G Model

ASOC 2618 1–20 ARTICLE IN PRESS

Applied Soft Computing xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

1 Water cycle algorithm for solving constrained multi-objective

2 optimization problems
3 Q1 Ali Sadollah a , Hadi Eskandar b , Joong Hoon Kim a,∗
a
4 School of Civil, Environmental and Architectural Engineering, Korea University, 136-713 Seoul, South Korea
b
5 Faculty of Mechanical Engineering, University of Semnan, Semnan, Iran
6

7
23 a r t i c l e i n f o a b s t r a c t
8
9 Article history: In this paper, a metaheuristic optimizer, the multi-objective water cycle algorithm (MOWCA), is presented
10 Received 14 October 2013 for solving constrained multi-objective problems. The MOWCA is based on emulation of the water cycle
11 Received in revised form process in nature. In this study, a set of non-dominated solutions obtained by the proposed algorithm is
12 11 September 2014
kept in an archive to be used to display the exploratory capability of the MOWCA as compared to other
13 Accepted 10 October 2014
efficient methods in the literature. Moreover, to make a comprehensive assessment about the robustness
14 Available online xxx
and efficiency of the proposed algorithm, the obtained optimization results are also compared with other
15
widely used optimizers for constrained and engineering design problems. The comparisons are carried
16 Keywords:
17 Multi-objective optimization
out using tabular, descriptive, and graphical presentations.
18 Water cycle algorithm © 2014 Published by Elsevier B.V.
19 Pareto optimal solutions
20 Benchmark function
21 Metaheuristics
22 Constrained optimization

24 1. Introduction based multi-objective simulated annealing with self-stopping cri- 45

terion (PDMOSA-I) [4], the vector immune algorithm (VIS) [12], the 46

25 In recent decades, solving real-world engineering design and elitist-mutation multi-objective particle swarm optimization (EM- 47

26 resource-optimization problems via multi-objective evolutionary MOPSO) [13], the weight-based multi-objective immune algorithm 48

27 algorithms (MOEAs) has become an attractive research area for (WBMOIA) [14], the orthogonal simulated annealing (OSA) [15], 49

28 many scientists and researchers [1]. Many optimization methods and the hybrid quantum immune algorithm (HQIA) [16]. 50

29 have been developed to deal with these kinds of problems [2,3]. Recently, some researchers have expressed enthusiasm 51

30 In contrast to single-optimization problems, the main goal of regarding immune-system algorithms for solving different types 52

31 evolutionary algorithms in multi-objective optimization problems of MOPs. In fact, many researchers have attempted to boost and 53

32 (MOPs) is to find a set of best solutions, so-called non-dominated amend the main characteristics of immune algorithms to increase 54

33 solutions or Pareto-optimal solutions. In addition, non-dominated the efficiency and convergence speed of these methods for solving 55

34 solutions obtained by different evolutionary algorithms are one of MOPs. 56

35 the most common ways to clarify and assess the robustness and Representatives of immune-based algorithms include the 57

36 capabilities of a proposed algorithm. immune forgetting multi-objective optimization algorithm 58

37 In this situation, metaheuristic methods as a component of (IFMOA) suggested by Zhang et al. [17], the immune dominance 59

38 evolutionary algorithms have been significant owing to their clonal multi-objective algorithm (IDCMA) developed by Jiao 60

39 fast convergence rate and accuracy [4]. Some of these methods et al. [18], and the adaptive clonal selection algorithm for multi- 61

40 include the strength Pareto evolutionary algorithm (SPEA) [5], objective optimization (ACSAMO) proposed by Wang and Mahfouf 62

41 SPEA2 [6], the Pareto archive evolution strategy (PAES) [7], the [19]. 63

42 micro-genetic algorithm (micro-GA) [8], the non-dominated sor- Furthermore, many studies prefer to combine metaheuristic 64

43 ting genetic algorithm (NSGA) [9], NSGA-II [10], the multi-objective methods to take advantage of the predominant features of multi- 65

44 particle swarm optimization (MOPSO) [11], the Pareto dominant ple methods simultaneously. These approaches are so-called hybrid 66

techniques. There have been many researchers in the past who have 67

tried to use this idea to handle MOPs. 68

Q2 ∗ Corresponding author. Tel.: +82 02 290 3316; fax: +82 232903316. For instance, Kaveh and Laknejadi [20] introduced the novel 69

E-mail address: [email protected] (J.H. Kim). hybrid charge system search and particle swarm multi-objective 70

http://dx.doi.org/10.1016/j.asoc.2014.10.042
1568-4946/© 2014 Published by Elsevier B.V.

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
2 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

71 optimization method (CSS-MOPSO). This multi-objective optimizer

72 is a hybridization of particle swarm optimization (PSO) and the
73 charged-system search method [20]. Another approach, recently
74 proposed by Narimani et al. [21], is called the HMPSO-SFLA method.
75 This hybrid optimization algorithm is based on the concepts of PSO
76 and the shuffle frog-leaping algorithms (SFLA) [21].
77 Deferential evolutionary for multi-objective optimization with
78 local search based on rough set theory (DEMORS) is another hybrid
79 method, presented by Coello et al. [22]. The DEMORS method has
80 also been used to solve constrained MOPs (CMOPs). Looking at the
81 mentioned algorithms, we can notice that the majority of these
82 approaches are classified as population-based methods.
83 Hence, there is enough proof to support the idea that
84 population-based algorithms are the most common way to solving
85 MOPs, primarily because this subject is linked to the characteristics
86 and potentials of these methods. In other words, these methods are
87 capable of handling both continuous and combinatorial optimiza- Fig. 1. Optimal Pareto solutions (A and B) for the 2D domain.
88 tion problems having high accuracy and satisfactory convergence
89 speed to the Pareto-optimal solutions [4].
can be considered as a non-dominated solution if and only if the 131
90 In this research, a recently developed population-based algo-
following conditions become satisfied by the solution as given: 132
91 rithm, the multi-objective water cycle algorithm (MOWCA), is used
92 to tackle CMOPs. The proposed algorithm was first presented by
(a) Pareto dominance: U = (u1 , u2 , u3 , . . ., un ) < V = (v1 , v2 , v3 , . . ., vn ) 133
93 Eskandar et al. [23] for ordinary optimization problems. The basic
if and only if U is partially less than V in the objective space, as 134
94 idea of the WCA is inspired by the real-world water cycle process
follows: 135
95 in nature, including the motion of rivers to the sea. The MOWCA 
96 algorithm is evaluated here by solving a set of engineering design fi (U) ≤ fi (V ) ∀i
97 problems and CMOPs, and the final optimization results obtained i = 1, 2, 3, . . ., n, (3) 136

98 by the MOWCA are compared with those of other metaheuristic fi (U) < fi (V ) ∃i
99 algorithms in the literature.
where n is the number of objective functions. 137
100 The remaining of the present paper is organized as follows. In
(b) Pareto optimal solution: vector U is said to be a Pareto optimal 138
101 Section 2, the definition of standard MOPs is given, and the perfor-
solution if and only if any other solutions cannot be determined 139
102 mance criteria used for quantitative assessments are described. In
to dominate U. A set of Pareto optimal solutions is called a Pareto 140
103 Section 3, a short description of the WCA, the definition of MOWCA,
optimal front (PFoptimal ). 141
104 and the concept of MOWCA are introduced in detail. Numerical
105 examples and benchmark functions considered in this paper are
Fig. 1 gives an overview of the concept of non-dominated solu- 142
106 provided in Section 4, along with their results and discussion.
tions in MOPs. It can be seen from Fig. 1 that among three solutions 143
107 Finally, conclusions are drawn in Section 5.
A, B, and C, solution C has the highest values for f1 and f2 . This 144

means that this solution is a solution dominated by solutions A 145

108 2. Multi-objective problems
and B. In contrast, both solutions A and B can be considered as 146

non-dominated solutions, as neither of them dominates each other. 147

109 Multi-objective optimization problems (MOPs) can be defined
110 as optimization problems for which at least two objective functions
2.1. Performance metric parameters 148
111 are to be optimized simultaneously. Mathematically, a MOP can be
112 formulated as follows:
To make fair quantitative evaluations and judgments among 149
T
113 F(X) = [f1 (X), f2 (X), . . ., fm (X)] , (1) different types of MOEAs, three performance parameters that are 150

widely used to evaluate the performance of metaheuristic algo- 151

114 where X = [x1 , x2 , x3 , . . ., xd ] is a vector of design variables (d is the
rithms are investigated in this paper. These criteria are defined in 152
115 number of design variables). One initial approach for solving MOPs
detail in the following subsections. 153
116 is to use weight factors to convert a MOP into a single-optimization
117 problem [24]. This technique can be formulated based on the fol-
2.1.1. Generational distance metric 154
118 lowing equation:
The generational distance (GD) metric was first presented by 155


