313304.
E-mail address: ctchen@nqu.edu.tw (C.-T. Chen).
http://dx.doi.org/10.1016/j.ins.2015.01.026
0020-0255/Ó 2015 Elsevier Inc. All rights reserved.
Please cite this article in press as: Y.-C. Chuang et al., A real-coded genetic algorithm with a direction-based crossover operator, Inform. Sci.
(2015), http://dx.doi.org/10.1016/j.ins.2015.01.026
INS 11379 No. of Pages 29, Model 3G
14 February 2015
2 Y.-C. Chuang et al. / Information Sciences xxx (2015) xxx–xxx
58 many successful applications have been reported in the literature [13,14,27,49,87,92]. Basically, the GA is a kind of popula-
59 tion-based stochastic searching approach that mimics the natural selection and survival of the fittest in the biological world.
60 We note that, among many available coding schemata such as Gray coding [71], integer genes [56], etc., the real-coded
61 genetic algorithms (RCGAs) [34,35] present to be the most intuitive type of GAs for solving real-parameter optimization
62 problems because of the ability of representing the numerical solution directly with a real number [50,63] and without
63 requiring the coding and decoding procedures. Furthermore, in addition to the theoretical studies [41,46,71,72] that have
64 been conducted to prove the solution efficiency and stability of RCGAs, several remarkable works [30,40,43,55,65,85] have
65 revealed that the RCGAs outperform several counterpart GAs in many real-world optimization applications, especially for
66 high-dimensional or high-accuracy problems.
67 From a technical point of view, the RCGAs essentially make use of three fundamental evolutionary operators – selection,
68 crossover, and mutation – to search for an optimal solution. Each operator is featured with a specific mechanism to approach
69 better solutions. The selection operator chooses those potential chromosomes among the whole population, and those
70 selected will receive the subsequent crossover manipulation. The crossover operator generates some new candidate chromo-
71 somes (offspring) by recombining the genetic information gained from the selected parent chromosomes. The mutation
72 operator randomly changes the gene of chromosomes in order to prevent premature convergence; as a consequence, it pro-
73 vides an opportunity to avoid being trapped by a local optimum. The conventional RCGAs commonly operate these three
74 kinds of operators in a series framework as shown in Fig. 1. Despite many successful applications of using the conventional
75 RCGAs were reported, we found from a literature survey that there has been an increasing interest in improving the RCGA’s
76 solution efficiency. As a brief summary, we classify the recent attempts made to cultivate the RCGAs into the following
77 categories: the determination of the optimal population size [1,28,37,48], the discovery of effective methods to initialize
Start
Initialize population
k = k+1
Selection Operation
Crossover Operation
Mutation Operation
Replacement Operation
No
Stop criteria met?
Yes
Stop
Fig. 1. The flow chart of the conventional RCGA optimization scheme.
14 February 2015
Y.-C. Chuang et al. / Information Sciences xxx (2015) xxx–xxx 3
78 population [10,58,59], the exploration of automatic schemes for operator’s parameter tuning [2,82,83], the control of pop-
79 ulation diversity [3,42,54,68], the improvement of existing crossover methods [23,63,67], the development of new crossover
80 schemes [8,20,24,30,37,66,75,90,91,95], the investigation of novel evolutionary strategies [8,16,21,79], the use of hybrid
81 operators [4,29,78,97], and some others [22,51,86].
82 1.2. Related works on the development of crossover operators
83 In view of the above-mentioned attempts, we discovered that a great emphasis has been placed on the efforts of devel-
84 oping or improving crossover operators. This is attributed to the fact that, among the three fundamental operators, the cross-
85 over operator is the most important evolution mechanism because it greatly assists the algorithms to achieve the optimum
86 by producing some new candidate chromosomes. To discover some more in details, we briefly review those representative
87 works related to the crossover developments as follows. Based on the concept of interval schemata, Eshelman and Schaffer
88 [30] proposed a blend crossover operator, named BLX-a, by which the parameter a is used to control the location where one
89 offspring solution is to be generated from the parents. Their numerical experiences reported that the best parameter value
90 for BLX-a operator is a = 0.5; this suggested a value indicates that the BLX-a prefers generating the offspring from the center
91 of the chosen two parents rather than other locations. Alternatively, Deb and Agrawal [20] developed a simulated binary
92 crossover (SBX) to generate two offspring at a time from two selected parents. Functioned in the continuous domain, the
93 SBX operator simply simulates the principle of the single-point crossover operator by incorporating some binary-coded
94 strings. Using the SBX operator, the offspring is favorably generated at the site closer to parents rather than away from
95 the chosen two parents. Based on the use of an ellipsoidal probability distribution, Ono and Kobayashi [66] proposed a uni-
96 modal normal distribution crossover (UNDX) operator by which two or more offspring solutions are created from three
97 selected parent chromosomes. Generally, the UNDX operator assigns a higher probability to create solutions that are close
98 to the center of the first two parents and a considerably lower probability for those adjacent to the parents themselves. By
99 using a technique of uniformly sampling, Tsutsui et al. [91] proposed a simplex crossover (SPX) operator to generate off-
100 spring from a simplex formed by three or more parent chromosomes. They concluded from some experiments that the
101 SPX is more suitable for functions having multimodality and/or epistasis as the simplex is formed with a medium number
102 of parents. For example, SPX prefers using 3-parent for a low dimensional function and 4 parents for a higher dimensional
103 function. Lately, Deb et al. [21] proposed a parent centric crossover operator, named PCX, for use in association with EAs.
104 They showed by some test problems that the PCX-based RCGA outperforms several comparative schemes in locating the glo-
105 bal optimum. More recently, Ripon et al. [75] proposed a jumping-gene RCGA to exploit local search heuristics by manipu-
106 lating the gene to jump from one position to another under multiple stresses. Deep and Thakur [24] proposed a specific
107 Laplace crossover operator (LX) that uses Laplace distribution as the density function to control the location where an off-
108 spring is generated. Alternatively, Yoon et al. [95] proposed an unbiased boundary-extended crossover that has an advantage
109 of removing the inherent bias caused by some existing crossover operators. More recently, Kuo and Lin [50] developed a
110 directed crossover operator by utilizing the reflection mechanism of the Nelder-Mead’s simplex method (NMSM). In addition
111 to the above developments, some remarkable comparison studies regarding the crossover operators have been reported,
112 which can be found in [8,21,24,37,75,95].
113 From the above brief survey, we found that the line segment connection and the distribution analysis of parent solutions
114 are the two most commonly used techniques to develop crossover operators. For example, UNDX, SPX, and BLX are mean-
115 centric approaches developed based on line-segment connection, and the LX, SBX and PCX are parent-centric ones derived
116 from the distribution analysis of parents. In spite of many successful applications, it has been pointed out in several related
117 studies [8,66,90] that the crossover operators featured by line segment connection or distribution analysis of parent solu-
118 tions may encounter difficulties when applied in some specific and stringent applications. First, as the size of population
119 is relatively small compared to the whole searching space, and/or the initial population distribution does not scatter uni-
120 formly over the admissible domain, these two kinds of crossover operators can generate offspring ineffectively in some spe-
121 cific regions that are uncovered by the initial population [66], which therefore significantly degrades the algorithm’s
122 performance. Second, these crossover operators may not work effectively in locating the global optimum as it is adjacent
123 to or exactly sits on the boundaries of the feasible search space [91].
124 1.3. The contribution of this work
125 To tackle with the above-mentioned difficulties, in this paper, we convey the idea of the gradient search [17] to develop a
126 direction-based crossover operator, named DBX, for the RCGAs. Quite different from the crossover operators developed based
127 on the line-segment connection or distribution analysis, the DBX takes the relative fitness information into account to build a
128 gradient (direction) for offspring generation. More precisely, by using the relative fitness information of two selected chro-
129 mosomes, the DBX operator is able to effectively direct the crossover toward a right direction to generate two offspring
130 whose fitness is much better than each of their parents. Thus, even the initial population distribution is not uniformly scat-
131 tered in the searching domain, the direction-based crossover operator still enables the derivation of effective crossover direc-
132 tions to generate potential offspring. Especially, the DBX carries out 2n 1 crossover scenarios to explore the possibly better
133 offspring, which therefore provides a higher success rate in locating the global optimum. Besides, to ensure a dynamic
134 crossover evolution, the DBX applies an adaptive scheme [17] to control the step size of the crossover over the successive
14 February 2015
135 generations. The step-size adjustment scheme is performed based on a measure that instantly reflects the ratio of fitness
136 distance of the paired parents and the maximum fitness difference extracted from the current population. Besides, a simple
137 strategy, which fixes the chromosomes that are guided to move across the boundary of the feasible region on their extreme
138 bounds, is appended to the function of DBX to increase the possibility of locating the global optimum that sits on the bound-
139 ary. In addition, the DBX is featured with a specific and simple scheme to avoid generating a null crossover direction; this
140 newly added feature makes the DBX distinct from its original version in [17]. Here, we note that since the DBX is able to
141 explore the best crossover direction from the generated 2n 1 scenarios, it is quite different from the directed crossover
142 operator constructed using the reflection mechanism of the NMSM [50] and the directed crossover developed based on a tree
143 structure [51]. Furthermore, the RGA-RDD equips a coordinator in the inner parallel loop to organize the operation of the
144 DBX and DRM in order to increase the population diversity. Because the DRM is able to adaptively regulate the admissible
145 mutation size during evolution, it not only prevents the searching from being trapped by a local optimum, but also has the
146 capability of improving the solution accuracy when the evolution is about to reach the final stage of convergence. In addition,
147 an RS operator [17] is integrated into the RGA-RDD to eliminate those chromosomes having worse fitness and, at the same
148 time, to reproduce the same amount of chromosomes that are with better fitness. To facilitate a better understanding of
149 these three RGA-RDD’s operators, in this paper, we perform a series of basic studies on a set of test problems. Besides, based
150 on the results of a parametric analysis, a parameter selection guideline for the proposed RGA-RDD is provided and the inter-
151 action of algorithm parameters is explored. Moreover, by applying a DDOS for parameter tuning, we report an optimal set of
152 RGA-RDD parameters for a better handling of the test problems. The effectiveness and applicability of the proposed RGA-
153 RDD are demonstrated through a variety of benchmark optimization problems, followed by an extensive comparison with
154 some existing state-of-the-art EAs. The extensive simulation results reveal that the proposed RGA-RDD is a simple yet pow-
155 erful algorithm for real-parameter numerical optimization; it enables a considerably faster convergence speed and a signif-
156 icantly higher success rate in locating the global optimum than the comparative methods, especially for unsolved
157 multimodal and high-dimensional hybrid functions. We note that the RGA-RDD presented in this paper significantly
158 advances our tentative work [17] of using the RS, DBX, and DRM operators by comprising several new features, ideas, and
159 novelties. First, a newly designed parallel-structured coordinator is used to organize the operation of the DRM and DBX; this
160 effort has been proven to be very effective in locating the global optimum. Second, the generational replacement (GR) with
161 an elitism strategy is additionally added into the algorithm to perform the survivor selection in order to preserve the best
162 found solution in the subsequent evolution. Third, a scheme of alleviating stagnation is implemented to avoid being trapped
163 by a local optimum and to divert the situation of fast converging to a local optimum. In addition to these major improve-
164 ments, we also modify the DBX by adding a simple strategy to prevent from generating a null crossover direction; this effort
165 makes the DBX very effective in generating better offspring. Furthermore, this paper completes a parametric analysis so as to
166 provide a useful parameter selection guideline for the proposed RGA-RDD. An experimental design using a data-driven opti-
167 mization technique is also performed in this paper to search for an optimal set of algorithm parameters. In addition to the
168 above-mentioned contributions, this paper greatly broadens its scope by evaluating the RGA-RDD’s performance on the
169 CEC2005 testbed and completing the performance comparisons with some existing state-of-the-art algorithms.
170 1.4. The organization of this work
171 The remainder of this paper is organized as follows. In Section 2, we present the proposed parallel-structured RGA-RDD,
172 its operational flowchart, and the associated pseudo code. Subsequently, in Section 3 we systematically investigate the pro-
173 posed parallel-structured RGA-RDD by exploring the effectiveness and applicability of each of the three specially designed
174 operators, DBX, DRM, and the RS. Therein, we complete a parametric analysis and an experimental design using the DDOS in
175 order to provide a parameter selection guideline and an optimal set of RGA-RDD parameters for a specific application. In
176 addition, the interaction of the algorithm parameters will be explored using a DOE and demonstrated with response surface
177 plots. In Section 4, we perform extensive comparison studies on benchmark optimization problems, where the experimental
178 setup and the algorithms to be compared are discussed in a systematic way. Finally, in Section 5 concluding remarks are
179 given to the proposed parallel-structured RGA-RDD for real-parameter numerical optimization.
180 2. The proposed RGA-RDD
181 2.1. The algorithm structure and the underlying ideas
182 Fig. 2 depicts the algorithm configuration of the proposed RGA-RDD, where the three evolutionary operators – RS, DBX,
183 and DRM – are integrated as a whole to mimic a special evolutionary process that has an inner parallel loop to coordinate the
184 operations of the DBX and DRM. We note that, in additional to the specially designed operators, the imbedded parallel oper-
185 ation loop is the essential core that distinguishes the proposed parallel-structured RGA-RDD from the conventional RCGAs
186 that operate the evolutionary operators in a series framework. Also noted is that the proposed algorithm structure is quite
187 different from that used in our previous work [17] since a different version of the parallel-structure coordinator is applied
188 and a stagnation prevention scheme is additionally added to keep the evolution away from reaching the condition of pre-
189 mature stagnation and to successfully locate the global optimum. Before presenting the specific operational functions of
14 February 2015
Fig. 2. The flow chart of the proposed RGA-RDD optimization scheme.
190 the RGA-RDD, we introduce the underlying design concept and notations associated with RCGAs as follows. Let
191 f ðxÞ : X # Rn ! R be the objective function to be minimized, where X is the admissible search space. If xopt is the optimum
192 solution which minimizes f ðxÞ, then we have f ðxopt Þ 6 f ðxÞ 8x 2 X. To search for xopt by using the proposed RGA-RDD, we
193 first let h ¼ ½x1 ; x2 ; . . . ; xn be a solution termed as chromosome/individual in the sense of RCGAs. In the chromosome, each
194 and n
xj ; j 2 n ¼ f1; 2; . . . ; ng, is called a gene and directly represented as a real number. Suppose that the lower and upper
195 bounds of each gene are known, the admissible parameter space for h can be characterized as follows:
196
198 Xh ¼ fh 2 Rn jx1;min 6 x1 6 x1;max ; x2;min 6 x2 6 x2;max ; . . . ; xn;min 6 xn 6 xn;max g: ð1Þ

199 For a chromosome h 2 Xh , we have hL 6 h 6 hU ,where hL x1;min ; x2;min ; . . . ; xn;min and hU ½x1;max ; x2;max ; . . . ; xn;max . Having
200 completed the description of those required notations and basic concept about the real-number expression of a gene, we
201 are ready to introduce the operational procedure and operators associated with the proposed RCA-RDD.
14 February 2015
202 2.2. The RS: a Ranking Selection operator
203 Among the diversified techniques developed for chromosome reproduction [9], for example the elitist, disruptive,
204 dynamic probability selection, and so on, the roulette wheel selection (RWS) and the tournament selection (TS) are the
205 two most commonly used mechanisms in association with RCGAs. The RWS can be visualized by imagining a wheel where
206 each solution occupies an area, and the size of the area is directly related to its fitness. When a spinning wheel stops, a fixed
207 marker determines which chromosome will be selected to enter into the mating pool to receive the subsequent crossover
208 operation. Since the RWS involves heavy computation at each generation, the speed of evolution would be decreased accord-
209 ingly. In contrast, the scheme of TS can be simply operated based on the competition of individual fitness; that is, with TS, a
210 few chromosomes are randomly picked up from the population to compete with each other, and the winner who has the best
211 fitness among the competitors is then selected and added into the mating pool for recombination. The procedure of TS is
212 repeatedly performed until the target size of mating pool is reached. This kind of selection mechanism is quite straightfor-
213 ward and has less computational complexity; as a result, some potential chromosomes might be missed and, in turn, it could
214 lead to a suboptimal solution [38]. To tackle with this difficulty, we equip the proposed RGA-RDD with an alternative selec-
215 tion mechanism called Ranking Selection (RS) [17]. The RS, which reaps the benefits of both selection schemes mentioned
216 above, is directly operated on the basis of the individual fitness. More precisely, the RS first ranks the chromosomes in
217 the population according to their fitness and, based on the fitness rankings, the RS then discards the amounts of bpNc chro-
218 mosomes that have relatively lower rankings (i.e., worse fitness), while at the same time reproducing the same amounts of
219 chromosomes that are with higher rankings. Note that p is a proportional parameter chosen from the interval (0, 0.5], which
220 is concluded from our previous numerical experience [17]. Since the RS is simply operated according to fitness rankings, it is
221 computationally more efficient than TS and RWS. Besides, using RS, the total number of chromosomes in the population is
222 kept constant because the same amount of chromosomes is discarded and reproduced at a time. In addition, because those
223 potential solutions in the population are preserved, the RS effectively makes the whole population to evolve toward a situ-
224 ation that has a better average fitness. In Fig. 3, we demonstrate the function of the RS operator performing on a population
225 of 10 chromosomes with respect to different proportional parameter p.
226 2.3. The DBX: a Direction-Based Crossover operator
227 As mentioned previously, the crossover operation, which acts as a mechanism to blend the parents’ information for gen-
228 erating potential offspring chromosomes, is widely recognized as the most important and effective searching scheme. Dif-
229 ferent from the commonly used line connection or metric-based crossover operators, the proposed DBX is simply
230 functioned based on the fitness rankings sorted by the RS. After the RS is performed, we sorted the chromosomes and
231 denoted them with the ordered sequence of H ¼ fh1 ; h2 ; . . . ; hN g, in which the chromosomes satisfy the relation of
232 h1 h2 hN . By the notation A B, we mean that the fitness of chromosome A is better or equal to that of chromosome
233 B. Without loss of generality, let the population number N be even and M ¼ N=2. Then, according to H, we divide the N chro-
234 mosomes into two groups as follows:
235 Group A(the leading group):
236

H A ¼ h1 ; h2 ; . . . ; hN=2

238 h1A ; h2A ; . . . ; hMA
Fig. 3. An example to illustrate the operation of the RS operator with p = 0.3 and p = 0.5.
14 February 2015
239 Group B(the group with lower rankings):

240

HB ¼ hN=2þ1 ; hN=2þ2 ; . . . ; hN

242 hB1 ; hB2 ; . . . ; hBM :
243 Note that the set H A is the leading group that has a relatively better fitness, whereas the set HB is the group formed by those
244 worse chromosomes. Let ri be a number randomly chosen from the interval (0, 1]. If ri > k, representing the crossover prob-
245 ability is greater than the specified probability threshold k, we perform the DBX operation to generate offspring chromo-
246 somes by applying the following rule:
247
hA ~i ;
hiA þ sc;i D i ¼ 1; 2; . . . ; M ð2aÞ
i
249 hB B
h þ sc;i D~i ; i ¼ 1; 2; . . . ; M ð2bÞ
i i
250 where sc;i is the crossover step-size given by

251
A
f h f hB
i i
sc;i ¼ ð3Þ
253 max ff ðHÞg min ff ðHÞg
254 ~i ¼ ½ Di;1
and D Di;2 Di;n is the crossover direction vector with its component being generated by
255 (
0 if r j < 0:5
Di;j ¼ ð4Þ
257 hi;jA hBi;j if r j P 0:5
258 ~ is dynamically adjusted by

for i ¼ 1; 2; . . . ; M and j ¼ 1; 2; . . . ; n. As shown in Eq. (3), the step-length of the direction vector D
259 the parameter sc;i according to the related information of the chosen parents. Unlike the conventional crossover operators
260 that use a fixed value or a random number as the step size, we dynamically update sc;i ; 0 6 sc;i 6 1, based on the measure
261 defined in Eq. (3), which is calculated from the ratio of the fitness difference of the chosen paired-parents and the maximum
262 fitness difference extracted from the current population. Besides, we operate the crossover directions generated by DBX
263 according to the difference information produced from each paired gene. In light of Eq. (4), the DBX produces a vector com-
264 ponent by using the information of the j-the paired genes if a random number rj is greater or equal to 0.5, otherwise DBX
265 allows no crossover direction being made from the selected genes. To avoid generating a null direction vector, we randomly
266 pick a non-zero crossover direction Di; j ¼ h A hB –0 from D ~i to execute the crossover operation if rj < 0:5; 8j 2 f1; 2; . . . ; ng.
i; j i; j
267 Such an effort enables the DBX operator to provide 2n-1 searching directions (scenarios) for exploring potential offspring
268 chromosomes. Since the DBX is equipped with a mechanism to adaptively regulate the step size of the crossover for explor-
269 ing better solutions, it is quite different from the tree-structured, directed crossover [51] and the operator developed based
270 on the reflection mechanism of the Nelder–Mead’s simplex method [50]. Besides, we note that the DBX is an advanced ver-
271 sion of our previous work [17] since the new DBX applies a more effective crossover direction generator, Eq. (4), and addi-
272 tionally adds a scheme to preclude the derivation of a null crossover direction.
273 In Fig. 4, we schematically demonstrate the DBX operation working on a two-dimensional (n = 2) case for a given paired
274 parents. In this figure, it is shown obviously that each chromosome is tentatively exploring along each of the three (22 1)
275 possible crossover directions derived by the DBX operator to generate better offspring. Besides, since the initial population is
276 rather diverse in the search space, according to Eq. (3), the DBX operator will make an aggressive move along the direction
277 ~ at the very beginning evolutionary stage. As the population has been gradually converging to a small region after
vector D
278 some successful generations, the DBX turns out to induce an insignificant crossover since the created direction vector is with
279 a rather small magnitude. We note that a specific situation of zero crossover can happen if the step size sc becomes zero, i.e.,

280 the paired parents’ chromosomes have the same properties, namely, hiA ¼ hBi and/or f hiA ¼ f hBi . To divert the situation and
281 to strengthen the population diversity, we introduce DRM to replace DBX through making use of the specifically designed
282 parallel-structured coordinator. On the other hand, to ensure the satisfaction of the required lower and upper bounds, we
283 adjust the genes by fixing their values to their individually extreme bounds if they are guided by DBX to move across the
284 boundary of the feasible region. This effort greatly enhances the probability of locating the optimal solution that sits on
285 the boundary, and thus prevents the drawbacks of conventional crossover operators. To conclude, we summarize the distinct
286 features of the proposed DBX operator as follows:
287 (i) According to fitness rankings, the DBX systematically divides the population into two groups and, based on these two
288 purposely divided groups, M pairs are formed and each pair of parents is guided to generate two candidate offspring.
289 (ii) The DBX uses the relative fitness information of the paired parents to adaptively control the crossover’s step size and
290 conducts 2n 1 possible searching directions (scenarios) to effectively explore potential offspring for fitness
291 improvement.
292 (iii) The function of the DBX is enhanced with several simple yet useful strategies; thus, the DBX is capable of producing
293 non-null and effective crossover directions and greatly increasing the possibility of locating the global optimum on the
294 boundary.
14 February 2015
Fig. 4. A schematic diagram to illustrate the operation of the DBX in a two-dimensional space.
295
296 2.4. The DRM: A Dynamic Random Mutation operator
297 Since mutation serves as a process to change the gene of a chromosome randomly, it is able to prevent the population
298 from premature convergence to a suboptimal solution. However, owing to the mutation contributes considerably less per-
299 formance improvement than other evolutionary operations, only a few efforts have been devoted to design real-value muta-
300 tion operators. To the authors’ limited knowledge, these exceptions are the random mutation (RM) [61], non-uniform
301 mutation (NUM) [61], polynomial mutation (PM) [22], adaptive mutation (AM) [81], and the directed mutation operators
302 (DMO) [47,50]. The mutation scheme of RM simply creates a new solution in the search space by means of a random number.
303 Along with the increment of the generation number, the NUM gradually increases the probability to create a solution that is
304 located nearest to the parent. Similar to NUM, the PM uses a polynomial function to control the probability distribution of
305 mutation. However, the PM uses a fixed probability instead of a dynamic one. Alternatively, the AM recommends the use of
306 adaptive probabilities of mutation, in association with an adaptive crossover, to maintain diversity in the population. The
307 DMO of Korejo et al. [47] applied the statistical information regarding the fitness and distribution of individuals in the cur-
308 rent population to guide the mutation. By adopting the expansion mechanism of NMSM, Kuo and Lin [50] developed a special
309 DMO by which two individuals are randomly selected from the population to generate a new individual in the direction
310 formed from the worse individual to the better individual. In this paper, we propose an alternative mutation scheme, named
311 the Dynamic Random Mutation (DRM) operator, to enhance the population diversity. The DRM applies the mutation rule of
312

hA
i hiA þ sm U0 hU hL ð5aÞ

314 hB
i hBi þ sm U0 hU hL ð5bÞ
n
315 where U0 is a random perturbation vector in the n-dimensional cube ½/0 ; þ/0 of which /0 is a user-specified number cho-
316 sen within the interval (0, 1]. To operate the DRM, we dynamically adjust the mutation step size sm by using the following
317 update rule [17]:
318
b
320 sm ¼ ð1 k=kmax Þ ð6Þ
321 where the parameter bð> 0Þ is used to control the decay rate of sm , and k and kmax denote, respectively, the current generation
322 number and the maximum number of generations.
323 Fig. 5 schematically depicts the progressive, feasible mutation regions that are dynamically produced by the DRM oper-
324 ator with respect to different decay rates. This characteristic plots show clearly that the decay rate b controls the shape of the
325 allowable mutation region, and that mutation range decreases as the number of generation k increases. The idea of using a
326 dynamic mutation size is adopted from simulated annealing [45,64], which describes how a thermodynamic process is
14 February 2015
327 stabilized from a chaotic condition to a stable state by decreasing the system temperature. In a similar concept, the function
328 of sm in the DRM mutation is analogous to the temperature decreasing procedure in a stabilizing thermodynamic process. In
329 view of Eq. (6), it is clear that the DRM operator offers a greater chance of population variations by producing a relatively
330 large allowable step size for the mutation at the very beginning evolution period. As an outcome of this greater variability,
331 the searching process provides a higher probability to escape from local traps. When the population has been gradually con-
332 verging to an optimal solution, a small feasible mutation region produced by the DRM operator would consequently enhance
333 the precision of the obtained solution. Notice that the DRM is operated if ri 6 k or the paired parents have the same prop-
334 erties. Like that mentioned in the DBX operation, those genes mutated outside Xh will also be individually fixed at their
335 extreme bounds to meet the constraint requirements.
336 2.5. The replacement operation
337 Since the population size is kept constant, the survivor selection of both parent and offspring population plays an impor-
338 tant role in preserving the current best found solution for subsequent evolution. We note that the generational replacement
339 (GR) and the steady-state replacement (SSR) are the two commonly used strategies to perform the survivor selection [54].
340 Technically, the GR systematically replaces the parent chromosomes with the offspring, whereas the SSR only replaces one
341 parent chromosome with the best offspring chromosome. In this paper, we implemented the GR with an elitism strategy by
342 which at each generation the offspring chromosome is individually compared with its parent chromosome (i.e., hA i is com-
343 pared with hiA , and hB B
i with hi ), and from each pair of competitors we select the better chromosome to be the survivor in the
344 population.
345 2.6. A simple scheme for alleviating stagnation
346 It is well known that premature convergence and stagnation may occur if the convergence speed of an evolutionary algo-
347 rithm runs too fast and/or the solution has been trapped by a local optimum. To prevent possible occurrence of the prema-
348 ture stagnation and make the RGA-RDD most effective in finding the true optimum, we refresh all chromosomes except the
349 current best one by regenerating a new set of chromosomes when a defined stagnation index (SI) is measured to be less than
350 a pre-specified criterion. In this paper, we use the standard deviation of the population fitness stdevðf ðHÞÞ as the SI to indi-
351 cate the condition of the stagnation. We note that a large value of SI means the population is in a state of high diversity,
352 whereas a small SI value indicates a situation that a premature stagnation might exist. As shown in Fig. 2, we activate
353 the above-mentioned stagnation prevention scheme if stdevð f ðHÞÞ 6 e, where e is a very small positive number. Although
354 other criteria for predicting the premature stagnation are available, our experiences reveal that using the suggested SI cri-
355 terion is a simple yet effective strategy to prevent prematurity of the population during evolution.
Fig. 5. Dynamic mutation regions provided by the DRM with respect to different decay rates.
14 February 2015
356 2.7. The stop criteria
357 We note again that if the evolved chromosome is outside XH , the original chromosome is retained. The proposed RGA-
358 RDD uses the following termination criteria: (1) the number of generations or the number of the function evaluations attains
359 the user-specified maximum number; (2) the family of the chromosomes converges to a final one that has the best value of
360 the objective function.
361 2.8. A summary of the proposed RGA-RDD
362 Up to this stage, we have described the operational principle and the underlying ideas of the proposed RGA-RDD’s oper-
363 ating configuration and the evolutionary operators. In essence, the operation of the proposed RGA-RDD comprises five major
364 functions: (i) the Ranking Selection (RS), (ii) the Direction-Based Crossover operation (DBX), (iii) the Dynamic Random Muta-
365 tion (DRM), (iv) the GR with an elitism strategy, and (v) the scheme of stagnation prevention. To sum up, we illustrate in
366 Fig. 6 the whole evolution algorithm of the proposed parallel-structured RGA-RDD in a pseudo-code. Here, it should be men-
367 tioned again that, in addition to some modifications made to the related operators, the RGA-RDD advances our previous work
368 [17] by adding several useful and new features to the algorithm, for example the scheme of alleviating stagnation, the GR
369 with an elitism strategy for survivor selection, and a specially designed parallel-structured coordinator that is used to orga-
370 nize the operations of the DBX and DRM for improving the effectiveness of locating a global optimum.
371 3. Basic study of the RGA-RDD and an experimental design for parameter tuning
372 The purposes of this section are three-folds. The first is to demonstrate the fundamental properties of the proposed RGA-
373 RDD through two classical benchmark functions. To provide a parameter selection guideline for the RGA-RDD, the second
374 task devotes to perform a parametric analysis to investigate how the parameters of the algorithm affect the solution perfor-
375 mance. In addition, the interaction of algorithm parameters will be explored using the technique of DOE. The third goal is to
376 complete a data-driven experimental design in order to determine an optimal set of RGA-RDD parameters for the benchmark
377 functions. We describe the related efforts and achievements of these three goals in the following individual subsections.
378 3.1. The designate experiments and the benchmark functions
379 The details of these two benchmark functions are shown in Appendix A. We note that both benchmarking functions, the
380 unimodal Sphere function and the multimodal Ackley function, have no specific value of their optimum since they are ran-
381 domly shifted in the search space. To complete the basic studies and the parametric analysis, we designated nine sets of
382 experiments for these two benchmark functions.
383 Table 1 lists the parameter settings for these designate experiments, denoted by Test-1 Test-9, and that will be per-
384 formed on both test functions to investigate the effects of parameters on algorithm performance. In this series of test runs,
385 we set the population size as N = 100, and the maximum numbers of generations for the unimodal Sphere function and the
386 multimodal Ackley function are assigned as kmax ¼ 1000 and kmax ¼ 10; 000, respectively. Besides, we executed each test run
387 100 times to gather the statistical performance data, and the initial population in each individual run was uniformly and
388 randomly initialized within the search space. Furthermore, to provide a fair basis of performance comparison, we recorded
389 the mean, best, median, and the worst solutions of 100 optimized objective function values in each test run.
390 3.2. The effectiveness of the operators and the parallel-structured coordinator
391 We note that, among the nine set of test experiments, the first four tests (Test-1 4) were specially designed to evaluate
392 the effectiveness and applicability of each operator and the parallel-structured coordinator used in the RGA-RDD. To achieve
393 this goal, we organized these four experiments as follows. The designated Test-1 intends to assess the performance of the
394 DBX operator. Thus, in Test-1 we applied only the DBX operator in the RGA-RDD, i.e., the RS and DRM operators were both
395 deactivated. In Test-2, we deactivated the RS and the DBX operators and solely implemented the DRM operator in the algo-
396 rithm to examine its performance. With the RS operator being disabled, the Test-3 activated both the DBX and DRM operators
397 in the algorithm in order to evaluate their cooperated performance. In Test-4, we normally operated the parallel-structured
398 RGA-RDD by activating all the three operators. The experimental results of these four test runs are listed in Table 2. By com-
399 paring the results of Test-1 with that of the Test-2, we found that the DBX operator provides a faster convergence speed than
400 the DRM operator does as they were individually used in the unimodal Sphere function problem. On the contrary, when deal-
401 ing with the multimodal Ackley function problem, the DRM operator turns out to present a superior performance than the
402 DBX operator, revealing that the DRM operator is more effective for not being trapped by a local optimum because of its spe-
403 cial feature of dynamic mutation. We also observed that the Test-3, which coordinated the use of the DBX and DRM opera-
404 tors, presents a better convergence performance than that of both Test-1 and Test-2 for the unimodal Sphere problem. As to
405 the multimodal Ackley problem, the best solution of Test-3 presents a much smaller objective function value than those solu-
406 tions obtained in either Test-1 or Test-2. The above observations obviously reveal that the coordinator in the inner parallel
14 February 2015
Fig. 6. The pseudo-code of the proposed RGA-RDD optimization scheme.
14 February 2015
Table 1
Parameter setting of designate tests.
Set no. Parameter setting