N Veldhuizen and Lamont [26]. The main objective of this criterion is 156
119 F= wn fn , (2) to clarify the capability of the different algorithms of finding a set 157

n=1 of non-dominated solutions having the lowest distance with the 158

Pareto optimal fronts (PFoptimal ). 159

120 where N is the number of objective functions and wn and fn are
Based on this definition, it can be understood that the algorithm 160
121 weighting factors and objective functions, respectively.
with the minimum GD has the best convergence to PFoptimal . This 161
122 It is worth mentioning that single-optimization problems have
evaluation factor is defined in mathematical form as can be seen in 162
123 just one point as the optimal solution. Hence, in order to find a set of
the following equations [27]: 163
124 solutions, Eq. (2) has to be solved by using a wide variety of weight
125 factors; this is extremely time consuming and must be taken into  npf 1/2
126 serious consideration as a major downside of this method. 1 
GD = di2 , (4) 164
127 In contrast, the most common way to solve MOPs is by keeping a npf
i=1
128 set of best solutions in an archive and updating the archive at each
129 iteration. In this approach, the best solutions are defined as non- where npf is the number of members in the generated Pareto front 165

130 dominated solutions or Pareto optimal solutions [25]. A solution (PFg ), and di is the Euclidean distance between member i in PFg and 166

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 3

Fig. 3. Schematic view of the S metric for MOPs.

Fig. 2. Schematic view of the GD criterion for MOPs.

distance between each point in PFg and the closest point in PFoptimal . 201

167 the nearest member in PFoptimal . Meanwhile, the Euclidean distance npf and d̄ are defined as the total number of members in PFg and the 202

168 (d) is calculated based on the following equation: average of all distances, respectively. 203

 1/2 Based on Eq. (7), it can be easily inferred that the value of the  204


n
2
metric is always greater than zero, and a small value of  means 205
169 d(p, q) = d(q, p) = (fiq − fip ) , (5) better distribution and spread of the solutions. In this condition, 206

i=1  = 0 is the perfect condition indicating that extreme solutions of 207

PFoptimal have been found and that di = d̄ for all non-dominated 208
170 where q = (f1q , f2q , f3q , . . ., fnq ) is a point on PFg , and P = (f1p , f2p ,
points. Fig. 4 shows a schematic view of the  performance metric 209
171 f3p , . . ., fnp ) is the nearest member to q in PFoptimal . Fig. 2 shows a
for a given Pareto front. 210
172 schematic view of this performance metric for the 2D space. The
173 best obtained value for the GD metric is equal to zero which corre-
174 sponds to PFg exactly covers the PFoptimal .
3. Multi-objective water cycle algorithm 211

175 2.1.2. Metric of spacing

3.1. Water cycle algorithm 212
176 The metric of spacing (S) was suggested by Scott [29]. The main
177 goal of this factor is to show the distribution of non-dominated
The water cycle algorithm (WCA) mimics the flow of rivers 213
178 solutions obtained by a specified algorithm. This criterion can show
and streams toward the sea and derived by the observation of 214
179 how well the obtained solutions are distributed among each other.
water cycle process. Let us assume that there are some rain or pre- 215
180 The mathematical definition of this parameter is given in the fol-
cipitation phenomena. An initial population of designs variables 216
181 lowing equation [30]:
(population of streams) is randomly generated after raining pro-
 217

npf cess. The best individual (i.e., the best stream), classified in terms
1 
218
2 of having the minimum cost function, is chosen as the sea [23].
182 S= (di − d̄) , (6) 219
npf − 1 Then, a number of good streams (i.e., cost function values close 220
i=1
to the current best record) are chosen as rivers, while all other 221

183 where npf is number of member in PFg and di is the Euclidean dis- streams flow to the rivers and sea. In an N dimensional optimization 222

184 tance between member ith in PFg and nearest member in PFg . The
185 smallest value of S gives the best uniform distribution in PFg . If all
186 non-dominated solutions are uniformly distributed in the PFg , then,
187 the values of di and d̄ are the same, therefore, the value of S metric
188 equals to zero. Fig. 3 shows a schematic view of the spacing metric
189 used in this paper.

190 2.1.3. Metric of spread

191 The third performance metric parameter is the spread metric
192 () proposed by Deb [30]. This assessment parameter has been
193 defined to determine the extent of spread attained by the non-
194 dominated solutions obtained from a specified algorithm. To be
195 more precise, this criterion can analyze how the solutions are
196 extended across PFoptimal . Hence, the  metric can be formulated
197 as follows [30]:
npf
df + dl + i=1
|di − d̄|
198 = , (7)
df + dl + (npf − 1)d̄

199 where df and dl are the Euclidean distances between the extreme
200 solutions in PFoptimal and PFg , respectively. Further, di is the Euclidean Fig. 4. Schematic view of the  metric [10].

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
4 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

223 problem, a stream is an array of 1 × N. This array is defined as fol- population): 251

224 lows: ⎡ ⎤
Stream1
⎢ ⎥
⎢ Stream2 ⎥
225 A stream candidate = [x1 , x2 , x3 , . . ., xN ], (8) ⎢ ⎥
⎢ ⎥
Population of Streams = ⎢ Stream3 ⎥ 252
⎢ ⎥
⎢ .. ⎥
226 where N is the number of design variables (problem dimension). To ⎣ . ⎦
227 start the optimization algorithm, an initial population representing
228 a matrix of streams of size Npop × N is generated. Hence, the matrix StreamNStream
229 of initial population, which is generated randomly, is given as (rows ⎡ ⎤
x11 x21 x31 ··· 1
xN
and column are the number of population and the number of design
230
⎢ ⎥
variables, respectively): ⎢ x12 x22 x32 ··· 2
xN ⎥
⎢ ⎥
231

=⎢ ⎥. 253

⎡ ⎤ ⎢ .. .. .. .. .. ⎥
Sea ⎣ . . . . . ⎦
N N N N
⎢ River1 ⎥ x1 Stream x2 Stream x3 Stream ··· xNStream
⎢ ⎥
⎢ ⎥ (13) 254
⎢ River2 ⎥
⎢ ⎥
⎢ River ⎥ 255
⎢ 3 ⎥ Depending on flow magnitude, each river absorbs water from
⎢ ⎥ 256

⎢ .. ⎥ streams. The amount of water entering a river and/or the sea, hence, 257
⎢ ⎥
⎢ . ⎥ varies from stream to stream. In addition, rivers flow to the sea 258
232 Total population = ⎢ ⎥ which is the most downhill location. The designated streams for
⎢ StreamNsr+1 ⎥ 259
⎢ ⎥ each rivers and sea are calculated using the following equation [23]:
⎢ Stream ⎥   
260

⎢ Nsr+2 ⎥
 
⎢ ⎥  Costn 
⎢ StreamNsr+3 ⎥ NSn = round   × NStream , n = 1, 2, . . ., Nsr ,
⎢ ⎥ 
(14) 261
Costi 
Nsr
⎢ ⎥
⎢ .. ⎥ i=1
⎣ . ⎦
where NSn is the number of streams which flow to the specific rivers 262
StreamNpop and sea. Based on the assumed concept for the WCA, rivers and sea 263
⎡ ⎤ absorb streams as part of movement of individuals toward best 264
x11 x21 x31 ··· 1
xN
(i.e., sea) and better (i.e., rivers) solutions. In fact, sea and rivers
⎢ ⎥
265

⎢ x12 x22 x32 ··· 2

xN ⎥ designate streams based on their intensity of flow (i.e., fitness value 266
⎢ ⎥
233 =⎢ ⎥, (9) or objective function value). 267
⎢ .. .. .. .. .. ⎥ Therefore, as sea is the best solution for the current population,
⎣ . . . . . ⎦
268

N N N N
based on Eq. (14), it absorbs (assigns) more streams than rivers 269
x1 pop x2 pop x3 pop ··· xNpop do. If the problem is minimization problem (cost function), more 270

streams flow to sea which has the lowest cost, and other streams 271

flow to rivers which have lower costs. 272

234 where Npop and N are the total number of population and the num-
For instance, if we have 50 streams (population size), one sea and 273
235 ber of design variables, respectively. Each of the decision variable
4 rivers, we can assume that 22 streams flow to sea (the lowest cost), 274
236 values (x1 , x2 , . . ., xN ) can be represented as floating point number
13 streams flow to the first best river, 7 streams flow to the second 275
237 (real values) or as a predefined set for continuous and discrete prob-
best river, 5 streams flow to the third best river, and 3 streams flow 276
238 lems, respectively. The cost of a stream is obtained by the evaluation
to the forth best river. The reason for using absolute value used 277
239 of cost function (C) given as follows:
in Eq. (14) is that to eliminate the effect of negative sign. Also, the 278

reason for using round operator is that we cannot assign 1.5 streams 279

240 Ci = Costi = f (x1i , x2i , . . ., xN

i
) i = 1, 2, 3, . . ., Npop . (10) to a river or sea. The numbers should be integer values. 280