p k b /0 e
Test-1 0 0 – – 1010
Test-2 0 1 4 0.5 1010
Test-3 0 0.1 4 0.5 1010
Test-4 0.01 0.1 4 0.5 1010
Test-5(a) (d) 0.03 0.1 0.1 4 0.5 1010
Test-6(a) (b) 0.01 0.3 0.5 4 0.5 1010
Test-7(a) (c) 0.01 0.1 0.5 2 0.5 1010
Test-8(a) (b) 0.01 0.1 4 0.25 1 1010
Test-9(a) (b) 0.01 0.1 4 0.5 1012 108
407 loop can effectively manage the operations of the DBX and DRM so that both the solution efficiency and the accuracy of the
408 obtained solution were significantly increased.
409 Furthermore, from the results of Test-4 shown in Table 2, it is clearly observed that the normally operated RGA-RDD pro-
410 vides an improved performance in both benchmark problems, apparently better than that of solely using the individual oper-
411 ator. Also observed is that the RS is very effective to assist the RGA-RDD to boost up the convergence speed, even though a
412 very small p value, say p = 0.01, is assigned. Note that, in this case of p = 0.01 and N = 100, the RS merely eliminates the worst
413 solution and reproduce only the best solution from the population at each generation. Based on the extensive comparisons of
414 these test runs, we conclude that the integration of RS, DBX, and DRM operators profoundly benefits both the solution effi-
415 ciency and accuracy of the proposed parallel-structured RGA-RDD.
416 3.3. The role of the RS operator
417 In Test-4, we have observed that the RS operator profoundly affects the convergence performance of the proposed RGA-
418 RDD. To further investigate the role of the RS operator, we designated a new test run, Test-5, where the RGA-RDD was nor-
419 mally operated using the three evolutionary operators, while different proportional parameters were assigned to the RS
420 operator.
421 The simulation results are listed in Table 3, showing that, when the proportional parameter value p is set too large, the RS
422 tends to slow down the convergence speed, especially for multimodal problems. A reason for this is that, with a large p value
423 assigned, the RS is prone to reaching a condition that a vast amount of the same chromosomes duplicated from the current
424 best chromosome will dominate the leading group and that will restrict the DBX operator to explore the offspring around
425 just a few top chromosomes, rather than a whole-range exploration. This therefore would reduce the possibility of locating
426 a global optimum and hence slow down the convergence speed. This situation could be even worse as a multimodal problem
427 is being dealt with. On the contrary, as a suitably small p value is assigned to the RS operator, the DBX will be reactivated to
428 perform an effective and wide-range exploring in the feasible region so that the convergence speed can be significantly
429 enhanced. From Table 3, we observe that the best p value could be simply set to 1/N; this makes the RS to merely eliminate
430 the worst solution and to reproduce only the best solution from the population at each generation. After gathering this basic
431 knowledge about the setting of the RS’s proportional parameter value, we further explored how the RS affects the DBX
432 operation.
Table 2
Performance comparison of designate tests-1 4.
Set no. Optimized f(x)

Best Median Worst Mean Std.
Shifted sphere function with n = 30
Test-1 1.102E1 4.071E1 1.567E+0 4.886E1 2.899E1
Test-2 4.885E1 2.971E+0 6.404E+0 2.979E+0 1.006E+0
Test-3 6.771E2 1.444E1 2.976E1 1.433E1 4.649E2
Test-4 3.778E6 3.436E5 1.873E4 4.527E5 3.4506E5
Shifted Ackley function with n = 30
Test-1 2.286E+0 3.007E+0 4.178E+0 3.082E+0 4.823E1
Test-2 4.022E5 5.715E5 8.784E5 6.006E5 1.171E5
Test-3 2.829E9 2.823E+0 3.802E+0 2.717E+0 5.629E1
Test-4 6.943E10 1.392E9 4.345E9 1.527E9 6.681E10
14 February 2015
433 In Fig. 7, we schematically illustrate the convergence stages of DBX influenced by the RS. This illustrative figure demon-
434 strates clearly that the DBX operator is able to conduct a whole range exploration because the search directions generated at
435 the early evolutionary stages, such as stages A and B, are widely scattered around those diversified leading chromosomes hiA .
436 In the subsequent convergence stages, denoted as stages B and C in the figure, the population has been gradually converging
437 to lie within a small region from which some more chromosomes in the leading group H A that have the same properties will
438 be duplicated by the RS operation. In such a situation, the DBX operator turns out to explore some more potential directions
439 around those top chromosomes and that will enhance a great probability in locating a much better and more accurate solu-
440 tion. Once a majority of chromosomes in the family of the chromosomes has evolved to the current best chromosome as indi-
441 cated in the stage D, the scheme of alleviating stagnation stated in Section 2.6 will be activated to refresh all the
442 chromosomes except the best, and that greatly helps increase the population diversity. Since a new set of chromosomes
443 is created by the stagnation prevention scheme, the searching process returns back to stage A so that the DBX operator will
444 again explore intensively on those scattered search directions made by the refreshed and mostly diversified chromosomes in
445 the population. Based on the above discussion and the results depicted in Table 3, we conclude that the combined use of the
446 RS and the DBX can substantially enhance the searching ability to locate a global optimum, and a suitably small p value for RS
447 substantially benefits to provide a good balance between the fast convergence speed and the stable searching performance.
448 3.4. Parametric studies and a parameter selection guideline
449 We note that the proposed RGA-RDD chiefly consists of six parameters: N; p, k, /0 ; b, and it e. Based on the results pre-
450 sented in the previous subsection, we suggest using p = 1/N for the RS to accelerate the convergence speed, where N is a
451 dimension-dependent population size given by N ¼ min ð10 n=2; 100Þ. To investigate the effects of the remaining four
452 parameters on the searching performance and then accomplish a parameter selection guideline for the RGA-RDD, we per-
453 form some more additional, designate experiments on the above-mentioned two benchmark problems. These designate tests
454 are denoted as Test-6 Test-9, and their results are listed in Tables 4–7, respectively. Recall that the threshold parameter k is
455 used to regulate the crossover and mutation probability. To ensure that almost every pair of the selected parents can go
456 through the DBX operation, we recommend the use of a small k value to lower the threshold.
457 As shown in Table 4, the RGA-RDD presented a relatively stable convergence performance on both the test problems as a
458 small value of k=0.1 was set. Regarding to the parameter /0 , we note that it controls the range of feasible mutation region of
459 the DRM operator. From the results of the parametric study listed in Table 5, we found that /0 has insignificant influence on
460 the RGA-RDD’s searching performance. However, to ensure that at least a half searching space in each dimension can be cov-
461 ered by the DRM operation at the initial evolutionary stage, we suggest using a conservative /0 with its value being chosen
462 within the range of 0.25 and 1.
463 The effects of the parameter b on the searching performance are listed in Table 6, indicating that the suitable b value
464 ranges from 0.5 to 4. Finally, we evaluate the effects of the convergence level parameter e on the removal of premature stag-
465 nation. The simulation results listed in Table 7 reveal that the parameter e significantly affects the performance of the RGA-
466 RDD and its suitable value for an effective stagnation prevention may lie within the range of 1012 and 106.
467 Finally, to characterize the interaction of algorithm parameters, we additionally performed a task of design of experi-
468 ments (DOE) on the above-mentioned two benchmark problems. Although many methods, like the orthogonal array [77],
469 and covering array [74,88], are available for performing the DOE, in this paper, we applied the uniform design (UD) method
470 [31] to allocate the design points. Since relatively few experiments are required to achieve the required robust uniform con-
471 dition, the UD has been proven to be an effective method for the DOE. To accomplish the parametric analysis, we conducted a
Table 3
Performance Comparison of Designate Test-5.
Set no. p Optimized f(x)

Test-3 0 6.771E2 1.444E1 2.976E1 1.433E1 4.649E2
Test-4 0.01 3.778E6 3.436E5 1.873E4 4.527E5 3.451E5
Test-5(a) 0.03 6.704E8 6.919E7 1.124E5 1.196E6 1.614E6
Test-5(b) 0.05 3.387E8 6.001E7 1.063E5 1.080E6 1.418E6
Test-5(c) 0.07 3.528E8 7.166E7 1.086E5 1.357E6 1.618E6
Test-5(d) 0.10 3.893E8 1.508E6 1.713E5 2.514E6 2.904E6
Test-3 0 2.829E9 2.823E+0 3.802E+0 2.717E+0 5.629E1
Test-4 0.01 6.943E10 1.392E9 4.345E9 1.527E9 6.681E10
Test-5(a) 0.03 2.338E9 4.235E9 9.342E1 9.342E3 9.342E2
Test-5(b) 0.05 2.525E9 3.096E8 9.313E1 9.322E3 9.313E2
Test-5(c) 0.07 3.339E9 8.696E8 1.155E+0 1.155E2 1.155E1
Test-5(d) 0.10 4.304E9 8.942E8 1.340E+0 5.860E2 2.421E1
14 February 2015
Fig. 7. A schematic diagram to illustrate the effects of the RS on the DBX operation with respect to different evolution stages.
472 DOE using a UD table of U 36 (44), where a total of 36 experiments was arranged with each of the four decision parameters
473 being divided into four levels; that is, k ¼ ½0:1; 0:3666; 0:6333; 0:9; :½½0:25; 0:5; 0:75; 1; ½b ¼ ½0:5; 1; 2; 4; and
474 e ¼ ½1012 ; 1010 ; 108 ; 106 . To assess the performance of each experiment, we calculated the signal-to-noise ratio gdB as
475 follows:
476 !
1X t
gdB ¼ 10log10 Y2 ð7Þ
478 t i¼1 i
14 February 2015
Table 4
Set no. k Optimized f(x)

Test-4 0.1 3.778E6 3.436E5 1.873E4 4.527E5 3.450E5
Test-6(a) 0.3 7.545E5 3.194E4 7.819E4 3.528E4 1.805E4
Test-6(b) 0.5 2.169E4 5.340E4 1.115E3 5.473E4 1.701E4
Test-4 0.1 6.943E10 1.392E9 4.345E9 1.527E9 6.681E10
Test-6(a) 0.3 1.735E9 3.833E9 9.313E1 9.314E3 9.313E2
Test-6(b) 0.5 2.632E9 2.186E6 1.778E+0 2.463E1 5.042E1
Table 5
Performance comparison of designate test-7.
Set no. /0 Optimized f(x)

Test-8(a) 0.25 2.480E6 2.101E5 1.073E4 2.677E5 1.858E5
Test-4 0.5 3.778E6 3.436E5 1.873E4 4.527E5 3.450E5
Test-8(b) 1 7.461E6 4.161E5 2.895E4 5.362E5 4.602E5
Test-8(a) 0.25 3.937E10 1.441E9 3.813E9 1.520E9 6.803E10
Test-4 0.5 6.943E10 1.392E9 4.345E9 1.527E9 6.681E10
Test-8(b) 1 5.665E10 1.393E9 4.041E9 1.539E9 6.243E10
Table 6
Set No. b Optimized f(x)

Test-7(a) 0.5 9.614E6 6.770E5 2.621E3 1.232E4 2.693E4
Test-7(b) 1 5.824E6 8.495E5 9.304E4 1.284E4 1.311E4
Test-7(c) 2 7.850E6 5.432E5 3.548E4 7.636E5 6.453E5
Test-4 4 3.778E6 3.436E5 1.873E4 4.527E5 3.450E5
Test-7(a) 0.5 7.875E10 1.492E9 4.647E9 1.688E9 7.314E10
Test-7(b) 1 5.051E10 1.652E9 4.511E9 1.761E9 7.653E10
Test-7(c) 2 6.064E10 1.536E9 4.116E9 1.652E9 6.799E10
Test-4 4 6.943E10 1.392E9 4.345E9 1.527E9 6.681E10
Table 7
Performance comparison of designate test-9.
Set no. e Optimized f(x)

Test-9(a) 1012 4.916E6 3.482E5 2.125E4 4.709E5 3.897E5
Test-4 1010 3.778E6 3.436E5 1.873E4 4.527E5 3.450E5
Test-9(b) 108 4.279E6 3.879E5 2.770E4 4.999E5 4.306E5
Test-9(a) 1012 1.997E11 4.066E11 9.313E1 9.313E3 9.313E2
Test-4 1010 6.943E10 1.392E9 4.345E9 1.527E9 6.681E10
Test-9(b) 108 8.307E9 2.185E8 9.110E8 2.486E8 1.310E8
14 February 2015
479 where t is the total number of the test runs in each experiment, and Y i is the optimized objective function value in each test
480 run. Note that the signal-to-noise ratio gdB is a commonly used measure (performance index) in association with DOE, and
481 the higher values of gdB , the smaller the value of Y i is obtained. Table 8 lists the 36 combinatorial arrangements and their
482 respective outcomes with t ¼ 100. Besides, we calculated the main effect D [77] of each parameter and list the results in
483 Table 9. Note that, in the design of experiments, the main effect is a factor to indicate the extent of the importance (influence)
484 of an independent variable on a dependent variable averaging across the levels of any other independent variables. From
485 Table 9, it is shown clearly that, among the algorithm parameters, the parameter k is the most significant factor to influence
486 the searching performance of both test functions. Furthermore, we depict in Fig. 8 the response surface plots to demonstrate
487 the interactions of parameters of the proposed algorithm, which provides a comprehensive basis for optimizing the RGA-
488 RDD parameters.
489 3.5. A data-driven experimental design for optimizing the algorithm parameters
490 Having characterized the parameter interactions and a guideline of parameter selection, in this subsection we further
491 applied the technique of data-driven experimental design (DDOS) to search for an optimal set of RGA-RDD parameters,
492 except that the dimension-dependent parameters were given by p = 1/N and N ¼ minð10 n=2; 100Þ. Based on the param-
493 eter selection guideline presented in the previous subsection, we set the feasible region for the RGA-RDD parameters as fol-
n o

494 lows: W ¼ R0:1 6 k 6 0:9; 0:3 6 /0 6 1; 0:5 6 b 6 4; 1012 6 e 6 106 ,where R ½k; /0 ; b; e. Although many advanced
495 methods, for example the binary covering array [74,89], are available in the literature for parameter fine-tuning, in this
496 paper, we applied the technique of DDOS [18] to search for an optimal set of RGA-RDD parameters. We note the DDOS
497 was originally developed to search for an optimal set of operating parameters that maximizes the film growth performance
498 of an MOCVD process. In essence, the DDOS incorporates a UD [31] for design of experiments and a shape-tunable neural
Table 8
The outcomes of a DOE using the UD Table U 36 (44).
Exp. Parameters gdB RgdB

k /0 b e Sphere Ackley
1 0.3666 0.25 0.5 1010 1.5959E+02 1.9289E+02 3.5248E+02
2 0.6333 1 2 108 6.0145E+01 1.6393E+02 2.2408E+02
3 0.9 1 4 106 4.5482E+01 3.0802E+00 4.8562E+01
4 0.3666 1 4 1012 1.5898E+02 2.3480E+02 3.9377E+02
5 0.3666 0.75 2 106 1.0109E+02 1.1745E+02 2.1853E+02
6 0.3666 0.25 2 1010 1.5667E+02 2.7600E+02 4.3267E+02
7 0.6333 0.5 4 108 9.0592E+01 1.6825E+02 2.5884E+02
8 0.6333 0.5 1 108 6.1944E+01 2.6205E+02 3.2399E+02
9 0.3666 0.75 0.5 106 1.0074E+02 2.9261E+01 1.3000E+02
10 0.1 0.25 1 108 1.3965E+02 2.7707E+02 4.1672E+02
11 0.6333 1 0.5 1010 1.8138E+01 1.2068E+01 6.0701E+00
12 0.1 0.5 0.5 108 1.4127E+02 1.4915E+02 2.9042E+02
13 0.3666 1 0.5 108 1.4029E+02 2.6849E+02 4.0878E+02
14 0.9 0.75 0.5 108 4.3166E+01 5.5428E+00 3.7623E+01
15 0.3666 0.5 1 1012 1.5253E+02 1.1756E+02 2.7009E+02
16 0.6333 0.75 2 106 6.0789E+01 1.1435E+01 7.2224E+01
17 0.9 0.5 1 106 2.3559E+01 1.4387E+01 9.1717E+00
18 0.9 0.25 1 1010 1.2233E+01 1.1920E+01 2.4153E+01
19 0.1 0.75 0.5 1012 2.1962E+02 1.7820E+01 2.3744E+02
20 0.1 1 2 1010 1.7784E+02 2.7646E+02 4.5430E+02
21 0.6333 0.25 4 1012 1.0020E+02 2.1336E+02 3.1356E+02
22 0.9 0.5 0.5 1012 3.6132E+01 5.0983E+00 4.1230E+01
23 0.3666 0.25 4 106 1.0374E+02 2.7587E+02 3.7961E+02
24 0.9 1 1 1010 3.1562E+01 3.4161E+00 3.4978E+01
25 0.3666 0.75 1 1012 1.5042E+02 2.7553E+02 4.2595E+02
26 0.9 0.5 4 1010 5.8266E+01 4.4878E+00 5.3778E+01
27 0.9 0.75 2 1012 1.3086E+00 3.8725E+00 2.5638E+00
28 0.1 0.25 2 1012 1.7974E+02 2.7702E+02 4.5676E+02
29 0.1 0.5 4 1010 2.2048E+02 2.7596E+02 4.9644E+02
30 0.6333 0.25 0.5 106 7.2302E+01 1.3077E+02 2.0307E+02
31 0.1 0.75 4 108 1.3856E+02 2.7520E+02 4.1377E+02
32 0.6333 0.75 4 1010 8.6624E+01 1.8914E+02 2.7576E+02
33 0.1 0.5 2 106 1.1702E+02 2.7580E+02 3.9282E+02
34 0.1 1 1 106 1.0311E+02 3.9993E+01 1.4310E+02
35 0.9 0.25 2 108 1.9017E+01 4.1586E+00 2.3175E+01
36 0.6333 1 1 1012 4.6914E+01 1.6949E+02 2.1641E+02
14 February 2015
Table 9
The main effect on response of gdB .
Problem Main effect D on response of gdB
k /0 B e
Sphere 364.9680 (0.62) 49.8364 (0.08) 103.9281 (0.18) 73.2161 (0.12)
Ackley 466.5241 (0.45) 179.4289 (0.17) 213.6025 (0.22) 168.9501 (0.16)
Note: The value in parenthesis denotes the importance of the algorithm parameter that affects the solution effectiveness of the problem.
Fig. 8. Response surface plots of the algorithm parameters.
499 network (NN) model [11] for auxiliary decision support. The step-by-step parameter-tuning procedure using the DDOS is
500 described as follows:
501 Step 0: Set the admissible search space and initialize the parameters for the NN decision support model.
502 Step 1: Apply the UD method to complete the initial design of experiments and calculate its individual performance index.
14 February 2015
503 Step 2: Train the NN decision support model based on the gathered response data.
504 Step 3: Perform an optimizer on the NN model to search for the optimal parameters.
505 Step 4: Execute the RGA-RDD to solve the test problems with the estimated parameters and calculate its performance index.
506 Step 5: Check the convergence of the objective function (performance index). If the stop criterion is satisfied, break the loop.
507 Otherwise, go back to Step 2 and add the current experiment result into the NN model for retraining.
508
509 The above-mentioned DDOS procedure applies the UD method [31] to allocate the orthogonal design points that are uni-
510 formly scattered in the searching domain. As to the decision support model, we utilized a shape-tunable neural network [11].
511 Unlike the conventional NNs that use fixed nonlinear function shape, the shape-tunable NN possesses the ability to adjust
512 the neuron shapes and thus has greater flexibility and higher nonlinearity capacity to relate the training data. Besides, the
513 MATLAB subroutine ‘‘fmincon’’ is applied as an optimizer working on the established NN decision support model to search
514 for an optimal set of the RGA-RDD parameters. For an initial design of experiments, we selected the uniform design table
515 U 9 (94); this UD table indicates that each of the 4 decision parameters has 9 levels and a total of 9 experiments are performed
516 initially. The structure of the NN auxiliary model is chosen as 4-10-1, i.e., the NN model has three layers with, respectively, 4,
517 10, and 1 nodes in the input, hidden, and output layers. We note that the initial NN model parameters were randomly
518 assigned and the admissible searching range of each parameter was normalized to the range of [0, 1]4 when applying the
P
519 DDOS. The objective function used is the sum of the signal-to-noise ratio, i.e., max. gdB . By using the above algorithm
520 setup, the DDOS scheme located the set of optimal parameter values R ¼ ½0:1; 0:413; 4; 1:036 1010 with the maximized
521 objective function value of 521.336. The convergence progress of the DDOS is plotted in Fig. 9, showing that the algorithm
522 successfully converged after 15 iterations (runs). To this end, we conclude that the DDOS is very effective to optimize the
523 RGA-RDD’s parameters, and the RGA-RDD equipped with the set of optimized parameters apparently presents a better per-
524 formance than that with arbitrarily assigned parameters. Having presented the DDOS for optimizing the RGA-RDD’s param-
525 eters, we are ready to assess its performance on a testbed of real-parameter benchmark optimization problems.
526 4. Benchmark study of The RGA-RDD
527 In this section, we aim to quantify the searching performance of the proposed RGA-RDD and compare its performance
528 with those of existing state-of-the-art optimization algorithms. To do so, we selected the testbed named CEC2005, which
529 consists of a set of 25 well-motivated benchmark optimization problems published in [84]. The chosen CEC2005 testbed
530 has been reported to be very suitable for systematic and sound comparisons of real-parameter optimization algorithms
531 because it does not have some common drawbacks of other frequently used benchmark functions [53], such as (1) the global
532 optimal solution lies at the center of the search space or the origin; (2) the global optimum has the same value of all decision
533 variables, and (3) the local optimal solution lies along the coordinate axes or has no linkage among the decision variables.
534 Before we proceed to execute the proposed RGA-RDD on the CEC2005 testbed, the properties of the benchmark functions,
535 evaluation criteria, and the algorithm setup are, respectively, described in the following subsections.
536 4.1. The properties of the benchmark Testbed
537 Recall that the CEC2005 testbed comprises 25 functions. According to function properties, we classified these functions into
538 the following three categories: unimodal functions (F 1 F 5 ), multimodal functions (F 6 F 12 ), and the hybrid composition
539 functions (F 15 F 25 ). Due to limited page space, in Table 10 we briefly summarize the properties of these benchmark
540 functions.
525
520
515
510
505
500
495
0 5 10 15 20
Iterations
Fig. 9. The convergence progress of the data-driven optimization scheme for parameter tuning.
14 February 2015
Table 10
Properties of benchmark functions [71].
No. Function name Properties

F1 Shifted Sphere Shifted; separable
F2 Shifted Schwefel’s Problem 1.2 Shifted; Non-separable
F3 Shifted Rotated High Conditioned Elliptic function Shifted; Rotated; Non-Separable; High conditioning
F4 Sifted Schwefel’s Problem 1.2 with Noise in fitness Shifted; Non-Separable; Noise in fitness
F5 Schwefel’s Problem 2.6 with Global Optimum on Non-Separable; The global optimum lies on the boundary
Bounds
F6 Shifted Rosenbrock’s Function Shifted; Non-separable; Having a very narrow valley from a local optimum to the
global optimum
F7 Shifted Rotated Griewank’s Function without Bounds Shifted; Rotated; Non-separable; No bounds for variables x
F8 Shifted Rotated Ackley’s Function with Global Optimum Shifted; Rotated; Non-separable; The global optimum lies on the boundary
on Bounds
F9 Shifted Rastrigin’s Function Shifted; Separable; Local optima’s number is huge
F 10 Shifted Rotated Rastrigin’s Function Shifted; Rotated; Non-separable; Local optima’s number is huge
F 11 Shifted Rotated Weierstrass Function Shifted; Rotated; Non-separable; Continuous but differentiable only on a set of
points
F 12 Schwefel’s Problem 2.13 Shifted; Non-separable
F 13 Expanded Extended Griewank’s plus Rosenbrock’s Shifted; Non-separable
Function (F8F2)
F 14 Shifted Rotated Expanded Scaffer’s F6 Shifted; Non-separable
F 15 Hybrid Composition Function Separable near the global optimum; A huge number of local optima
F 16 Rotated Hybrid Composition Function Rotated; Non-separable; A huge number of local optima
F 17 Rotated Hybrid Composition Function with Noise in Rotated; Non-separable; A huge number of local optima; With Gaussian noise in
Fitness fitness
F 18 Rotated Hybrid Composition Function Rotated; Non-separable; A huge number of local optima; A local optimum is sat on
the origin
F 19 Rotated Hybrid Composition Function with a Narrow Non-separable; A huge number of local optima; A local optimum is sat on the origin;
Basin for the Global Optimum A narrow basin for the global optimum
F 20 Rotated Hybrid Composition Function with the Global Non-separable; A huge number of local optima; A local optimum is sat on the origin;
Optimum on the Bounds The global optimum lies on the boundary
F 22 Rotated Hybrid Composition Function with High Non-separable; A huge number of local optima; The global optimum lies on the
Condition Number Matrix boundary
F 23 Non-Continuous Rotated Hybrid Composition Function Non-separable; A huge number of local optima; noncontinuous; Global optimum lies
on the boundary
F 25 Rotated Hybrid Composition Function without Bounds Non-separable; A huge number of local optima; The global optimum lies on the
boundary; No bounds for variables x
541 For a detailed description about these benchmark functions, please refer to [84]. We note that all these benchmark func-
542 tions are scalable with the dimension of the decision variables and most functions have no specific value of their optimal
543 solution xopt since the optimal solution is randomly shifted in x-space for each test run. As a consequence, each of the test
544 functions has an artificially chosen optimal function value f opt that is randomly shifted in the f-space. With the above
545 -mentioned properties, we confirm that this chosen testbed does not have the above-mentioned common drawbacks like
546 other testbeds, and thus a fair basis of performance evaluation and comparisons is provided.
547 4.2. The evaluation criteria
548 For benchmarking the performance of algorithms, we set up the condition of the experiments and the evaluation criteria
549 as follows. In each benchmark function, the dimensions of search space considered are 10 and 30, respectively. Besides, the
550 maximum number of function evaluations (MaxFEs) was set to be n 105 , where n is the dimension of the function. To
551 gather the statistical performance data, we executed each test set 25 runs. Note that the proposed RGA-RDD as well as those
552 comparative algorithms is uniformly and randomly initialized within the search space for each benchmark function, except
553 for F 7 and F 25 where the initialization is specially specified. A reason for this is that all other test functions have the global
554 optimum that lies inside the given bounds, but F 7 and F 25 are exceptions because they do not have a pre-defined search
555 domain and their global optimum are located outside the specified initialization region. For performance comparison, in each

556 run, we recorded function error value f ðxÞ f xopt at each generation and the number of function evaluations (FEs)
557 needed to achieve the desired accuracy level of each function [71]. Besides, the performance measures used in this paper
558 for benchmarking are the success rate (SR) and success performance (SP), which are, respectively, defined as:
559
561 SR ¼ ð#RT Þ=ðtotal runsÞ ð8Þ
562 and
563
565 SP ¼ RTs ðtotal runsÞ=ð#RT Þ: ð9Þ
14 February 2015
566 In the above quantities, #RT is the number of successful runs that achieved the desired accuracy level specified in [46], and
567 RTs is the average FEs for successful runs.
568 4.3. Algorithms setup
569 Based on the parameter selection guideline presented in Section 3.4 and the use of the DDOS on the CEC2005 testbed, we

570 set up the optimized RGA-RDD parameters as follows: k ¼ 0:126; /0 ¼ 0:503; b ¼ 4:371; e ¼ 0:983 1010 , p = 1/N,
571 and N ¼ minð10 n=2; 100Þ. Note that this set of parameters was used to solve each benchmark function of the CEC2005
572 testbed and we did not readjust them even though the function to be solved is with different dimensions. Moreover, as a
573 basis of rigorous performance comparison, we selected the following ten state-of-the-art algorithms: BLX-GL50 [32], BLX-
574 MA [62], CoEVO [70], DE [76], DMS-L-PSO [52], EDA [97], G-CMA-ES [5], K-PCX [80], L-SaDE [73], and SPC-PNX [6]. These
575 algorithms were chosen based on the following criteria: (1) these algorithms had been performed on the same testbed
576 and they all presented an outstanding performance in the competition of CEC2005; and (2) their optimization schemes
577 are similar to or in the same category of the proposed RGA-RDD. According to the operational functions and evolution mech-
578 anisms used, we further categorized these comparative algorithms into the following five groups: (i) ES strategy: CoEVO, G-
579 CMA-ES, and K-PCX; (ii) RCGA algorithm: BLX-GL50, BLX-MA, and SPC-PNX; (iii) Swarm intelligence: DMS-L-PSO; (iv) DE
580 algorithms: DE, and L-SaDE; and (v) Estimation of distribution algorithm: EDA.
581 4.4. Results and discussion
582 4.4.1. The performance evaluation on CEC2005 testbed