As it happens in nature, streams are created from the raindrops 281

and join each other to generate new rivers. Some stream may even 282

241 At the first step, Npop streams are created. A number of Nsr from flow directly to the sea. All rivers and streams end up in the sea that 283

242 the best individuals (minimum values) are selected as a sea and corresponds to the current best record. 284

243 rivers. The stream which has the minimum value among others is Let us assume that there are Npop streams of which Nsr − 1 are 285

244 considered as the sea. In fact, Nsr is the summation of number of selected as rivers and 1 is selected as the sea. Fig. 5a shows the 286

245 rivers (which is a user parameter) and a single sea as given in Eq. schematic view of a stream flowing toward a specific river along 287

246 (11). The rest of the population (streams which flow to the rivers or their connecting line. 288

247 may directly flow to the sea) is calculated using Eq. (12) as follows: The distance X between the stream and the river may be ran- 289

domly updated as the following relation: 290

Nsr = Number of Rivers + 1 , (11) X ∈ (0, C × d), C > 1, (15)


248 291

Sea where 1 < C < 2 and the best value for C may be chosen as 2; d is the 292

current distance between stream and river. The value of X in rela- 293

249 NStream = Npop − Nsr . (12) tion (15) corresponds to a random number (uniformly distributed 294

or determined from any appropriate distribution) between 0 and 295

(C × d). Setting C > 1 allows streams to flow in different directions 296

250 Eq. (13) shows the population of streams who flow to the rivers toward rivers. This concept may also be used to describe rivers flow- 297

or sea. Indeed, Eq. (13) is part of Eq. (9) (i.e., total individual in ing to the sea. Therefore, as the exploitation phase in the WCA, the 298

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 5

where dmax is a small number close to zero. After evaporation, 321

the raining process is applied and new streams are formed in the 322

different locations (similar to mutation in the GAs). Indeed, the 323

evaporation operator is responsible for the exploration phase in the 324

WCA. The following equation is used to specify the new locations 325

of the newly formed streams: 326

 new = LB
X  + rand × (U B
 − LB),
 (19) 327
Stream

where LB and UB are lower and upper bounds defined by the given 328

problem, respectively. Similarly, the best newly formed stream is 329

considered as a river flowing to the sea. The rest of new streams are 330

assumed to flow into the rivers or may directly flow into the sea. 331

A large value for dmax prevents extra searches and small values 332

encourage the search intensity near the sea. Therefore, dmax con- 333

trols the search intensity near the sea. The value of dmax adaptively 334

decreases as follows: 335

i
dmax
i+1 i
dmax = dmax − . (20) 336
Max Iteration
The development of the WCA optimization process is illustrated 337

by Fig. 5b where circles, stars, and the diamond correspond to 338

streams, rivers, and sea, respectively. The white (empty) shapes 339

denote the new positions taken by streams and rivers. In addition, 340

Table 1 shows the pseudo-code and step by step processes of the 341

WCA in detail. 342

3.1.1. PSO versus WCA 343

In this subsection, similarity and differences of WCA with the 344

particle swarm optimization (PSO) [31] are highlighted. Indeed, 345

every metaheuristic algorithm has its own approach and method- 346

ology in finding global optimum solution. As a similarity between 347

the WCA and PSO, we can say that all methods are categorized as 348

population-based metaheuristic algorithms; population of parti- 349

cles in the PSO and population of streams in the WCA. 350

Except of this similarity, their concepts, parameters and oper- 351

Fig. 5. (a) Schematic description of the stream’s flow to a specific river and (b) ators are different with each other. The PSO’s concept is based on 352
schematic of the WCA optimization process.
the movement of particles (e.g., fishes, birds, etc.) and their per- 353

sonal and best individual experiences [31]. The WCA’s notions are 354
299 new position for streams and rivers may be given in the following derived by the water cycle process in nature and the observation 355
300 equations [23]: of how streams and rivers flow to the sea. 356

 i+1 = X
i i i The updating formulations for the positions of rivers and 357
301 XStream Stream + rand × C × (XRiver − XStream ), (16)
streams differ from the updating formulations used in the PSO. The 358

i
 i+1 = X i i WCA does not use the concept of moving directly to the best solu-
X Stream + rand × C × (XSea − XStream ), (17)
302 359
Stream
tion (global best) as used in the PSO. In fact, the WCA utilizes the 360
303 X i
 i+1 = X i − X
+ rand × C × (X i ), (18) concept of moving indirectly from streams to the rivers and from 361
River River Sea River
rivers to the sea (i.e., the best temporal obtained optimum solution). 362
304 where rand is an uniformly distributed random number between
In the WCA, rivers (a number of best selected solutions except Q3 363
305 0 and 1. Eqs. (16) and (17) are for streams which flow to their cor-
the best one (sea), (Eq. (11)) act as guidance points for guiding other 364
306 responding rivers and sea, respectively. Notations having vector
individuals in the population (streams) toward better positions (see 365
307 sign correspond to vector values, otherwise the rest of notations
Fig. 5b) and to avoid the search in inappropriate regions in near- 366
308 and parameters are considered as scalar values. If the solution
optimum solutions (see Eq. (16)). 367
309 given by a stream is better than its connecting river, the pos-
It is worth pointing out that rivers, themselves, move toward 368
310 itions of river and stream are exchanged (i.e., the stream becomes
the sea (i.e., best obtained solution). They are not fixed points (see 369
311 a river and the river becomes a stream). A similar exchange can
Eq. (18)). In fact, this procedure (moving streams to the rivers and, 370
312 be done for a river and the sea. The evaporation process operator
then moving rivers to the sea) leads to indirect movements toward 371
313 also is introduced to avoid premature convergence to local optima
the best solution by the WCA. 372
314 (exploitation phase). Basically, evaporation causes sea water to
On other hand, in the PSO, individuals (particles) based on their 373
315 evaporate as rivers/streams flow to the sea. This leads to new pre-
personal and best experiences attempt to find the best solution as 374
316 cipitations. Therefore, we have to check if the river/stream is close
the searching approach is moving directly toward the best optimal 375
317 enough to the sea to make the evaporation process occur. For that
solution [31]. In addition, in the WCA, a number of near-best to best 376
318 purpose, the following criterion is utilized for evaporation condi-
319 selected solutions (rivers + sea) attract other individuals of popula- 377
tion:
  tion (streams) based on their goodness of the function values (i.e., 378
if X Sea
i
i − X
River
 < dmax or rand < 0.1 i = 1, 2, 3, . . ., Nsr − 1 intensity of flow) using Eq. (14). However, in the classical PSO, this 379

, process is not used. 380

320 Perform raining process using Eq. (19)
Another difference between the WCA and PSO is the existence 381
end of evaporation condition and raining process in the WCA which 382

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
6 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

Table 1
Pseudo-code of the WCA.

• Set user parameter of the WCA: Npop, Nsr, dmax, and Maximum_Iteration.
• Determine the number of streams (individuals) which flow to the rivers and sea
using Eqs. (11) and (12).
• Create randomly initial population.
• Define the intensity of flow (How many streams flow to their corresponding
rivers and sea) using Eq. (14).
while (t < Maximum_Iteration) or (any stopping condition)
for i = 1 : Population Size (Npop)
Stream flows to its corresponding rivers and sea using Eqs. (16) and (17)
Calculate the objective function of the generated stream
if F_New_Stream < F_river
River = New_Stream;
if F_New_Stream < F_Sea
Sea = New_Stream;
end if
end if
River flows to the sea using Eq. (18)
Calculate the objective function of the generated river
if F_New_River < F_Sea
Sea = New_River;
end if
end for
for i = 1 : number of rivers (Nsr)
if (distance (Sea and River) < dmax) or (rand < 0.1)
New streams are created using Eq. (19)
end if
end for
Reduce the dmax using Eq. (20)
end while
Postprocess results and visualization

383 corresponds to the exploration phase. The evaporation condition average side length of the cuboid, for point i [10]. A lower value for 410

384 and raining process provide an escape mechanism for the WCA to the crowding-distance indicates greater distribution of the solu- 411

385 avoid getting trapped in local optimum solutions, while in the PSO, tions in a specific region. In MOPs, this parameter is calculated in 412

386 the exploration mechanism (formulation) is different. objective spaces. Hence, to compute this parameter for each non- 413

387 In the PSO, inertia weight (w) (i.e., a user parameter) in the dominated solution, all non-dominated solutions must be sorted in 414

388 updating equation (movement equation) is responsible for the terms of values for one of the objective functions. 415

389 exploration phase and reduces at each iteration [31]. Table 2 sum- Selection of the sea and rivers from the obtained population 416

390 marizes the differences of two reported optimizers in terms of as the best-guide solution for other solutions at each iteration is 417

391 applied strategies. a vital step in the MOWCA. It affects the convergence capability 418

of the algorithm and maintains a good spread of non-dominated 419

392 3.2. Proposed MOWCA solutions. Therefore, for all iterations, the crowding-distance for all 420

393 In order to convert the WCA to an efficient multi-objective

394 optimization algorithm, it is crucially important to define the pre-
395 dominant features of the algorithm (i.e., sea and rivers) correctly.
396 In ordinary optimization problems for the WCA, only one objective
397 function needs to be minimized and in this condition, a number of
398 best obtained solutions in the population are selected as a sea (best
399 obtained solution so far) and rivers.
400 Nevertheless, in MOPs, there is more than one function to be
401 minimized (maximized). Therefore, the definition in the WCA for
402 selecting sea and rivers has to be changed in the multi-objective
403 space. To select the most efficient (best) solutions in the population
404 as a sea and rivers, a crowding-distance mechanism is used. The
405 concept of crowding-distance mechanism was first defined by Deb
406 et al. [10].
407 This parameter is a criterion for showing the distribution of
408 non-dominated solutions around a particular non-dominated solu-
409 tion. Fig. 6 illustrates how to calculate the crowding-distance, the Fig. 6. Schematic view of crowding-distance calculation.