583 The experimental results of the RGA-RDD performed on the CEC2005 testbed are listed in Tables 11–16, therein we report
584 the error values of each function after 103; 104; 105; 106 FEs in the case of 10-dimension, and 3 103; 3 104; 3 105;
585 3 106 FEs in 30-dimensional case. Moreover, the errors of each function were sorted in an ascending order, and the 1st
586 (best), 7th, 13th (median), 19th, 25th (worst), mean, and the standard deviation for the 25 runs are also presented in the
587 respective tables.
588 From the results of the experiments, we observed that the proposed RGA-RDD successfully and easily solved the test func-
589 tions F 1 and F 2 to meet any required level of solution accuracy. As for F 3 , the RGA-RDD resulted in a shrunk convergence
590 speed because the function is highly ill-conditioned. On F 4 , we found that the additional random noise that was added to
591 the function value did not hinder the RGA-RDD from locating the true optimum, as indicated by the obtained results that
Table 11
Error values of functions F 1 F 8 in 10-dimension.
FEs F1 F2 F3 F4 F5 F6 F7 F8
1E3 Best 4.870E+02 1.668E+03 2.127E+07 3.470E+03 1.277E+03 1.755E+07 1.724E+01 2.038E+01
7th 1.235E+03 4.950E+03 3.705E+07 1.067E+04 3.518E+03 3.447E+07 4.991E+01 2.065E+01
Median 1.577E+03 6.718E+03 4.097E+07 1.281E+04 4.401E+03 4.391E+07 6.899E+01 2.073E+01
19th 2.211E+03 8.065E+03 6.300E+07 1.544E+04 6.101E+03 8.215E+07 8.641E+01 2.078E+01
Worst 3.310E+03 1.425E+04 1.561E+08 1.975E+04 7.541E+03 1.894E+08 1.335E+02 2.091E+01
Mean 1.728E+03 6.739E+03 5.460E+07 1.256E+04 4.681E+03 6.086E+07 7.270E+01 2.071E+01
Std 6.968E+02 2.995E+03 3.580E+07 4.629E+03 1.634E+03 3.996E+07 3.034E+01 1.259E01
1E4 Best 1.765E09 6.291E02 1.530E+05 3.387E+01 1.059E09 2.536E01 6.176E02 2.021E+01
7th 5.206E07 4.831E01 2.675E+05 7.800E+01 6.065E08 4.830E+00 2.341E01 2.038E+01
Median 9.432E07 1.583E+00 4.562E+05 1.240E+02 1.783E07 1.148E+01 4.521E01 2.049E+01
19th 1.661E06 4.317E+00 5.784E+05 2.215E+02 1.410E06 6.965E+01 7.598E01 2.054E+01
Worst 5.586E05 3.350E+01 1.374E+06 5.952E+02 1.735E03 2.215E+02 1.023E+00 2.065E+01
Mean 4.445E06 4.585E+00 4.881E+05 1.780E+02 7.829E05 4.822E+01 4.954E01 2.046E+01
Std 1.232E05 7.320E+00 2.898E+05 1.474E+02 3.460E04 6.703E+01 3.197E01 1.168E01
1E5 Best 0.000E+00 5.684E14 2.703E+03 5.684E14 0.000E+00 6.063E07 4.673E02 2.011E+01
7th 0.000E+00 5.684E14 7.668E+03 1.137E13 0.000E+00 4.272E03 2.336E01 2.028E+01
Median 5.684E14 1.137E13 2.277E+04 5.116E13 0.000E+00 5.080E02 4.163E01 2.034E+01
19th 5.684E14 1.137E13 5.084E+04 1.359E11 0.000E+00 3.987E+00 7.263E01 2.038E+01
Worst 5.684E14 3.411E13 1.263E+05 9.721E09 0.000E+00 4.027E+00 9.943E01 2.050E+01
Mean 4.093E14 1.046E13 3.498E+04 4.024E10 0.000E+00 1.265E+00 4.651E01 2.032E+01
Std 2.605E14 6.702E14 3.372E+04 1.942E09 0.000E+00 1.774E+00 2.918E01 8.113E02
1E6 Best 0.000E+00 0.000E+00 2.198E+01 0.000E+00 0.000E+00 1.359E11 4.673E02 2.000E+01
7th 0.000E+00 0.000E+00 3.543E+02 0.000E+00 0.000E+00 1.125E10 2.336E01 2.000E+01
Median 0.000E+00 0.000E+00 9.645E+02 0.000E+00 0.000E+00 1.678E10 4.163E01 2.001E+01
19th 0.000E+00 0.000E+00 1.951E+03 0.000E+00 0.000E+00 3.169E10 7.263E01 2.002E+01
Worst 0.000E+00 0.000E+00 1.399E+04 5.684E14 0.000E+00 1.039E09 9.943E01 2.018E+01
Mean 0.000E+00 0.000E+00 1.750E+03 1.137E14 0.000E+00 2.700E10 4.651E01 2.003E+01
Std 0.000E+00 0.000E+00 2.835E+03 2.321E14 0.000E+00 2.613E10 2.918E01 4.834E02
14 February 2015
Table 12
Error values of functions F 9 —F 17 in 10-dimension.
FEs F9 F 10 F 11 F 12 F 13 F 14 F 15 F 16 F 17
1E3 Best 3.855E+01 5.356E+01 9.495E+00 6.673E+03 2.443E+00 4.098E+00 4.259E+02 3.129E+02 3.371E+02
7th 4.968E+01 7.087E+01 1.092E+01 1.585E+04 4.665E+00 4.320E+00 7.016E+02 3.791E+02 4.213E+02
Median 5.424E+01 7.919E+01 1.132E+01 2.205E+04 5.996E+00 4.413E+00 7.499E+02 4.174E+02 4.641E+02
19th 6.115E+01 8.277E+01 1.184E+01 3.051E+04 6.869E+00 4.484E+00 7.750E+02 4.613E+02 5.208E+02
Worst 6.841E+01 1.032E+02 1.310E+01 4.572E+04 1.096E+01 4.531E+00 9.260E+02 5.487E+02 5.603E+02
Mean 5.445E+01 7.796E+01 1.128E+01 2.381E+04 6.003E+00 4.385E+00 7.177E+02 4.171E+02 4.637E+02
Std 7.979E+00 1.126E+01 9.658E01 9.741E+03 1.788E+00 1.116E01 1.273E+02 6.270E+01 6.050E+01
1E4 Best 1.489E02 5.970E+00 5.095E+00 4.080E+00 4.839E01 3.094E+00 1.430E+02 1.289E+02 1.545E+02
7th 2.985E+00 1.393E+01 6.157E+00 3.057E+01 6.296E01 3.527E+00 1.986E+02 1.457E+02 1.722E+02
Median 4.975E+00 1.990E+01 7.469E+00 1.111E+02 8.035E01 3.646E+00 2.165E+02 1.504E+02 1.822E+02
19th 7.960E+00 2.189E+01 8.088E+00 9.263E+02 1.014E+00 3.760E+00 2.698E+02 1.640E+02 1.955E+02
Worst 1.393E+01 3.681E+01 9.487E+00 7.174E+03 1.302E+00 4.047E+00 3.502E+02 1.809E+02 2.146E+02
Mean 5.573E+00 1.949E+01 7.165E+00 9.205E+02 8.371E01 3.635E+00 2.337E+02 1.529E+02 1.824E+02
Std 3.287E+00 7.993E+00 1.233E+00 1.760E+03 2.422E01 1.957E01 5.720E+01 1.377E+01 1.696E+01
1E5 Best 5.684E14 5.970E+00 2.329E+00 1.447E04 3.589E01 2.620E+00 6.644E01 8.859E+01 9.531E+01
7th 1.137E13 1.393E+01 3.236E+00 9.991E03 4.714E01 3.103E+00 4.166E+01 1.023E+02 1.144E+02
Median 1.705E13 1.990E+01 4.493E+00 1.005E+01 5.398E01 3.153E+00 6.407E+01 1.097E+02 1.239E+02
19th 3.411E13 2.189E+01 5.188E+00 4.225E+02 7.407E01 3.447E+00 8.039E+01 1.207E+02 1.314E+02
Worst 7.390E13 3.283E+01 7.018E+00 2.463E+03 1.221E+00 3.689E+00 1.085E+02 1.374E+02 1.552E+02
Mean 2.478E13 1.859E+01 4.471E+00 3.898E+02 6.374E01 3.234E+00 5.760E+01 1.113E+02 1.227E+02
Std 1.961E13 6.776E+00 1.435E+00 7.258E+02 2.300E01 2.704E01 3.457E+01 1.369E+01 1.503E+01
1E6 Best 0.000E+00 5.970E+00 1.381E+00 2.640E09 1.874E01 2.107E+00 0.000E+00 5.963E+01 9.450E+01
7th 5.684E14 1.194E+01 2.746E+00 1.273E08 2.828E01 2.915E+00 0.000E+00 1.015E+02 1.121E+02
Median 5.684E14 1.691E+01 3.219E+00 8.164E04 3.232E01 3.054E+00 0.000E+00 1.097E+02 1.226E+02
19th 5.684E14 2.089E+01 4.223E+00 1.884E+01 3.568E01 3.180E+00 0.000E+00 1.152E+02 1.297E+02
Worst 5.684E14 2.786E+01 5.562E+00 1.694E+03 6.318E01 3.523E+00 1.421E14 1.374E+02 1.546E+02
Mean 4.548E14 1.609E+01 3.500E+00 1.291E+02 3.419E01 3.003E+00 1.137E15 1.080E+02 1.212E+02
Std 2.321E14 5.886E+00 1.056E+00 3.802E+02 9.993E02 3.492E01 3.935E15 1.570E+01 1.489E+01
Table 13
Error values of functions F 18 —F 25 in 10-Dimension.
FEs F 18 F 19 F 20 F 21 F 22 F 23 F 24 F 25
1E3 Best 8.705E+02 8.853E+02 9.504E+02 1.332E+03 9.909E+02 9.906E+02 1.044E+03 1.030E+03
7th 9.953E+02 1.047E+03 1.046E+03 1.360E+03 1.053E+03 1.222E+03 1.123E+03 1.071E+03
Median 1.064E+03 1.063E+03 1.064E+03 1.399E+03 1.132E+03 1.298E+03 1.254E+03 1.143E+03
19th 1.087E+03 1.106E+03 1.090E+03 1.418E+03 1.183E+03 1.324E+03 1.332E+03 1.307E+03
Worst 1.123E+03 1.149E+03 1.103E+03 1.464E+03 1.374E+03 1.360E+03 1.448E+03 1.430E+03
Mean 1.040E+03 1.063E+03 1.061E+03 1.392E+03 1.123E+03 1.260E+03 1.235E+03 1.190E+03
Std 6.797E+01 6.140E+01 3.683E+01 3.774E+01 8.947E+01 9.217E+01 1.253E+02 1.338E+02
1E4 Best 6.309E+02 3.539E+02 5.139E+02 3.658E+02 3.188E+02 5.595E+02 2.019E+02 2.009E+02
7th 7.080E+02 7.015E+02 7.340E+02 5.027E+02 7.783E+02 5.595E+02 2.066E+02 2.030E+02
Median 7.430E+02 8.002E+02 8.298E+02 5.170E+02 7.873E+02 1.078E+03 2.159E+02 2.070E+02
19th 9.961E+02 9.247E+02 9.945E+02 8.195E+02 7.966E+02 1.171E+03 2.268E+02 2.140E+02
Worst 1.038E+03 1.030E+03 1.042E+03 1.156E+03 8.727E+02 1.232E+03 5.119E+02 3.360E+02
Mean 8.134E+02 8.117E+02 8.383E+02 6.633E+02 7.230E+02 9.078E+02 2.364E+02 2.188E+02
Std 1.430E+02 1.573E+02 1.619E+02 2.442E+02 1.674E+02 2.967E+02 6.689E+01 3.132E+01
1E5 Best 3.000E+02 3.539E+02 4.386E+02 3.000E+02 3.000E+02 5.595E+02 2.000E+02 2.000E+02
7th 5.228E+02 6.791E+02 6.172E+02 3.000E+02 7.445E+02 5.595E+02 2.000E+02 2.000E+02
Median 6.681E+02 7.344E+02 6.861E+02 5.000E+02 7.604E+02 5.595E+02 2.000E+02 2.000E+02
19th 7.113E+02 8.001E+02 8.000E+02 5.000E+02 7.640E+02 1.078E+03 2.000E+02 2.000E+02
Worst 9.961E+02 9.504E+02 9.667E+02 1.144E+03 8.375E+02 1.135E+03 5.000E+02 2.000E+02
Mean 6.366E+02 7.195E+02 6.908E+02 5.268E+02 6.914E+02 7.546E+02 2.120E+02 2.000E+02
Std 1.632E+02 1.433E+02 1.501E+02 2.603E+02 1.756E+02 2.414E+02 6.000E+01 2.161E11
1E6 Best 3.000E+02 3.000E+02 3.000E+02 3.000E+02 3.000E+02 5.595E+02 2.000E+02 2.000E+02
7th 3.000E+02 3.000E+02 3.000E+02 3.000E+02 3.000E+02 5.595E+02 2.000E+02 2.000E+02
Median 3.557E+02 3.716E+02 3.579E+02 5.000E+02 3.000E+02 5.595E+02 2.000E+02 2.000E+02
19th 4.069E+02 4.581E+02 4.533E+02 5.000E+02 7.457E+02 7.212E+02 2.000E+02 2.000E+02
Worst 8.000E+02 8.000E+02 8.000E+02 9.665E+02 8.118E+02 1.101E+03 5.000E+02 2.000E+02
Mean 4.161E+02 4.626E+02 4.325E+02 4.609E+02 4.484E+02 6.622E+02 2.120E+02 2.000E+02
Std 1.680E+02 1.987E+02 1.794E+02 1.943E+02 2.213E+02 1.653E+02 6.000E+01 1.610E12
14 February 2015
Table 14
FEs F1 F2 F3 F4 F5 F6 F7 F8
1E3 Best 3.510E+03 1.776E+04 1.866E+08 3.250E+04 1.580E+04 1.832E+09 4.761E+02 2.098E+01
7th 5.183E+03 2.861E+04 3.030E+08 4.448E+04 1.912E+04 3.462E+09 7.864E+02 2.114E+01
Median 5.669E+03 3.078E+04 3.867E+08 4.801E+04 2.041E+04 4.283E+09 8.670E+02 2.116E+01
19th 6.244E+03 3.327E+04 4.567E+08 5.627E+04 2.121E+04 5.098E+09 9.490E+02 2.121E+01
Worst 1.047E+04 4.654E+04 6.415E+08 7.123E+04 2.532E+04 7.128E+09 1.215E+03 2.127E+01
Mean 6.091E+03 3.068E+04 3.752E+08 4.986E+04 2.018E+04 4.272E+09 8.766E+02 2.116E+01
Std 1.662E+03 6.304E+03 1.137E+08 9.345E+03 2.157E+03 1.296E+09 1.691E+02 7.289E02
1E4 Best 1.408E04 8.587E+02 2.435E+06 6.442E+03 3.235E+03 5.968E+03 2.610E+00 2.097E+01
7th 2.466E04 1.893E+03 6.569E+06 1.250E+04 3.962E+03 1.518E+04 5.528E+00 2.102E+01
Median 6.405E04 2.552E+03 9.878E+06 1.313E+04 4.413E+03 2.145E+04 7.399E+00 2.106E+01
19th 1.088E03 3.475E+03 1.133E+07 1.642E+04 5.098E+03 3.775E+04 9.368E+00 2.109E+01
Worst 4.063E03 6.525E+03 2.110E+07 2.149E+04 5.741E+03 1.335E+05 1.616E+01 2.113E+01
Mean 9.059E04 2.859E+03 9.669E+06 1.365E+04 4.547E+03 3.194E+04 7.569E+00 2.105E+01
Std 9.635E04 1.531E+03 4.245E+06 3.951E+03 6.825E+02 3.255E+04 3.231E+00 4.823E02
1E5 Best 1.137E13 2.377E08 3.157E+05 8.176E+02 2.506E+03 7.151E+00 7.405E03 2.082E+01
7th 2.274E13 2.685E07 6.596E+05 2.258E+03 2.969E+03 2.089E+01 9.866E03 2.095E+01
Median 3.411E13 1.592E06 7.520E+05 4.299E+03 3.492E+03 7.556E+01 1.003E02 2.096E+01
19th 7.390E13 2.186E05 1.100E+06 5.527E+03 4.063E+03 7.922E+01 2.457E02 2.098E+01
Worst 6.594E12 1.571E+03 2.450E+06 1.332E+04 4.781E+03 8.677E+01 7.868E02 2.102E+01
Mean 7.594E13 7.295E+01 8.945E+05 4.281E+03 3.504E+03 5.895E+01 1.859E02 2.095E+01
Std 1.297E12 3.140E+02 4.346E+05 2.975E+03 6.222E+02 3.034E+01 1.609E02 4.615E02
1E6 Best 5.684E14 1.137E13 2.951E+04 5.824E06 1.960E+03 2.842E13 7.396E03 2.000E+01
7th 5.684E14 2.842E13 5.111E+04 3.882E05 2.618E+03 1.104E10 9.857E03 2.004E+01
Median 5.684E14 3.411E13 1.002E+05 7.785E05 3.023E+03 1.399E10 9.865E03 2.012E+01
19th 5.684E14 4.548E13 1.385E+05 2.683E04 3.428E+03 2.011E10 2.457E02 2.016E+01
Worst 1.137E13 7.958E13 1.893E+05 1.665E03 4.052E+03 1.462E09 7.866E02 2.033E+01
Mean 5.912E14 3.661E13 9.591E+04 2.160E04 3.049E+03 2.085E10 1.851E02 2.011E+01
Std 1.137E14 1.540E13 4.996E+04 3.547E04 5.803E+02 2.777E10 1.611E02 8.043E02
Table 15
FEs F9 F 10 F 11 F 12 F 13 F 14 F 15 F 16 F 17
1E3 Best 2.337E+02 3.371E+02 3.911E+01 7.165E+05 2.394E+01 1.381E+01 5.974E+02 4.117E+02 4.000E+02
7th 2.590E+02 3.681E+02 4.169E+01 7.970E+05 2.783E+01 1.388E+01 7.790E+02 4.930E+02 4.000E+02
19th 2.838E+02 3.979E+02 4.297E+01 8.983E+05 3.091E+01 1.409E+01 9.210E+02 6.092E+02 4.000E+02
Worst 3.127E+02 4.136E+02 4.507E+01 1.088E+06 3.408E+01 1.419E+01 9.987E+02 7.437E+02 4.000E+02
Mean 2.733E+02 3.798E+02 4.233E+01 8.494E+05 2.931E+01 1.399E+01 8.523E+02 5.556E+02 4.613E+02
Std 2.078E+01 2.162E+01 1.571E+00 8.546E+04 2.867E+00 1.185E01 1.015E+02 8.431E+01 6.852E+01
1E4 Best 2.565E+01 6.088E+01 2.199E+01 2.102E+04 2.609E+00 1.231E+01 2.506E+02 8.402E+01 1.139E+02
7th 2.962E+01 7.779E+01 2.821E+01 3.192E+04 3.803E+00 1.275E+01 3.564E+02 1.377E+02 1.589E+02
19th 3.798E+01 1.141E+02 3.258E+01 5.762E+04 5.101E+00 1.349E+01 4.735E+02 1.885E+02 2.643E+02
Worst 4.822E+01 1.550E+02 3.726E+01 1.024E+05 7.279E+00 1.387E+01 5.736E+02 4.258E+02 3.126E+02
Mean 3.345E+01 9.741E+01 2.993E+01 5.033E+04 4.546E+00 1.312E+01 4.189E+02 1.827E+02 4.224E+02
Std 6.370E+00 2.514E+01 3.606E+00 1.915E+04 1.152E+00 4.710E01 9.372E+01 8.314E+01 1.306E+02
1E5 Best 4.975E+00 5.870E+01 2.095E+01 1.172E+02 1.798E+00 1.224E+01 2.049E+02 6.843E+01 7.386E+01
7th 9.957E+00 7.462E+01 2.694E+01 5.967E+03 3.023E+00 1.261E+01 3.138E+02 9.582E+01 7.845E+01
19th 1.592E+01 1.134E+02 3.167E+01 1.646E+04 4.286E+00 1.322E+01 4.052E+02 1.649E+02 1.970E+02
Worst 1.990E+01 1.492E+02 3.638E+01 3.734E+04 6.111E+00 1.385E+01 5.109E+02 4.022E+02 2.393E+02
Mean 1.303E+01 9.338E+01 2.869E+01 1.151E+04 3.622E+00 1.292E+01 3.722E+02 1.510E+02 4.088E+02
Std 3.915E+00 2.524E+01 4.073E+00 8.849E+03 1.130E+00 4.357E01 9.345E+01 8.620E+01 1.544E+02
1E6 Best 5.684E14 5.870E+01 1.922E+01 1.378E+01 9.988E01 1.223E+01 2.000E+02 6.843E+01 7.253E+01
7th 1.137E13 7.462E+01 2.389E+01 4.892E+02 1.311E+00 1.245E+01 3.000E+02 9.576E+01 7.650E+01
Median 1.137E13 8.557E+01 2.694E+01 1.543E+03 1.623E+00 1.273E+01 4.000E+02 1.207E+02 9.724E+01
19th 1.137E13 1.134E+02 2.889E+01 2.909E+03 1.813E+00 1.294E+01 4.051E+02 1.649E+02 1.882E+02
Worst 1.705E13 1.492E+02 3.207E+01 1.002E+04 2.910E+00 1.334E+01 5.044E+02 4.022E+02 2.318E+02
Mean 1.069E13 9.338E+01 2.631E+01 2.695E+03 1.671E+00 1.270E+01 3.571E+02 1.509E+02 4.079E+02
Std 2.499E14 2.524E+01 3.359E+00 2.890E+03 4.692E01 3.147E01 1.052E+02 8.622E+01 1.557E+02
14 February 2015
Table 16
FEs F 18 F 19 F 20 F 21 F 22 F 23 F 24 F 25
1E3 Best 4.000E+02 9.061E+02 9.120E+02 9.187E+02 1.066E+03 1.095E+03 1.138E+03 1.166E+03
7th 4.000E+02 9.061E+02 9.123E+02 9.188E+02 1.085E+03 1.100E+03 1.157E+03 1.193E+03
Median 4.000E+02 9.068E+02 9.125E+02 9.188E+02 1.089E+03 1.107E+03 1.157E+03 1.206E+03
19th 4.000E+02 9.068E+02 9.134E+02 9.194E+02 1.094E+03 1.108E+03 1.164E+03 1.210E+03
Worst 4.000E+02 9.089E+02 9.134E+02 9.196E+02 1.095E+03 1.109E+03 1.164E+03 1.214E+03
Mean 6.011E+02 9.830E+02 1.019E+03 1.084E+03 1.161E+03 1.191E+03 1.233E+03 1.258E+03
Std 2.616E+02 9.298E+01 9.045E+01 1.048E+02 6.789E+01 6.154E+01 5.842E+01 5.051E+01
1E4 Best 4.015E+02 9.090E+02 9.110E+02 5.001E+02 8.864E+02 5.342E+02 2.003E+02 2.003E+02
7th 4.046E+02 9.090E+02 9.136E+02 5.002E+02 9.170E+02 5.547E+02 2.004E+02 2.005E+02
Median 4.554E+02 9.090E+02 9.141E+02 5.002E+02 9.380E+02 5.595E+02 2.009E+02 2.009E+02
19th 4.723E+02 9.097E+02 9.143E+02 5.010E+02 9.468E+02 5.608E+02 2.019E+02 2.013E+02
Worst 4.767E+02 9.102E+02 9.151E+02 1.031E+03 9.835E+02 1.122E+03 2.129E+02 2.014E+02
Mean 7.033E+02 9.840E+02 1.020E+03 1.027E+03 1.130E+03 1.105E+03 1.042E+03 1.059E+03
Std 2.630E+02 9.214E+01 8.970E+01 2.429E+02 1.155E+02 2.500E+02 4.302E+02 4.395E+02
1E5 Best 4.874E+02 9.085E+02 9.103E+02 5.000E+02 8.855E+02 5.342E+02 2.000E+02 2.000E+02
7th 5.000E+02 9.091E+02 9.129E+02 5.000E+02 9.107E+02 5.342E+02 2.000E+02 2.000E+02
Median 5.000E+02 9.105E+02 9.137E+02 5.000E+02 9.309E+02 5.342E+02 2.000E+02 2.000E+02
19th 5.000E+02 9.110E+02 9.151E+02 5.000E+02 9.398E+02 5.467E+02 2.000E+02 2.000E+02
Worst 5.031E+02 9.112E+02 9.152E+02 1.047E+03 9.825E+02 5.547E+02 2.000E+02 2.000E+02
Mean 7.969E+02 9.843E+02 1.020E+03 1.032E+03 1.130E+03 1.080E+03 1.041E+03 1.059E+03
Std 2.335E+02 9.191E+01 8.940E+01 2.446E+02 1.171E+02 2.779E+02 4.315E+02 4.398E+02
1E6 Best 5.065E+02 9.085E+02 9.103E+02 5.000E+02 8.854E+02 5.342E+02 2.000E+02 2.000E+02
7th 5.520E+02 9.091E+02 9.129E+02 5.000E+02 9.046E+02 5.342E+02 2.000E+02 2.000E+02
Median 5.762E+02 9.105E+02 9.137E+02 5.000E+02 9.309E+02 5.342E+02 2.000E+02 2.000E+02
19th 5.951E+02 9.121E+02 9.159E+02 5.000E+02 9.398E+02 5.342E+02 2.000E+02 2.000E+02
Worst 6.014E+02 9.122E+02 9.187E+02 1.051E+03 9.825E+02 5.342E+02 2.000E+02 2.000E+02
Mean 8.792E+02 9.844E+02 1.020E+03 1.037E+03 1.129E+03 1.078E+03 1.041E+03 1.059E+03
Std 1.775E+02 9.180E+01 8.905E+01 2.457E+02 1.176E+02 2.805E+02 4.315E+02 4.398E+02
592 the proposed algorithm can achieve a very small error value. On F 5 , the RGA-RDD demonstrated an excellent performance in
593 the case of 10-dimension in finding out the global optimum that locates exactly on the boundary. On F 6 , the RGA-RDD pre-
594 sented a superior performance to against local traps and accurately located the global optimum that lies in a very narrow and
595 sharp valley. On F 7 , although the global optimum is located outside the initialization range, the RGA-RDD still successfully
596 directed the searching toward the right region and effectively found out a solution which is very close to the global optimum.
597 On F 8 , the RGA-RDD, like all of these comparative algorithms, totally failed to locate the optimum solution on this specific
598 version of multimodal Ackley functions. This can be explained by the fact that, as a result of the linear transformation by a
599 condition number of 100, the composition of F 8 results in a flat but unachievable region in the search space. As a conse-
600 quence, the F 8 function shapes like a needle in the haystack problem, and that brings difficulties to the RGA-RDD as well
601 as to all the comparative algorithms. On F 9 , the RGA-RDD presented a good performance for not being trapped by a huge
602 number of local optima; as a result, a very high success rate of locating the global solution was obtained. However, the func-
603 tion F 10 , which is a rotated version of F 9 , causes an obstacle for the RGA-RDD to achieve the desired accuracy level within the
604 given budget of MaxFEs. Besides, we found that the functions ranging from F 11 to F 15 are hard to be solved by most of the
605 comparative optimization algorithms because of their complexity and multimodal characteristics. As a consequence, the
606 RGA-RDD, like other comparative algorithms, did not perform satisfactorily on the hybrid functions from F 16 to F 25 . This dif-
607 ficulty is attributed to the fact that these hybrid functions can lead algorithms to divert away from the global region due to
608 their complicated profiles such as a huge number of local optima and a narrow global basin surrounded by a large number of
609 local maxima.
610 4.4.2. The performance evaluation of paired benchmark functions