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 7

Table 2
Differences between two optimization methods in terms of their approaches for finding global optimum solution.

Strategy PSO WCA

Population Particles (e.g., fishes, birds) Streams (including sea and rivers)

User parameter - Npopa - Npop

- w (Inertia weight) - Nsr (Number of rivers + sea)
- c1 (Personal learning constant) - dmax (Maximum allowable distance between
river and sea)
- c2 (Global learning constant)
- Vmax (Maximum velocity)

Global search Inertia weight (first term of the movement equation): Evaporation condition:
w × vti i
if norm(xSea − xRiver
i
) < dmax
Perform raining based on Eq. (19)
end if

Local search The second and third terms of the movement equation: Moving streams to the rivers and rivers to the
 i ) + c2 r2 (gBesti − X
c 1 r1 (pBesti − X i ) streams (Eqs. (16)–(18))

Selection if Fb New < F Old if F Stream < F River

Accept New Particle River = Stream;
if F New < F Best if F Stream < F Sea
Accept New Particle Stream = Sea;
end if end if
end if end if
if F River < F Sea
River = Sea;
end if
a
Npop , total number of population.
b
F, value of objective function.

421 non-dominated solutions must be calculated to determine which potential of the MOWCA to search a wide range of design space, 448

422 solutions have the highest crowding-distance values. while concentrating to near optimum non-dominated solutions. 449

423 Afterwards, the obtained non-dominated solutions are desig-

424 nated as a sea and rivers and, furthermore, the intensity of flow for 3.3. Constraint-handling approach 450
425 rivers and sea is calculated based on the crowding-distance values.
426 In this situation, some non-dominated solutions will most likely Many MOPs are subject to a set of constraints (e.g., inequal- 451
427 be created around sea and rivers at the next iteration, and their ity, equality, linear, nonlinear, and the like). Therefore, it is vital 452
428 crowding-distance values will be amended and reduced. to find good, simple strategies to handle the imposed constraints 453
429 Moreover, it is very important to save non-dominated solu- and detect solutions in the feasible area. Hence, a simple approach 454
430 tions in an archive to generate the Pareto front sets. This archive is defined here to apply on the MOWCA. 455
431 is updated at each iteration, and any dominated solutions are In this mechanism, after generating a set of solutions at each 456
432 eliminated from the archive. Based on the literature and due to iteration, all constraints are checked and some solutions that are in 457
433 having fair comparison, the size of Pareto archive (number of non- the feasible area are selected. Afterward, non-dominated solutions 458
434 dominated solutions in the archive) is set to be the same with other are selected from the feasible solutions and inserted into the Pareto 459
435 reported optimizers (see Table 3). Therefore, whenever the number archive. Finally, sea and rivers are chosen from this archive for the 460
436 of members in the Pareto archive becomes greater than the Pareto next iteration based on the strategy mentioned in Section 3.2. 461
437 archive size, the crowding-distance is applied again to eliminate as
438 many as necessary of the non-dominated solutions having the low- 3.4. Steps of the MOWCA 462
439 est crowding-distance values among the Pareto archive members.
440 The MOWCA has the great potential for exploitation in the The steps of the MOWCA are summarized as follows: 463
441 design space as it focuses in the near optimum solutions and
442 exploits better solutions. Indeed, the exploration and exploitation Step 1: Choose the initial parameters for the MOWCA: Nsr , dmax , 464
443 phases in the MOWCA are combined and mixed. To clarify fur- Npop , Max Iteration, and Pareto archive size. 465
444 ther, mostly the MOWCA starts with exploitation approach, moving Step 2: Generate random initial population and form the initial 466
445 streams toward rivers and rivers toward the sea. In between these streams, rivers, and sea using Eqs. (9), (11), and (12). 467
446 movements, the MOWCA applies evaporation condition in order Step 3: Calculate the value of multi-objective functions for each 468
447 to perform exploration phase. This trend can be considered as the stream using Eq. (10). 469

Table 3
Parameters used for optimization of mathematical and engineering MOPs.

Constrained problems Engineering design problems

NFEs Archive size NFEs Archive size

CONSTR 25,000 100 4-bar truss 8000 100

BNH 25,000 100 Speed reducer 24,000 100
KITA 5000 100 Gear train 10,000 100
TNK 25,000 100 Spring 10,000 100
OSY 40,000 200 Welded beam 15,000 100
SRN 25,000 100 Disk brake 5000 100

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
8 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

470 Step 4: Determine the non-dominated solutions in the initial pop- In addition, in order to have fair and reliable comparisons with 533

471 ulation and save them in the Pareto archive. other methods, the number of function evaluations (NFEs) used 534

472 Step 5: Determine the non-dominated solutions among the feasi- and the Pareto archive size for all MOPs are the same as in other 535

473 ble solutions and save them in the Pareto archive. previous studies. In fact, the maximum number of NFEs is taken as 536

474 Step 6: Calculate the crowding-distance for each Pareto archive the stopping condition, similarly to other the methods in this paper. 537

475 member. In the literature, researchers utilized one (i.e., constrained 538

476 Step 7: Select a sea and rivers based on the crowding-distance problem 5) or two (i.e., constrained problem 2, four-bar truss, 539

477 value. speed-reducer, disk brake, and welded beam design) performance 540

478 Step 8: Determine the intensity of the flow for rivers and sea based metrics for their examples using their optimization methods. In 541

479 on the crowding distance values using Eq. (14). some cases, they used all three performance metrics for constrained 542

480 Step 9: Streams flow into the rivers using Eq. (16). problems 1, 3, 4, and 6. 543

481 Step 10: Exchange positions of river with a stream which gives the Therefore, in order to have comparative study with other 544

482 best solution. optimization methods, we applied the MOWCA with existing per- 545

483 Step 11: Some streams may directly flow into the sea using Eq. formance metrics for each example and compared with the results 546

484 (17). obtained by other researchers in the literature. 547

485 Step 12: Exchange positions of sea with a stream which gives the These user parameters for all the MOPs are shown in Table 3. The 548

486 best solution. chosen values for the Pareto archive size are extracted from litera- 549

487 Step 13: Rivers flow into the sea using Eq. (18). ture [12–14,20] due to have fair comparisons with other optimizers 550

488 Step 14: Exchange positions of sea with a river which gives the in this study. 551

489 best solution.