611 On the basis of the previous simulation experiments, we further explored the effectiveness of the proposed RGA-RDD by
612 performing some more comparisons made on those intentionally designate, paired benchmark functions. For example, from
613 the comparison between F 1 ; F 2 , and F 3 , we found that a higher condition number can deteriorate the performance of the pro-
614 posed algorithm, especially when handling the 30-dimensional problems. Besides, we also observed from the comparison
615 between F 2 and F 4 that additional noise would affect the performance of the algorithm because the proposed DBX operation
616 makes the direct use of the relative fitness information to produce candidate solutions. On the other hand, we discovered
617 from the comparison results made on F 5 and F 7 that the DBX operation is very effective in finding the global optimum
618 solution which lies exactly on the boundaries or outside the initialization range. A reason for this is that the directional
619 search vector generated by the DBX operation can effectively direct the searching toward the region or boundary where
14 February 2015
Table 17
Comparisons of normalized SP for 10-dimension solved unimodal functions.
Algorithms F1 F2 F3 F4 F5
1000 2400 6500 2900 5900
RGA-RDD 10.0(25) 11.5(25) – 20.8(25) 1.6(25)
BLX-GL50 19.0(25) 17.1(25) – 14.5(25) 4.7(25)
BLX-MA 12.0(25) 15.4(25) – 25.9(24) –
CoEVO 23.0(25) 11.3(25) 6.8(25) 16.2(25) –
DE 29.0(25) 19.2(25) 18.5(20) 17.9(25) 6.9(25)
DMS-L-PSO 12.0(25) 5.0(25) 1.8(25) – 18.6(20)
EDA 10.0(25) 4.6(25) 2.5(23) 4.1(25) 4.2(25)
G-CMA-ES 1.6(25) 1.0(25) 1.0(25) 1.0(25) 1.0(25)
K-PCX 1.0(25) 1.0(25) – 19.7(21) –
L-SaDE 10.0(25) 4.2(25) 8.0(16) 15.9(24) –
SPC-PNX 6.7(25) 12.9(25) – 10.7(25) 6.8(25)
Table 18
Comparisons of normalized SP for 30-dimension solved unimodal functions.
Algorithms F1 F2 F3 F4 F5
2700 12,000 43,000 59,000 66,000
RGA-RDD 15.1(25) 37.0(25) – – –
BLX-GL50 21.5(25) 13.3(25) – – –
BLX-MA 11.9(25) – – – –
CoEVO 519 (3) 70.0(8) – – –
DE 51.9(25) – – – –
DMS-L-PSO 1.9(25) 10.8(25) 7.9(21) – –
EDA 55.6(25) 13.3(25) 5.1(25) 3.4(25) –
G-CMA-ES 1.7(25) 1.1(25) 1.0(25) 1.0(10) 1.0(25)
K-PCX 1.0(25) 1.0(25) – – –
L-SaDE 7.4(25) 12.5(24) – 9.2(13) –
SPC-PNX 11.1(25) 26.7(22) – 6.1(19) –
Table 19
Comparisons of normalized SP for 10-dimensional solved multimodal functions.
Algorithms F6 F7 F9 F 10 F 11 F 12 F 15
7100 4700 17,000 55,000 190,000 8200 33,000
RGA-RDD 23.1(25) – 2.7(25) – – 57.9(15) 4.5(22)
BLX-GL50 7.3(25) 12.3(9) 10.0(3) – – 12.1(13) –
BLX-MA – – 5.7(18) – – – 8.5(5)
CoEVO – – – – – – –
DE 6.6(24) 255(2) 10.6(11) – 1.0(12) 8.8(19) 75.8(1)
DMS-L-PSO 7.7(25) 126(4) 2.1(25) – – 6.6(19) 1.7(22)
EDA 9.6(22) 404(1) – – 2.9(3) 4.3(10) –
G-CMA-ES 1.5(25) 1.0(25) 4.5(19) 1.2 (23) 1.4(6) 4.0(22) –
K-PCX 1.0(22) – 2.9(24) 1.0(22) – 1.0(14) –
L-SaDE 6.9(25) 36.2(6) 1.0(25) – – 3.9(25) 1.0(23)
SPC-PNX – 383(1) – – 5.8(1) – –
620 the global optimum locates. Moreover, from the comparison results of F 6 ; F 7 ; F 9 , and F 15 , we confirmed that the proposed
621 RGA-RDD is able to provide a superior ability to against the local traps, especially for those multimodal functions with an
622 adequate global structure in the feasible region. As previously mentioned in Subsection 3.2 as well as in Fig. 7, this superior
623 performance is mainly contributed by the combined use of the RS, DBX, and DRM operated in the parallel-structured
624 algorithm that significantly enhances the possibility of locating the global optimum.
625 4.4.3. Further performance evaluation of multimodal and hybrid functions