490 Step 15: Check the evaporation condition using the pseudo-code 4.1. Multi-objective benchmark problems 552
491 in Section 3.1.
492 Step 16: If the evaporation condition is satisfied, the raining pro- In this subsection, twelve benchmark CMOPs and engineering 553
493 cess will occur using Eq. (19). design problems are used to assess the capability and efficiency 554
494 Step 17: Reduce the value of dmax which is a user defined parameter of the proposed algorithm for handling multifaceted mathematical 555
495 using Eq. (20). problems. 556
496 Step 18: Determine the new feasible solutions in the population.
497 Step 19: Determine the new non-dominated solutions among the 4.1.1. Constrained problem 1: CONSTR 557
498 feasible solutions and save them in the Pareto archive. CONSTR is a mathematical problem having two design variables. 558
499 Step 20: Eliminate any dominated solutions in the Pareto archive. This problem is defined as follows [30]: 559
500 Step 21: If the number of member in the Pareto archive is more ⎧
501 than the determined Pareto archive size, go to the Step 22, other- ⎨ f1 (x) = x1
502 wise, go to the Step 23. minimize , (21)
⎩ f1 (x) = 1 + x2
560
503 Step 22: Calculate the crowding-distance value for each Pareto
x1
504 archive member and remove as many members as necessary with
505 the lowest crowding-distance value. subject to 561
506 Step 23: Calculate the crowding-distance value for each Pareto
507 archive member to select new sea and rivers. g1 (x) = 6 − (x2 + 9x1 ) ≤ 0
508 Step 24: Check the convergence criteria. If the stopping criterion g2 (x) = 1 + x2 − 9x1 ≤ 0
509 is satisfied, the algorithm will be stopped, otherwise return to the . (22) 562

510 Step 9. 0.1 ≤ x1 ≤ 1

0 ≤ x2 ≤ 5

511 4. Numerical examples and results The Pareto optimal front for this problem is convex and defined 563

as follows [30]: 564

512 In this section, a set of well-known constrained and engineering

513 design MOPs are used to clarify the performance of the proposed For 0.39 ≤ x1∗ ≤ 0.67 : x2∗ = 6 − 9x1∗ , (23) 565

514 MOWCA. These problems are various types of constrained multi- For 0.67 ≤ x1∗ ≤1: x2∗ = 0. (24) 566
515 objective problems (CMOPs) with diverse features that have been
516 selected from a set of credible research studies. Moreover, the The final statistical optimization results obtained by the 567

517 MOWCA as a recently developed optimizer is compared with other reported methods for the performance measurements are shown 568

518 prominent methods employed in previous studies, such as the in Table 4. As can be seen from Table 4, it can be inferred that the 569

519 NSGA-II, MOPSO, CMOPSO, EM-MOPSO, MOCSO, Micro-GA, PAES, MOWCA could find the best mean solutions in terms of the GD value 570

520 VIS, WBMOIA, and HQIA. having the minimum distance from PFoptimal . 571

521 The MOWCA was coded in MATLAB programming software, and Similarly, the non-dominated solutions obtained by the 572

522 the task of optimizing each test function was executed using 30 MOWCA have the best distribution and spread, which can be 573

523 independent runs. For all benchmark problems, the initial param- inferred from the mean values of the S and  metrics (see Table 4). It 574

524 eters for the MOWCA (Ntotal , Nsr , and dmax ) were selected as 50, 8, is worth mentioning that the MOWCA shows about 68% reduction 575

525 and 1E−5, respectively. in the  value compared with the value obtained by the NSGA- 576

526 The natures of the mentioned problems include various types II, which is in the second rank in terms of the  value after the 577

527 of objective functions having different numbers of design vari- MOWCA. 578

528 ables and nonlinear constraints. In addition, some of the problems Therefore, the proposed MOWCA shows superiority over other 579

529 considered, such as gear train and spring design problems, are algorithms for this problem, especially for the  metric. These 580

530 categorized as discrete optimization problems (combinatorial opti- observations from Table 4 will be completed by the results pre- 581

531 mization problems). In this study, both continuous, discrete, and sented in Fig. 7a. Looking at Fig. 7a, one can conclude that the 582

532 integer MOPs are investigated using the MOWCA. MOWCA was successful in covering the optimal Pareto front, having 583

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 9

Fig. 7. Comparisons of the optimal Pareto fronts and generated Pareto front and distribution of non-dominated solutions using the MOWCA for the: (a) CONSTR, (b) TNK, (c)
KITA, (d) SRN, (e) BNH, and (f) OSY (solid lines and cross points represent the optimal and generated (obtained) Pareto front, respectively).

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
10 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

Table 4
Comparison of several algorithms based on the mean and standard deviation (SD) values of performance metrics for the CONSTR problem. “N/A” stands for not available.

Methods GD  S

Mean SD Mean SD Mean SD

NSGA-II 5.1349E−03 2.4753E−04 0.54863 2.7171E−02 0.0437 0.0041

MOPSO 4.5437E−03 6.8558E−04 0.94312 3.6719E−01 N/A N/A
CMOPSO 2.9894E−03 8.3111E−03 0.57586 2.2894E−02 N/A N/A
EM-MOPSO N/A N/A N/A N/A 0.0406 0.0017
MWCA 8.7102E−04 5.7486E−05 0.17452 2.0441E−02 0.0350 0.0010

Table 5
Comparison of the various algorithms based on the mean and SD values of the performance metrics for the TNK problem.

Methods GD 

Mean SD Mean SD

NSGA-II 4.0488E−03 4.3465E−04 0.82286 2.8678E−04

MOPSO 5.0877E−03 4.5564E−04 0.79363 5.1029E−02
CMOPSO 5.4811E−04 7.9634E−05 0.25871 2.7272E−02
MWCA 1.3067E−03 6.1979E−05 0.60211 2.8148E−02

584 satisfactory distribution and spread for the non-dominated solu- 4.1.3. Constrained problem 3: KITA 607

585 tions. In general, Fig. 7 demonstrates the graphical comparison KITA function, first introduced by Kita et al. [33], has been widely 608

586 between the exact and computed Pareto fronts using the proposed analyzed [11,34]. The mathematical formulation proposed by Kita 609

587 optimizer for solving the CMOPS studied in this paper. et al. [33] is as follows: 610

⎧
588 4.1.2. Constrained problem 2: TNK ⎨ f1 (x) = −x12 + x2
The second constrained problem having two design variables maximize , (27) 611
589
⎩ f2 (x) = 1 x1 + x2 + 1
590 is suggested in the form of a mathematical description as follows 2
591 [32]:
 subject to 612
f1 (x) = x1
592 minimize , (25)
f2 (x) = x2 1 13
g1 (x) = x1 + x2 − ≤0
6 2
subject to 1 15
593
g2 (x) = x1 + x2 − ≤ 0. (28)
  x  2 2 613

1
g1 (x) = −x12 − x22 + 1 + 0.1 cos 16 arctan ≤0 g3 (x) = 5x1 + x2 − 30 ≤ 0
x2
594 . (26) 0 ≤ x1 , x2 ≤ 7
g2 (x) = 0.5 − (x1 − 0.5)2 − (x2 − 0.5)2 ≤ 0

0 ≤ x1 , x2 ≤  The Pareto optimal front for the KITA function is convex and 614

continuous. Despite the fact that KITA is a simple problem, most 615

595 This problem has a discontinuous Pareto optimal front that lies algorithms are unable to find all of the points on PFoptimal [11]. 616

596 on the boundary of the first constraint and also includes some con- The statistical optimization results of six algorithms for the per- 617

597 vex regions [30]. Table 5 compares the reported mean solutions for formance metrics for this problem are shown in Table 6. 618

598 the NSGA-II, MOPSO, CMOPSO, and MOWCA. According to Table 5, From Table 6, it is apparent that the MOWCA has the minimum 619

599 in terms of the mean values of the GD metric, the CMOPSO has GD and that there is also a huge difference between the mean GD 620

600 outperformed the other methods, and the MOWCA is placed in the value obtained by the MOWCA and those obtained by the other 621

601 second rank, while having better stability for the obtained solutions methods (e.g., the PAES and Micro-GA). Moreover, the MOWCA not 622

602 (a lower SD value) compared with the other optimizers. only could reach the best GD value but also possesses the first rank 623

603 Likewise, the MOWCA has the second best mean solutions after in terms of having the minimum values for the metrics of spacing 624

604 the CMOPSO, judging by the mean values of the  metric. Moreover, and spread. Furthermore, based on Fig. 7c, it can be emphasized 625

605 looking at Fig. 7b, we can see that PFoptimal is finely covered by the that the non-dominated solutions detected by the MOWCA have 626

606 MOWCA with a good distribution of the non-dominated solutions. the highest accuracy with the best distribution for this problem. 627

Table 6
Comparison of the various algorithms based on the mean and SD values of the performance metrics for the KITA problem.

Methods GD  S

Mean SD Mean SD Mean SD

NSGA-II 0.04 0.044 0.7863 0.1951 0.1462 0.1518

MOPSO 0.0467 0.0535 0.9925 0.1176 0.3184 0.4778
MOCSO 0.0274 0.0324 1.0169 0.1147 0.1592 0.2338
Micro-GA 0.1507 0.0897 N/A N/A 0.3150 0.4217
PAES 0.1931 0.0332 N/A N/A 0.1101 0.0995
MWCA 0.0049 0.0045 0.3764 0.0744 0.0485 0.0478

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 11

Table 7
Comparison of the various algorithms based on the mean and SD values of the performance metrics for the SRN problem.