626 Finally, to compare the results of these comparative algorithms in a more systematic way, we regrouped the benchmark
627 functions F 1 F 25 into the following three categories: (1) solved unimodal functions: F 1 F 5 ; (2) solved multimodal func-
628 tion: F 6 F 7 ; F 9 F 12 ; F 15 ; and (3) never solved multimodal functions and hybrid functions: F 8 ; F 13 ; F 14 ; F 16 F 25 . According
629 to the comparative criteria stated in [84,96], all statistical performance data were calculated from the error values of 25 runs
14 February 2015
630 at the end of n 104 FEs. The comparison results are, respectively, listed in Tables 17–22 for different groups of functions
631 under different dimensions. We note that the table entries for these solved problems are the normalized success perfor-
632 mance and the number of successful runs (in round brackets) of the algorithm, while the table entries for those unsolved
633 problems are the rankings of the algorithms sorted according to the median error values.
634 For comparison, we depict in Tables 17 and 18 the performance of the solved unimodal functions in the case of 10-dimen-
635 sion and that of 30-dimension, respectively. From the comparison of these two tables, we found that almost every algorithm
636 has nearly the same success rate and success performance in 10-dimensional unimodal functions. However, when the prob-
637 lem dimension was increased to 30, the RGA-RDD and these comparative algorithms all failed to reach the desired accuracy
638 level for F 3 , F4, and F 5 , but the RGA-RDD still maintained a relatively high success rate in F 1 and F 2 , apparently better than
639 these comparative algorithms.
Table 20
Comparisons of normalized SP for 30-dimension solved multimodal functions.
Algorithms F6 F7 F9 F 10 F 11 F 12 F 15
60,000 6100 99,000 450,000 5,000,000 180,000 –
RGA-RDD 23.0(25) 76.3(14) 7.7(25) – – – –
BLX-GL50 3.7(25) 10.2(25) – – – – –
BLX-MA – – 6.7( 9) – – – –
CoEVO – 93.4(11) – – – – –
DE – 32.8(22) – – – – –
DMS-L-PSO 5.5(24) 9.8(24) – – – 8.3( 4) –
EDA – 21.3(25) – – – – –
G-CMA-ES 1.0(25) 1.0(25) 8.0( 9) 5.3( 3) 1.0( 1) 1.3( 8) –
K-PCX 1.1(14) 2.5(10) 3.3(18) 1.0(14) – 1.0( 5) –
L-SaDE – 21.3(20) 1.0(25) – – – –
SPC-PNX 86.7(1) 60.7(16) – – – – –
Table 21
Rank of median error value for 10-dimension never solved functions.
Algorithms F8 F 13 F 14 F 16 F 17 F 18 F 19 F 20 F 21 F 22 F 23 F 24 F 25
RGA-RDD 8 3 8 8 6 6 6 6 1 7 4 1 1
BLX-GL50 9 6 2 2 2 1 1 1 1 3 5 1 7
BLX-MA 5 7 1 6 8 7 7 7 10 5 10 1 6
CoEVO 6 9 11 11 11 11 10 11 9 11 9 10 2
DE 10 10 9 10 9 1 1 1 1 10 1 1 11
DMS-L-PSO 1 2 3 4 3 7 7 7 1 6 8 1 10
EDA 7 11 10 9 10 1 1 1 1 9 5 1 5
G-CMA-ES 1 4 6 1 7 1 1 1 1 2 1 1 3
K-PCX 1 5 4 3 1 10 11 10 11 1 11 11 8
L-SaDE 1 1 5 5 4 7 7 7 1 4 3 1 4
SPC-PNX 11 8 7 7 5 1 1 1 1 8 5 1 9
Table 22
Rank of median error value for 30-dimension never solved functions.
Algorithms F8 F 13 F 14 F 16 F 17 F 18 F 19 F 20 F 21 F 22 F 23 F 24 F 25
RGA-RDD 9 5 3 5 4 1 6 7 1 6 2 1 1
BLX-GL50 8 3 1 3 1 4 3 3 1 2 6 8 3
BLX-MA 3 1 2 9 5 3 2 2 1 5 4 1 5
CoEVO 5 7 6 8 9 9 9 9 6 8 8 9 9
DE 4 6 7 7 8 7 7 6 8 7 7 6 8
DMS-L-PSO N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A
EDA 7 9 8 6 7 8 8 8 7 3 5 5 7
G-CMA-ES 1 2 4 1 6 5 5 4 1 1 1 7 2
K-PCX 1 8 9 2 2 2 1 1 9 9 9 4 6
L-SaDE N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A N/A
SPC-PNX 6 4 5 4 3 6 3 5 1 4 2 1 3
N/A indicates the data are not available in the literature.
14 February 2015
640 In addition, we list in Tables 19 and 20 the comparative performances of the solved multimodal functions in 10-dimen-
641 sion and 30-dimension, respectively. From the results shown in these two tables, we observed that the proposed RGA-RDD
642 presents a superior performance in solving both F 9 and F 15 , better than these comparative algorithms. Especially, by com-
643 paring the results of the RGA-RDD with other comparative RCGAs such as BLX-GL50, BLX-MA, and SPC-PNX, we found that
644 the RGA-RDD is able to provide a considerably higher success rate and a better success performance, especially when dealing
645 with multimodal functions.
646 Moreover, we list in Tables 21 and 22 the rankings of these algorithms sorted by the median of the error values of the
647 obtained results. From these two tables, it can be shown obviously that the RGA-RDD offered a moderate performance in
648 10-dimensional cases and its performance in the case of 30-dimension is rather remarkable, especially for the functions
649 of F 18 ; F 21 ; F 23 ; F 24 , and F 25 . The above simulation results also reveal that, though the RGA-RDD may not be on the top list
650 to solve all kinds of optimization problems, its parallel-structured coordinator and the specially designed evolutionary oper-
651 ators greatly assist the proposed algorithm to locate the global optimum of a real-parameter optimization problem, espe-
652 cially for those unsolved multimodal and the high-dimensional hybrid functions. Thus, as shown in Tables 21 and 22, the
653 parallel-structured RGA-RDD ranks first among the comparison algorithms in solving F21, F24, and F25.
654 5. Conclusions
655 In this paper, we have developed a parallel-structured RGA-RDD to solve real-parameter optimization problems. Unlike
656 conventional RCGAs that manage evolutionary operators in a series framework, the proposed RGA-RDD incorporates three
657 specially designed RS, DBX, and DRM as a whole to mimic a specific evolutionary process that has an inner parallel structure.
658 The effectiveness and applicability of each operator have been demonstrated through a series of basic studies, showing that
659 the combined use of the RS, DBX, and DRM profoundly benefits the proposed parallel-structured RGA-RDD to locate an opti-
660 mal solution. Besides, based on a parametric analysis, a guideline for parameter settings of RGA-RDD has been provided for
661 practical applications. Furthermore, we have successfully applied a data-driven optimization scheme that incorporates the
662 uniform design of experiments and a shape-tunable neural network to search for an optimal set of the algorithm parameters.
663 As a rigorous base of performance comparison, the CEC2005 testbed, which contains 25 well-motivated benchmark func-
664 tions, has been used to evaluate the performance of the proposed RGA-RDD as well as the state-of-the-art evolutionary algo-
665 rithms. The extensive comparison results show that the RGA-RDD significantly outperforms the comparative RCGAs such as
666 BLX-GL50, BLX-MA, and SPC-PNX, especially for some multimodal and hybrid functions. Besides, the simulation results
667 reveal that, though the RGA-RDD may not be on the top list to solve all kinds of optimization problems, it provides an excel-
668 lent ability to solve real-parameter optimization problems, especially for those unsolved multimodal and high-dimensional
669 hybrid functions.
670 Based on the benchmarking results obtained in this paper and the techniques developed herein, we summarize the dis-
671 tinct features and advantages of the proposed parallel-structured RGA-RDD as follows: (1) Being operated simply using the
672 information of fitness rankings, the RS presents to be a very simple and effective mechanism to assist the DBX to produce
673 potential search directions around the current best solution. (2) The DBX systematically divides the population into N/2 pairs
674 according to fitness rankings, and then it makes the direct use of the relative fitness information of each pair of parents to
675 conduct 2n 1 crossover directions for exploring potential offspring chromosomes. The concept of the proposed DBX is sim-
676 ple and distinct from the conventional crossover operators developed using the techniques of line segment connection and
677 the distribution analysis. (3) The DRM dynamically adjusts the mutation size through successive generations so as to effec-
678 tively prevent the evolution from being trapped by a local optimal and at the same time enhance the precision of the
679 obtained optimal solution. (4) The evolutionary operation of the RGA-RDD does not just follow a sequential cycle like con-
680 ventional RCGAs do; it employs an inner-parallel coordinator to manage the operation of the DBX and DRM when the two
681 selected parents have the same fitness and/or same genetic expressions. This effort makes every evolutionary step count and
682 effective in producing better offspring and has been proven to be successful in locating the global optimum. Due to its
683 excellent performance, effective evolutionary operators, and the significant parallel-structured configuration, the proposed
684 RGA-RDD appears to be an effective, applicable, and promising scheme, alternative to existing RCGAs in the literature, for
685 real-parameter numerical optimization. Finally, we mention that the proposed RGA-RDD has been successfully applied to
686 assist the recently developed cell evolution method [12] in solving the reliability-based design optimization problems.
687 6. Uncited references
688 [93,94].
689 Acknowledgements
690 This work was supported by the National Science Council of Taiwan under Grants NSC99-2221-E-035-090 and NSC100-
691 2221-E-035-039. The authors thank the anonymous reviewers for their valuable and instructive comments that greatly help
692 improve the quality and completeness of this paper.
14 February 2015
693 Appendix A
694 The benchmarking functions for basic studies in Section 3 are defined as follows:
695 (1) Sphere unimodal function

696
X
n
f sphere ðxÞ ¼ z2j ; x 2 ½100; 100n
698 j¼1
699 (2) Ackley multimodal function

700 0 vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
u X
X
u1 n 2 1 n
f ackley ðxÞ ¼ 20 exp @0:2t zj A exp cos 2 p z j þ 20 þ exp ð1Þ x 2 ½32; 32n
n j¼1 n j¼1
702
703 where z x o; x ¼ ½x1 ; x2 ; . . . ; xn , and o ¼ ½o1 ; o2 ; . . . ; on is the shifted global optimum.
704 References
705 [1] M. Affenzeller, S. Wagner, S. Winkler, Self-adaptive population size adjustment for genetic algorithms, Lect. Notes Comput. Sci. 4739 (2007) 820–828.
706 [2] N. Amjady, H. Nasiri-Rad, Solution of nonconvex and nonsmooth economic dispatch by a new adaptive real coded genetic algorithm, Expert Syst. Appl.
707 37 (2010) 5239–5245.
708 [3] J. Andre, P. Siarry, T. Dognon, An improvement of the standard genetic algorithm fighting premature convergence in continuous optimization, Adv. Eng.
709 Softw. 32 (2001) 49–60.
710 [4] M.S. Arumugam, M.V.C. Rao, R. Palaniappan, New hybrid genetic operators for real coded genetic algorithm to compute optimal control of a class of
711 hybrid systems, Appl. Soft Comput. 6 (2005) 38–52.
712 [5] A. Auger, N. Hansen, A restart CMA evolution strategy with increasing population size, in: 2005 IEEE Congress on Evolutionary Computation, vol. 2,
713 Edinburgh, UK, 2005, pp. 1769–1776.
714 [6] P.J. Ballester, J. Stephenson, J.N. Carter, K. Gallagher, Real-parameter optimization performance study on the CEC-2005 benchmark with SPC-PNX, in:
715 2005 IEEE Congress on Evolutionary Computation, vol. 1, Edinburgh, UK, 2005, pp. 498–505.
716 [7] H.G. Beyer, The Theory of Evolution Strategies, Springer-Verlag, Berlin, Germany, 2000.
717 [8] H.G. Beyer, K. Deb, On self-adaptive features in real-parameter evolutionary algorithms, IEEE Trans. Evolution. Comput. 5 (2001) 250–270.
718 [9] T. Blickle, L. Thiele, A comparison of selection schemes used in evolutionary algorithms, Evolution. Comput. 4 (1996) 361–394.
719 [10] E. Cantú-Paz, On random numbers and the performance of genetic algorithms, in: Proceedings of Genetic and Evolutionary Computation Conference,
720 New York, NY, USA, 2002, pp. 311–318.
721 [11] C.T. Chen, W.D. Chang, A feedforward neural network with function shape autotuning, Neural Netw. 9 (1996) 627–641.
722 [12] C.T. Chen, M.H. Chen, W.T. Horng, A cell evolution method for reliability-based design optimization, Appl. Soft Comput. 15 (2014) 67–79.
723 [13] C.T. Chen, C.K. Wu, C. Hwang, Optimal design and control of CPU heat sink processes, IEEE Trans. Compon. Pack. Technol. 31 (2008) 184–195.
724 [14] C.T. Chen, Y.C. Chuang, An intelligent run-to-run control strategy for chemical–mechanical polishing processes, IEEE Trans. Semicond. Manuf. 23
725 (2010) 109–120.
726 [15] G. Chen, C.P. Low, Z. Yang, Preserving and exploiting genetic diversity in evolutionary programming algorithms, IEEE Trans. Evol. Comput. 13 (2009)
727 661–673.
728 [16] Z.Q. Chen, R.L. Wang, Two efficient real-coded genetic algorithms for real parameter optimization, Int. J. Innov. Inf. Control 7 (2001) 4871–4883.
729 [17] Y.C. Chuang, C.T. Chen, A study on real-coded genetic algorithm for process optimization using ranking selection, direction-based crossover and
730 dynamic mutation, in: Proceedings of the IEEE Congress on Evolution Computation, pp. 2488–2495, New Orleans, USA, 2011.
731 [18] Y.C. Chuang, C.T. Chen, Mathematical modeling and optimal design of an MOCVD reactor for GaAs film growth, J. Taiwan Inst. Chem. Engrs. 45 (2014)
732 254–267.
733 [19] S. Das, P.N. Suganthan, Differential evolution: a survey of the state-of-the-art, IEEE Trans. Evol. Comput. 15 (2011) 4–31.
734 [20] K. Deb, R.B. Agrawal, Simulated binary crossover for continuous search space, Complex Syst. 9 (1995) 115–148.
735 [21] K. Deb, A. Anand, D. Joshi, A computationally efficient evolutionary algorithm for real-parameter optimization, Evolution. Comput. 10 (2002) 371–395.
736 [22] K. Deb, M. Goyal, A combined genetic adaptive search (GeneAS) for engineering design, Comp. Sci. Inform. 26 (1996) 30–45.
737 [23] K. Deep, K.N. Das, Performance improvement of real coded genetic algorithm with quadratic approximation based hybridisation, Int. J. Intell. Def. Supp.
738 Syst. 2 (2009) 319–334.
739 [24] K. Deep, M. Thakur, A new crossover operator for real-coded genetic algorithms, Appl. Math. Comput. 218 (2012) 4710–4730.
740 [25] M. Dorigo, Optimization, Learning and Natural Algorithms, Ph.D. thesis, Politecnico di Milano, Italy, 1992.
741 [26] M. Dorigo, M. Birattari, T. Stutzle, Ant colony optimization, IEEE Comput. Intell. Mag. 1 (2006) 28–39.
742 [27] J.D. Dyer, R.J. Hartfield, G.V. Dozier, J.E. Burkhalter, Aerospace design optimization using a steady state real-coded genetic algorithm, Appl. Math.
743 Comput. 218 (2012) 4710–4730.
744 [28] A.E. Eiben, R. Hinterding, Z. Michalewicz, Parameter control in evolutionary algorithms, IEEE Trans. Evol. Comput. 3 (1999) 124–141.
745 [29] T.A. El-Mihoub, A.A. Hopgood, L. Nolle, A. Battersby, Hybrid genetic algorithm: a review, Eng. Lett. 13 (2006) 124–137.
746 [30] L.J. Eshelman, J.D. Schaffer, Real-coded genetic algorithms and interval-schemata, in: Proceedings of the Workshop on Foundations of Genetic
747 Algorithms, Vail, CO, USA, 1993, pp. 187–202.
748 [31] K.T. Fang, D.K. Lin, P. Winker, Y. Zhang, Uniform design: theory and application, Technometrics 42 (2000) 237–248.
749 [32] C. Garcia-Martinez and M. Lozano, Hybrid real-coded genetic algorithms with female and male differentiation, in: 2005 IEEE Congress on Evolutionary
750 Computation, vol. 1, Edinburgh, UK, 2005, pp. 896–903.
751 [33] S. Ghosh, S. Das, S. Roy, S.K.M. Islam, P.N. Suganthan, A differential covariance matrix adaptation evolutionary algorithm for real parameter
752 optimization, Inf. Sci. 182 (1997) 199–219.
753 [34] D.E. Goldberg, Genetic Algorithm in Search Optimization and Machine Learning, Addison-Wesley, 1989.
754 [35] D.E. Goldberg, Real-coded genetic algorithm, virtual alphabets, and blocking, Complex Syst. 5 (1991) 139–167.
755 [36] D.E. Goldberg, The Design of Innovation: Lessons From and Component Genetic Algorithms, Kluwer Academic Publishers, Boston, MA, 2002.
756 [37] D.E. Goldberg, K. Deb, J.H. Clark, Genetic algorithms, noise, and the sizing of population, Complex Syst. 6 (1992) 333–362.
757 [38] D.E. Goldberg, K. Deb, D. Thierens, Toward a better understanding of mixing in genetic algorithms, J. Soc. Instrum. Control Eng. 32 (1993) 10–16.
758 [39] N. Hansen, A. Ostermeier, Completely derandomized self-adaptation in evolution strategies, Evolution. Comput. 9 (2001) 159–195.
759 [40] F. Herrera, M. Lozano, J.L. Verdegay, Tackling real-coded genetic algorithms: operators and tools for behavioural analysis, Artif. Intell. Rev. 12 (1998)
760 265–319.
14 February 2015
761 [41] T. Higuchi, S. Tsutsui, M. Yamamura, Theoretical analysis of simplex crossover for real-coded genetic algorithms, in: Proceedings of the Sixth
762 International Conference on Parallel Problem Solving from Nature, 2000, pp. 365–374.
763 [42] O. Hrstka, A. Kucerova, Improvements of real coded genetic algorithms based on differential operators preventing premature convergence, Adv. Eng.
764 Softw. 35 (2004) 237–246.
765 [43] C.Z. Janikow, Z. Michalewicz, An experimental comparison of binary and floating point representations in genetic algorithms, in: Proceedings of the
766 Fourth International Conference on Genetic Algorithms, 1991, pp. 31–36.
767 [44] J. Kennedy, R. Eberhart, Particle swarm optimization, In: Proceedings of IEEE International Conference on Neural Networks IV, 1995, pp. 1942-1948.
768 [45] S. Kirkpatrick, C.D. Gelatt Jr., M.P. Vecchi, Optimization by simulated annealing, Science 220 (1983) 671–680.
769 [46] H. Kita, I. Ono, S. Kobayashi, Theoretical analysis of the unimodal normal distribution crossover for real-coded genetic algorithms, in: Proceedings of
770 the International Conference on Evolutionary Computation, 1998, pp. 529–534.
771 [47] I. Korejo, S.X. Yang, C. Li, A Directed Mutation Operator for Real Coded Genetic Algorithms, Applications of Evolutionary Computation, Springer, Berlin
772 Heidelberg, 2010.
773 [48] V.K. Koumousis, C.P. Katsaras, A saw-tooth genetic algorithm combining the effects of variable population size and reinitialization to enhance
774 performance, IEEE Trans. Evol. Comput. 10 (2006) 19–28.
775 [49] J.R. Koza, Genetic Programming, MIT Press, Cambridge, 1992.
776 [50] H.C. Kuo, C.H. Lin, A directed genetic algorithm for global optimization, Appl. Math. Comput. 219 (2013) 7348–7364.
777 [51] W. Langdon, Directed crossover within genetic programming, Advances in Genetic Programming 2 (1995). Number RN/95/71.
778 [52] J.J. Liang, P.N. Suganthan, Dynamic multi-swarm particle swarm optimizer with local search, in: 2005 IEEE Congress on Evolutionary Computation, vol.
779 1, Edinburgh, UK, 2005, pp. 522–528.
780 [53] J.J. Liang, P.N. Suganthan, K. Deb, Novel composition test functions for numerical global optimization, in: 2005 IEEE Swarm Intelligence Symposium,
781 Pasadena, CA, USA, 2005, pp. 68–75.
782 [54] M. Lozano, F. Herrera, J.R. Cano, Replacement strategies to preserve useful diversity in steady-state genetic algorithms, Inf. Sci. 178 (2008) 4421–4433.
783 [55] M. Lozano, F. Herrera, N. Krasnogor, D. Molina, Real-coded memetic algorithms with crossover hill-climbing, Evolution. Comput. 12 (2004) 273–302.
784 [56] D.J.C. MacKay, Information Theory, Inference, and Learning Algorithms, Cambridge University Press, Cambridge, 2003.
785 [57] R. Mallipeddi, S. Mallipeddi, P.N. Suganthan, Ensemble strategies with adaptive evolutionary programming, Inf. Sci. 180 (2010) 1571–1581.
786 [58] M.M. Meysenburg, J.A. Foster, The quality of pseudo-random number generators and simple genetic algorithm performance, in: Proceedings of Seventh
787 International Conference on Genetic Algorithms (ICGA-7), East Lansing, MI, USA, 1997, pp. 276–281.
788 [59] M.M. Meysenburg, D. Hoelting, D. McElvain, J.A. Foster, How random generator quality impacts genetic algorithm performance, in: Proceedings of
789 Genetic and Evolutionary Computation Conference, New York, NY, USA, 2002, pp. 480–487.
790 [60] E. Mezura-Montes, C.A. Coello Coello, A simple multimembered evolution strategy to solve constrained optimization problems, IEEE Trans. Evol.
791 Comput. 9 (2005) 1–17.
792 [61] Z. Michalewicz, Genetic Algorithm + Data Structure = Evolution Programs, Springer-Verlag, Berlin Heidelbreg, New York, 1992.
793 [62] D. Molina, F. Herrera, M. Lozano, Adaptive local search parameters for real-coded memetic algorithms, in: 2005 IEEE Congress on Evolutionary
795 [63] H. Nakanishi, H. Kinjo, N. Oshiro, T. Yamamoto, Searching performance of a real-coded genetic algorithm using biased probability distribution function
796 and mutation, Artif. Life Robot. 11 (2007) 37–41.
797 [64] A. Nolte, R. Schrader, A note on the finite time behavior of simulated annealing, Math. Operat. Res. 25 (2000) 476–484.
798 [65] I. Ono, H. Kita, S. Kobayashi, A robust real-coded genetic algorithm using unimodal normal distribution crossover augmented by uniform crossover:
799 effects of self-adaptation of crossover probabilities, in: Proceedings of the Genetic and Evolutionary Computation Conference, 1999, pp. 496–503.
800 [66] I. Ono, S. Kobayashi, A real-coded genetic algorithm for functional optimization using unimodal normal distribution crossover, in: Proceedings of
801 Seventh International Conference on Genetic Algorithms (ICGA-7), East Lansing, MI, USA, 1997, pp. 246–253.
802 [67] D. Ortiz-Boyer, C. Hervás-Martínez, N. García-Pedrajas, Improving crossover operator for real-coded genetic algorithm using virtual parents, J. Heurist.
803 13 (2007) 265–314.
804 [68] T. Park, K.R. Ryu, A dual-population genetic algorithm for adaptive diversity control, IEEE Trans. Evol. Comput. 14 (2010) 865–884.
805 [69] R. Poli, W.B. Langdon, N.F. McPhee, A Field Guide to Genetic Programming [FREE Online Book], 2011 <http://www.gp-fieldguide.org.uk>.
806 [70] P. Posik, Real-parameter optimization using the mutation step co-evolution, in: 2005 IEEE Congress on Evolutionary Computation, vol. 1, Edinburgh,
807 UK, 2005, pp. 872–879.
808 [71] W.H. Press, S.A. Teukolsky, W.T. Vetterling, B.P. Flannery, Section 22.3. Gray Codes, Numerical Recipes: The Art of Scientific Computing, third ed.,
809 Cambridge University Press, New York, 2007.
810 [72] A. Qi, F. Palmieri, Theoretical analysis of evolutionary algorithms with an infinite population size in continuous space. Part I: basic properties of
811 selection and mutation, IEEE Trans. Neural Netw. 5 (1994) 102–119.
812 [73] A.K. Qin, P.N. Suganthan, Self-adaptive differential evolution algorithm for numerical optimization, in: 2005 IEEE Congress on Evolutionary
814 [74] N. Rangel-Valdez, J. Torres-Jimenéz, J. Bracho-Ríos, P. Quiz-Ramos, Problem and algorithm fine-tuning – a case of study using bridge club and
815 simulated annealing, in: Proceedings of IJCCI, 2009, pp. 302–305.
816 [75] K.S.N. Ripon, S. Kwong, K.F. Man, A real-coding jumping gene genetic algorithm (RJGGA) for multiobjective optimization, Inf. Sci. 177 (2007) 632–654.
817 [76] J. Ronkkonen, S. Kukkonen, K.V. Price, Real-parameter optimization with differential evolution, in: 2005 IEEE Congress on Evolutionary Computation,
818 vol. 1, Edinburgh, UK, 2005, pp. 506–513.
819 [77] R.K. Roy, A Primer on the Taguchi Method, second ed., Society of Manufacturing Engineers, 2010.
820 [78] A.M. Sanchez, M. Lozano, P. Villar, F. Herrera, Hybrid crossover operators with multiple descendents for real-coded genetic algorithms: combining
821 neighborhood-based crossover operators, Int. Intell. Syst. 24 (2009) 540–567.
822 [79] H. Satoh, M. Yamaura, S. Kobayashi, Minimal generation gap model for GAs considering both exploration and exploitation, in: Proceedings of IIZUKA:
823 Methodologies for the Computation, Design, an Application of Intelligent System, 1996, pp. 494–497.
824 [80] A. Sinha, S. Tiwari and K. Deb, A population-based, steady-state procedure for real-parameter optimization, in: 2005 IEEE Congress on Evolutionary
825 Computation, vol. 1, Edinburgh, UK, 2005, pp. 514-521.
826 [81] M. Srinivas, L.M. Patnaik, Adaptive probabilities of crossover and mutation in genetic algorithms, IEEE Trans. Syst., Man Cybernet. 24 (1994) 656–667.
827 [82] P. Subbaraj, R. Rengaraj, S. Salivahanan, Enhancement of combined heat and power economic dispatch using self adaptive real-coded genetic
828 algorithm, Appl. Energy 86 (2009) 915–921.
829 [83] P. Subbaraj, R. Rengaraj, S. Salivahanan, Enhancement of self-adaptive real-coded genetic algorithm using Taguchi method for Economic dispatch
830 problem, Appl. Soft Comput. 11 (2011) 83–92.
831 [84] P.N. Suganthan, N. Hansen, J.J. Liang, K. Deb, A. Chen, Y.P. Auger, S. Tiwari, Problem Definitions and Evaluation Criteria for the CEC 2005 Special Session
832 on Real-Parameter Optimization, Technical Report, Nanyang Technological University, Singapore, 2005 <http://www.ntu.edu.sg/home/EPNSugan>.
833 [85] P.D. Surry, N. Radcliffe, Real representations, in: Proceedings of the Workshop on Foundations of Genetic Algorithms, 1996, pp. 343–363.
834 [86] P.H. Tang, M.H. Tseng, Adaptive directed mutation for real-coded genetic algorithms, Appl. Soft Comput. 13 (2013) 600–614.
835 [87] C.W. Tasi, C.L. Lin, C.H. Huang, Microbrushless DC motor control design based on real-coded structural genetic algorithm, IEEE/ASME Trans. Mechatron.
836 16 (2011) 151–159.
837 [88] T.K. Todd, Error Correction Coding, John Wiley & Sons, New Jersey, 2005.
838 [89] J. Torres-Jimenez, E. Rodriguez-Tello, New bounds for binary covering arrays using simulated annealing, Inf. Sci. 185 (2012) 137–152.
839 [90] S. Tsutsi, D.E. Goldberg, Search space boundary extension method in real-coded genetic algorithms, Inf. Sci. 133 (2001) 229–247.
14 February 2015
840 [91] S. Tsutsui, M. Yamamura, T. Higuchi, Multi-parents recombination with simplex crossover in real-coded genetic algorithms, in: Proceedings of the
841 Genetic and Evolutionary Computation Conference (GECCO-99), Orlando, FL, USA, 1999, pp. 675–664.
842 [92] K. Valarmathi, D. Devaraj, T.K. Radhakrishnan, Real-coded genetic algorithm for system identification and controller tuning, Appl. Math. Model. 33
843 (2009) 3392–3401.
844 [93] D.R. White, A. Arcuri, J.A. Clark, Evolutionary improvement of programs, IEEE Trans. Evol. Comput. 15 (2011) 515–538.
845 [94] A.H. Wright, Genetic algorithms for real parameter optimization, in: Proceedings of the Workshop on Foundations of Genetic Algorithms, 1991, pp.
846 205–218.
847 [95] Y. Yoon, Y.-H. Kim, A. Moraglio, B.-R. Moon, A theoretical and empirical study on unbiased boundary-extended crossover for real-valued
848 representation, Inf. Sci. 183 (2012) 48–65.
849 [96] B. Yuan, M. Gallagher, Experimental results for the special session on real-parameter optimization at CEC 2005: a simple, continuous EDA, in: 2005
850 IEEE Congress on Evolutionary Computation, vol. 2, Edinburgh, UK, 2005, pp. 1792–1799.
851 [97] Q. Yuan, F. Qian, W. Du, A hybrid genetic algorithm with the Baldwin effect, Inf. Sci. 180 (2010) 2815–2833.
852

Information Sciences: Yao-Chen Chuang, Chyi-Tsong Chen, Chyi Hwang

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Information Sciences: Yao-Chen Chuang, Chyi-Tsong Chen, Chyi Hwang

Uploaded by

Copyright:

Available Formats

INS 11379 No.

of Pages 29, Model 3G

Contents lists available at ScienceDirect

3 A real-coded genetic algorithm with a direction-based

51 1.1. A review of conventional real-coded genetic algorithms

Fig. 1. The ﬂow chart of the conventional RCGA optimization scheme.

82 1.2. Related works on the development of crossover operators

124 1.3. The contribution of this work

170 1.4. The organization of this work

180 2. The proposed RGA-RDD

181 2.1. The algorithm structure and the underlying ideas

Fig. 2. The ﬂow chart of the proposed RGA-RDD optimization scheme.

202 2.2. The RS: a Ranking Selection operator

226 2.3. The DBX: a Direction-Based Crossover operator

239 Group B(the group with lower rankings):

250 where sc;i is the crossover step-size given by

258 ~ is dynamically adjusted by

336 2.5. The replacement operation

345 2.6. A simple scheme for alleviating stagnation

356 2.7. The stop criteria

361 2.8. A summary of the proposed RGA-RDD

378 3.1. The designate experiments and the benchmark functions

Fig. 6. The pseudo-code of the proposed RGA-RDD optimization scheme.

Set no. Parameter setting

416 3.3. The role of the RS operator

Set no. Optimized f(x)

448 3.4. Parametric studies and a parameter selection guideline

Set no. p Optimized f(x)

Set no. k Optimized f(x)

Set no. /0 Optimized f(x)

Set No. b Optimized f(x)

Set no. e Optimized f(x)

Exp. Parameters gdB RgdB

Problem Main effect D on response of gdB

Fig. 8. Response surface plots of the algorithm parameters.

526 4. Benchmark study of The RGA-RDD

536 4.1. The properties of the benchmark Testbed

No. Function name Properties

547 4.2. The evaluation criteria

568 4.3. Algorithms setup

581 4.4. Results and discussion

582 4.4.1. The performance evaluation on CEC2005 testbed

610 4.4.2. The performance evaluation of paired benchmark functions

625 4.4.3. Further performance evaluation of multimodal and hybrid functions

N/A indicates the data are not available in the literature.

687 6. Uncited references

695 (1) Sphere unimodal function

699 (2) Ackley multimodal function

703 where z x o; x ¼ ½x1 ; x2 ; . . . ; xn , and o ¼ ½o1 ; o2 ; . . . ; on is the shifted global optimum.

You might also like