Methods GD  S

Mean SD Mean SD Mean SD

NSGA-II 3.7069E−03 5.1034E−04 0.3869 2.5115E−02 1.5860 0.1337

MOPSO 2.7623E−03 2.0794E−04 0.6655 7.2196E−02 N/A N/A
CMOPSO 2.5331E−02 5.2561E−03 0.1965 2.4527E−02 N/A N/A
EM-MOPSO N/A N/A N/A N/A 1.2439 0.1055
MWCA 2.5836E−02 5.0102E−03 0.1477 1.3432E−02 0.4164 0.0779

628 4.1.4. Constrained problem 4: SRN Table 8

Comparison of the various algorithms in terms of the mean and SD values of the S
629 This problem, presented by Srinivas and Deb [35], has two
metric for the BNH problem.
630 design variables with a continuous Pareto optimal front. It is
631 defined mathematically as follows: Methods S
 Mean SD
f1 (x) = 2 + (x1 − 2)2 + (x2 − 1)2
632 minimize , (29) NSGA-II 0.7756 0.0727
f2 (x) = 9x1 − (x2 − 1)2 EM-MOPSO 0.6941 0.0385
MWCA 0.5680 0.0265
633 subject to

g1 (x) = x12 + x22 − 225 ≤ 0 minimum value of 0.5680 for the mean S metric and better stability 664

634 g2 (x) = x1 − 3x2 + 10 ≤ 0 . (30) for finding non-dominated solutions (lower SD value). 665

Furthermore, Fig. 7e depicts the final solutions achieved by the 666

−20 ≤ x1 , x2 ≤ 20 MOWCA versus PFoptimal . As can be seen from Fig. 7e, the final solu- 667

tions have suitable matches in terms of the GD metric with PFoptimal . 668
635 The obtained values of three performance parameters are given
636 in Table 7 for the NSGA-II, MOPSO, CMOPSO, EM-MOPSO, and
637 MOWCA. The mean values of GD for all of the algorithms indicate 4.1.6. Constrained problem 6: OSY 669

638 that the MOPSO and NSGA-II have the best performances. Never- The sixth constrained function was first suggested by Osyczka 670

639 theless, their values of mean S and  indicate that these algorithms and Kundu [37] and has been widely investigated [12,16]. The 671

640 were unable to find solutions having an acceptable distribution and mathematical formulation of this problem is as follows: 672

641 spread. ⎧ 2 2 2
642 In contrast, the MOWCA has surpassed those reported optimi- ⎨ f 1 (x) = −[25(x1 − 2) + (x2 − 2) + (x3 − 1)
643 zers in term of the metrics of spacing and spread; however, the minimize +(x4 − 4)2 + (x5 − 1)2 ] , (35) 673

non-dominated solutions attained by the MOWCA are slightly far

⎩
644
f2 (x) = x12 + x22 + x32 + x42 + x52 + x62
645 from PFoptimal . Overall, it can be said that the MOWCA outperformed
646 other algorithms in two out of three performance metrics. In addi- subject to 674

647 tion, Fig. 7d depicts how well the MOWCA could cover PFoptimal .
g1 (x) = 2 − x1 − x2 ≤ 0
648 4.1.5. Constrained problem 5: BNH g2 (x) = x1 + x2 − 6 ≤ 0
649 This problem is a 2D problem subject to two constraints. This
650 problem was previously presented by Binh and Korn [36] and is g3 (x) = x2 − x1 − 2 ≤ 0
651 given as follows: g4 (x) = x1 − 3x2 − 2 ≤ 0

f 1 (x) = 4x12 + 4x22 g5 (x) = x4 + (x3 − 3)2 − 4 ≤ 0 . (36) 675
652 minimize , (31)
2 2
f 2 (x) = (x1 − 5) + (x2 − 5) g6 (x) = 4 − x6 − (x5 − 3) ≤ 0 2

653 subject to 0 ≤ x1 , x2 , x6 ≤ 10
2
g1 (x) = (x1 − 5) + x22 − 25 ≤ 0 1 ≤ x3 , x5 ≤ 5
2 2
g2 (x) = 7.7 − (x1 − 8) − (x2 + 3) ≤ 0 0 ≤ x4 ≤ 6
654 . (32)
0 ≤ x1 ≤ 5 The OSY problem has six design variables, and its Pareto optimal 676

front is a concatenation of five regions, as shown in Fig. 8. For all 677

0 ≤ x2 ≤ 3
the Pareto optimal regions, x4 = x6 = 0. Accordingly, Table 9 shows 678

655 It is worth mentioning that the second constraint (g2 (x) in Eq. other values of the design variables for each region [30]. 679

656 (32)) does not have any effect on the boundary of the infeasible The statistical optimization results from Table 10 illustrate the 680

657 region [30]. In addition, the Pareto optimal front for this problem values of the performance parameters obtained by the NSGA-II, VIS, 681

658 is as follows [30]: WBMOIA, HQIA, and MOWCA. Looking at the GD values summa- 682

rized in Table 10, we can see that the best GD values obtained by 683
659 For 0 ≤ x1∗ ≤ 3 : x2∗ = x1∗ , (33) the NSGA-II and MOWCA are 9.89E−01 and 9.68E−02, respectively. 684

Hence, in this problem, the MOWCA is placed in the second rank

For 3 ≤ x1∗ ≤5: x2∗ = 3.
685
660 (34)
in terms of mean GD values after the HQIA. However, the distribu- 686

661 The statistical results attained by the NSGA-II, EM-MOPSO, and tion of the non-dominated solutions reported by the MOWCA is 687

662 MOWCA are represented in Table 8. From Table 8, it can be seen that acceptable, so the MOWCA is located in the first rank with respect 688

663 the MOWCA has surpassed other methods in terms of having the to the S metric (see Table 10). Furthermore, Fig. 7f shows the final 689

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
12 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

Table 9
Design variable values for the Pareto optimal front for the OSY problem.

Regions Optimal values of design variables

x1∗ x2∗ x3∗ x5∗

AB 5 1 (1, . . ., 5) 5
BC 5 1 (1, . . ., 5) 1
CD (4.056, . . ., 5) (x1∗ − 2)/3 1 1
DE 0 2 (1, . . ., 3.732) 1
EF (0, . . ., 1) 2 − x1∗ 1 1

Table 10
Comparison of the various algorithms based on the mean and SD of the performance metrics for the OSY problem.

Methods GD S

Mean SD Mean SD

NSGA-II 9.89E−01 9.78E−01 1.14E+00 2.75E−01

VIS 3.15E−01 2.38E−01 8.30E−01 3.51E−01
WBMOIA 9.85E−02 1.00E−01 9.46E−01 2.27E−01
HQIA 3.49E−02 1.54E−02 7.21E−01 1.60E−01
MWCA 9.68E−02 7.18e−02 5.22E−01 9.52E−02

690 non-dominated solutions obtained by the MOWCA compared with subject to 702

691 PFoptimal for the OSY problem. F  F 

≤ x1 ≤ 3 ×
 
692 4.2. Multi-objective engineering design problems √ F  F 
2× ≤ x2 ≤ 3 ×
 
693 In this subsection, six mechanical component design opti- F  F  , (38) 703

mization problems with multiple objectives are studied to show √

694
2× ≤ x3 ≤ 3 ×
695 the efficiency and performance of the MOWCA in finding non-  
696 dominated solutions having an acceptable distribution and spread. F  F 
≤ x4 ≤ 3 ×
 
697 4.2.1. Four-bar truss design problem where 704
698 The four-bar truss design problem, depicted in Fig. 9, has been 5 2
F = 10 KN, E = (2)10 KN/cm
699 widely used to evaluate and validate different methods [11,27]. The . (39) 705
700 mathematical formulation of this problem is as follows: L = 200 cm,  = 10 KN/cm3
⎧  √
⎪
⎨ f 1 (x) = L(2x1 + 2x2 + x3 + x4 )
As can be seen from Eq. (37), the volume of the four-bar truss and 706

701 minimize √ √ ! , (37)

the joint displacement are to be optimized simultaneously. Table 11 707

⎪ FL 2 2 2 2 2 2 gives an overview of final optimization results reported by different

⎩ f 2 (x) = E
708
+ − +
x2 x2 x3 x4 algorithms. 709

Based on the GD values summarized in Table 11, the MOWCA 710

has the best performance in finding non-dominated solutions with 711

a suitable match with PFoptimal (having the minimum distance to 712

PFoptimal ). In addition to having a lower GD value, the MOWCA was 713

able to maintain a suitable distance among the generated solutions 714

that were shown in the S metric values. 715

Although, the NSGA-II and MOPSO have slightly better perform- 716

ances regarding the mean values of S, the SD value of the MOWCA 717

for the S metric shows better stability of solutions corresponding to 718

Fig. 8. Schematic view of the Pareto optimal front for the OSY problem [30]. Fig. 9. Schematic view of a four-bar truss design problem.

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 13

Fig. 10. Comparison of non-dominated solutions obtained by the MOPSO, Micro-GA, NSGA-II, MOPSO, PAES, and MOWCA for the four-bar truss (Solid lines and dot points
represent the optimal and generated (obtained) Pareto fronts, respectively).

719 the distribution of the non-dominated solutions. Moreover, Fig. 10 4.2.2. Speed reducer design problem 723

720 shows a graphical comparison of the NSGA-II, MOPSO, Micro-GA, The speed reducer design problem, shown in Fig. 11, has seven 724

721 PAES, and MOWCA. The discussion regarding Table 11 is confirmed design variables and has been studied widely in the literature 725

722 in Fig. 10. [27,38–40]. 726

Table 11
Comparison of the values for the mean and SD values for the used performance Table 12
metrics for the four-bar truss problem. Comparison of the various algorithms based on the mean and SD values of the
performance metrics for the speed-reducer design problem.
Methods GD S
Methods GD S
Mean SD Mean SD
Mean SD Mean SD
NSGA-II 0.3601 0.0470 2.3635 0.2551
MOPSO 0.3741 0.0422 2.5303 0.2275 NSGA-II 9.843702 7.08103039 2.765449155 3.53493787
Micro-GA 0.9102 1.7053 8.2742 16.8311 Micro-GA 3.117536 1.67810867 47.80098 32.80151572
PAES 0.9733 1.8211 3.2314 5.9555 PAES 77.99834 4.21026087 16.20129 4.26842769
MOWCA 0.2076 0.0055 2.5816 0.0298 MOWCA 0.98831 0.17894217 16. 68520 2.69694436

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
14 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

Table 13
Comparison of the various algorithms based on the mean and SD values of the
performance metrics for the disk-brake design problem.

Methods GD 

Mean SD Mean SD

NSGA-II 3.0771 0.10782 0.79717 0.06608

pa␧-ODEMO 2.6928 0.24051 0.84041 0.20085
MOWCA 0.0244 0.12314 0.46041 0.10961

discrepancy between the mean GD values obtained by the MOWCA 736

and the Micro-GA as the second best method corresponding to 737

values of 0.98 and 3.11, respectively (see Table 12). 738

In Fig. 12, graphical presentations are given for comparison pur- 739
Fig. 11. Schematic view of a speed reducer design optimization problem.
poses among the reported methods. Likewise, from Fig. 12, we can 740

say that the MOWCA has the best performance among the methods, 741
727 The mathematical formulation used for modeling this problem and that the NSGA-II and PAES were unsuccessful in finding a set of 742
728 is as follows [41]: solutions close to PFoptimal . Based on Fig. 12 and given the results in 743
⎧ Table 12, it can be concluded that the MOWCA is the best algorithm
⎪ f 1 (x) = 0.7854x1 x22 (10x32 /3 + 14.933x3 744
⎪
⎪ regarding the mean S and GD values, and that it would be unwise
⎪
⎪
745

⎪
⎨ − 43.0934) − 1.508x1 (x62 + x72 ) to introduce the NSGA-II as the best approach by considering the 746

729 minimize + 7.477(x63 + x73 ) + 0.7854(x4 x62 + x5 x72 ) , (40) mean S value for this problem. 747

⎪
⎪ 
⎪
⎪
⎪
⎪
2 7 4.2.3. Disk brake design problem
⎩ f 2 (x) = (745.0x4 /x2 x3 ) 3 + 1.69 × 10
748

0.1x6 This problem has four design variables first studied by Ray and 749

Liew [42]. The main objective of this problem is to minimize the 750
730 subject to stopping time and the mass of a brake. The mathematical formula- 751

1.0 1.0 tion of this problem is as follows: 752

g1 (x) = − ≤0 ⎧ −5 2
x1 x22 x3 27.0 ⎪ 2
⎨ f 1 (x) = 4.9 × 10 (x2 − x1 )(x4 − 1)
1.0 1.0
g2 (x) = − ≤0 minimize 9.82 × 106 (x22 − x12 ) , (42) 753
x1 x2 x2
2 3
397.5 ⎪
⎩ f 2 (x) = 3 3
x3 x4 (x2 − x1 )
x43 1.0
g3 (x) = − ≤0
x2 x3 x64 1.93 subject to 754

x53 1.0 g1 (x) = 20 + x1 − x2 ≤ 0

g4 (x) = − ≤0
x2 x3 x74 1.93
g2 (x) = 2.5(x4 + 1) − 30 ≤ 0
g5 (x) = x2 x3 − 40.0 ≤ 0
x3
x1 g3 (x) = 2
− 0.4 ≤ 0
g6 (x) = − 12 ≤ 0 3.14(x22 − x12 )
x2
2.22 × 10−3 x3 (x23 − x13 )
x1 g4 (x) = −1≤0
g7 (x) = 5 − ≤0 2
x2 (x22 − x12 )
2.66 × 10−2 x3 x4 (x23 − x13 ) , (43) 755
g8 (x) = 1.9 − x4 + 1.5x6 ≤ 0 g5 (x) = 900 − ≤0
2 2
(x2 − x1 )
731
g9 (x) = 1.9 − x5 + 1.1x7 ≤ 0 . (41)
 55 ≤ x1 ≤ 80
2
(745x4 /x2 x3 ) + 1.69 × 107 75 ≤ x2 ≤ 110
g10 (x) = − 1300 ≤ 0
0.1x63
 1000 ≤ x3 ≤ 3000
2
(745x5 /x2 x3 ) + 1.575 × 108
g11 (x) = − 1100 ≤ 0 2 ≤ x4 ≤ 20
0.1x73
2.6 ≤ x1 ≤ 3.6 where x1 , x2 , x3 , and x4 are the inner radius of the disk, the outer 756

radius of the disk, the engaging force, and the number of fric- 757
0.7 ≤ x2 ≤ 0.8
tion surfaces, respectively. This engineering problem has been 758

17 ≤ x3 ≤ 28 optimized previously using the NSGA-II and pa␧-ODEMO [28]. 759

Comparisons of the performance metric parameters related to 760

7.3 ≤ x4 ≤ 8.3 the reported methods compared with the MOWCA are given in 761

7.3 ≤ x5 ≤ 8.3 Table 13.Judging by Table 13, the MOWCA could find a wide vari- 762

ety of solutions having uniform spread and the smallest deviation 763
2.9 ≤ x6 ≤ 3.9 from PFoptimal . In this problem, the MOWCA proved its efficiency 764

against other optimizers. The mean values of the GD and  metrics 765
5.0 ≤ x7 ≤ 5.5
obtained by the MOWCA, 0.02 and 0.46, respectively, support this 766

732 Table 12 represents the comparison of values for the considered claim. In addition, Fig. 13 depicts graphical comparisons with the 767

733 performance metrics for the speed reducer problem using the vari- other reported methods. 768

734 ous algorithms. In Table 12, the MOWCA shows its superiority over Moreover, it is worth mentioning that the non-dominated solu- 769

735 other methods in terms of the GD values. Note that there is a large tions obtained by the MOWCA are in the range of (0.1274, 16.6549) 770

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 15

Fig. 12. Comparison of distribution of non-dominated solutions obtained by the Micro-GA, PAES, NSGA-II, and MOWCA for the speed reducer (solid lines and dot points
represent the optimal and generated (obtained) Pareto fronts, respectively).

771 and (2.7176, 2.0828), while those obtained by the NSGA-II are in the The first constraint in Eq. (45), g1 (x), is related to the shear stress 789

772 range of (0.1293, 17.598) and (2.7915, 2.1127). Similarly, this range imposed on the support location of the beam, which must be less 790

773 for the pa␧-ODEMO is (0.1274, 16.6549) and (2.8684, 2.0906) [28]. than the shear strength of the material (13,600 psi). In addition, 791

774 These results show that the brake mass of 2.8684 with minimum the normal stress generated at the support location of the beam 792

775 stopping time of 2.0906 obtained by the pa␧-ODEMO is domi- (the second constraint, g2 (x), in Eq. (45)) must be smaller than the 793

776 nated by (2.7176, 2.0828) attained by the MOWCA, confirming the allowable yield strength of the material (30,000 psi). Also, the third 794

777 exploratory capabilities of the MOWCA for finding accurate results. constraint, g3 (x), indicates that the thickness of the beam must be 795

larger than the weld thickness. 796

778 4.2.4. Welded beam design problem For the last constraint, g4 (x), the allowable buckling load (along 797

779 The welded beam design problem, shown in Fig. 14, has four the thickness direction) of the beam must be greater than the value 798

780 design variables and is selected from Ray and Liew [42]. In this of the applied load (F). The shear stress, normal stress, and other 799

781 case, the fabrication cost and end deflection of the beam must be parameters used in the constraints, Eq. (45), are defined as follows 800

782 minimized subject to imposed constraints on shear stress, bending [28]: 801

stress, and buckling load [43].

"
783

784 The two objective functions and four nonlinear constraints

l 
785 related to this problem are as follows: (x) = ( )2 + ( )2 + 
⎧ 2
0.25(l2 + (h + t) )
⎨ f 1 (x) = 1.10471h2 l + 0.0481tb(14.0 + t) 6000
 = √
786 minimize , (44) 2hl
⎩ f 2 (x) = 2.1952 
t3b 6000(14.0 + 0.5l) 0.25(l2 + +(h + t) )
2
 =
2[0.707hl((l2 /12) + 0.25(l2 + +(h + t) ))] . (46)
787 subject to 2 802

504, 000
g1 (x) = (x) − 13, 600 ≤ 0 (x) =
t2b
g2 (x) = (x) − 30, 000 ≤ 0
Pc(x) = 64, 764.022(1 − 0.0282346t)tb3
788 . (45)
g3 (x) = h − b ≤ 0
0.125 ≤ h, b ≤ 5.0
g4 (x) = 6000 − Pc(x) ≤ 0 0.1 ≤ l, t ≤ 10.0

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
16 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

Fig. 13. Distribution of non-dominated solutions obtained by the NSGA II, pa␧-ODEMO, and MOWCA for the disk brake design problem.

803 Final statistical optimization results obtained by applying the NSGA-II are 2.8959 units with a deflection of 0.0131 inches and 814

804 NSGA-II, pa␧-ODEMO, and MOWCA are collected in Table 14. 3.0294 units with a deflection of 0.0088 inches, respectively [28]. 815

805 Regarding the mean GD values in Table 14, the MOWCA having a GD Hence, the MOWCA has the best performance based on the 816

806 value of 0.04909 and the pa␧-ODEMO with a GD value of 0.09169 fabrication cost (having minimum fabrication cost). Likewise, the 817

807 have stood in the first and second ranks, respectively. Similarly, the minimum deflection related to the fabrication cost obtained by the 818

808 MOWCA and pa␧-ODEMO could maintain a uniform distribution of NSGA-II and pa␧-ODEMO are (37.4018, 0.000439) and (36.6172, 819

809 the solutions with a good spread, as is revealed from the  values 0.00044), respectively [28]. These statistics indicate that the pro- 820

810 given in Table 14. posed approach was successful in finding a wide variety of 821

811 Moreover, the extreme solutions (ranges) obtained by the Pareto-optimal solutions. For more assessments, Fig. 15 demon- 822

812 MOWCA are (2.5325, 0.0115) and (37.9686, 0.000439). In contrast, strates graphical comparisons among the reported multi-objective 823

813 the minimum fabrication cost obtained by the pa␧-ODEMO and optimizers. 824

4.2.5. Spring design problem 825

The spring problem requires a design based on minimum stress 826

and minimum volume. This problem has three design variables: 827

number of spring coils (N: x1 ), an integer variable; wire diameter 828

(d: x2 ), a discrete variable; and spring diameter (D: x3 ), a continuous 829

decision variable. This problem has two objective functions subject 830

Table 14
Comparison of the various algorithms based on the mean and SD values of the
performance metrics for the welded-beam design problem.

Methods GD 

Mean SD Mean SD

NSGA-II 0.16875 0.08030 0.88987 0.11976

pa␧-ODEMO 0.09169 0.00733 0.58607 0.04366
MOWCA 0.04909 0.02821 0.22478 0.09280
Fig. 14. Schematic view of a welded beam design optimization problem.

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 17

Fig. 15. Distribution of the non-dominated solutions obtained by the NSGA II, pa␧-ODEMO, and MOWCA for the welded beam design problem.

831 to eight constraints. The mathematical formulation of this problem where the parameters used in the objective functions and con- 836

832 is as follows [43]: straints (47) and (48), respectively, are as follows [43]: 837

⎧ 2 3 4C − 1 0.615x2
⎨ f 1 (x) = 0.25 x2 x3 (x1 + 2) K= + P = 300 lb Dmax = 3 in.
4C − 4 x3
833 minimize 8KPmax x3 , (47)
⎩ f 2 (x) = Gx24
x23 k= Pmax = 1000 lb ıw = 1.25 in.
8x1 x33 . (49) 838

subject to P
834 ıp = lmax = 14 in. ıpm = 6 in.
k
Pmax D
g1 (x) = 1.05x2 (x1 + 2) + − lmax ≤ 0 S = 189, 000 kpsi dmin = 0.2 in. C=
k d
g2 (x) = dmin − x2 ≤ 0 In this problem, x2 is a discrete variable whose values will be 839

selected from the set F = [0.009, 0.0095, 0.0104, 0.0118, 0.0128, 840
g3 (x) = x2 + x3 − Dmax ≤ 0
0.0132, 0.014, 0.015, 0.0162, 0.0173, 0.018, 0.020, 0.023, 0.025, 841

g4 (x) = 3 − C ≤ 0 0.028, 0.032, 0.035, 0.041, 0.047, 0.054, 0.063, 0.072, 0.080, 0.092, 842

835 , (48) 0.105, 0.120, 0.135, 0.148, 0.162, 0.177, 0.192, 0.207, 0.225, 0.244, 843
g5 (x) = ıp − ıpm ≤ 0
0.263, 0.283, 0.307, 0.331, 0.362, 0.394, 0.4375, 0.5]. In addition, x1 844

Pmax − P will be taken to be integer values in the range of Refs. [1,33]. In 845
g6 (x) = ıw − ≤0
k addition, the lower and upper bounds for x3 are 0 and 3 inches, 846

8KPmax x3 respectively. There were no statistical results for performance 847

g7 (x) = −S ≤0 metrics (i.e., the GD, S, and ) for this problem in literature. There- 848
x23
fore, Table 15 compares extreme solutions obtained by the NSGA-II 849
g8 (x) = [0.252 x23 x3 (x1 + 2)] − Vmax ≤ 0 and MOWCA in terms of interval span. 850

Table 15
Comparison of the extreme solutions for the objective functions (interval span) for the spring design problem.

Algorithms Solutions Design variables Objective functions

x1 x2 x3 f1 f2

NSGA-II Min. volume 5 0.307 1.619 2.690 187,053

Min. stress 24 0.5 1.865 24.189 61,949

MOWCA Min. volume 9 0.283 1.227 2.668 188,448

Min. stress 19 0.5 2.083 26.983 58,752

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
18 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

is provided. The results of the comparison for extreme solutions 873

obtained by the NSGA-II and MOWCA are tabulated in Table 16. 874

By observing Table 16, there is an increase in the range of non- 875

dominated solutions obtained by applying the MOWCA compared 876

with the NSGA-II. This feature is more obvious for the f2 function 877

(see Table 16). 878

To clarify further, the MOWCA could improve the minimum 879

error ratio from the 1.83E−08 value obtained by the NSGA-II to 880

a value of 4.50E−09. Furthermore, a similar trend related to the 881

second extreme solution (F2 ) is observed. Hence, these intervals 882

support the idea that the MOWCA is better able to find a wider 883

range of solutions compared with the NSGA-II. Indeed, wider range 884

means giving more choices to the decision maker to select his or 885

her optimum design from the set of non-dominated solutions. 886

Fig. 16. Schematic view of a gear train design optimization problem. As the WCA have shown its superiority for tackling single objec- 887

tive functions [23], in this article, we demonstrated the advantages 888

of the concepts and strategy of WCA for handling constrained MOPs. 889
851 Judging by Table 15, the MOWCA offers a wider range of non-
The MOWCA, so for the WCA, proves its efficiency for solving MOPs. 890
852 dominated solutions compared to the NSGA-II. The minimum
Offering wider range for non-dominated solutions (disk brake, 891
853 volume reported by the MOWCA is 2.668 units, which is 0.022 units
welded beam, spring design, and gear train design problems), 892
854 less than the minimum volume attained by the NSGA-II. Likewise,
obtaining the generated Pareto front close to the optimal Pareto 893
855 the minimum stress as the second bound of the Pareto front gained
front (i.e., for Section 4.2), and good distribution of the non- 894
856 by the NSGA-II and MOWCA reveals the same results. In fact, the
dominated solutions (metric of spacing) compared with other 895
857 MOWCA detected a wider range for both objective functions, and
optimizers, discussed earlier (Sections 4.2 and 4.3), are considered 896
858 the minimum values of f1 and f2 obtained by the MOWCA are less
as advantages of MOWCA for handling MOPs. 897
859 than the values found using the NSGA-II.

860 4.2.6. Gear train design problem

861 The main purpose of the gear train design problem, shown in 4.3. Sensitivity analysis for initial parameters of MOWCA 898

862 Fig. 16, is to find the number of teeth in each gear (x1 : Td , x2 : Tb , x3 :
863 Ta , and x4 : Tf ). In fact, this problem tries to minimize the maximum The main objective of this subsection is to evaluate and study 899

864 size of any of the four gears and minimize the error between the the effects of each initial parameter used in the MOWCA. One of 900

865 obtained gear ratio and a required gear ratio of 1/6.931 [43]. Hence, the most important concerns related to optimization algorithms is 901

866 this problem is as follows [43]: to find the most efficient user parameters. Performing sensitivity 902

⎧ # 1 $ analyses show the stability of the methods and importance of initial 903
⎨ f (x) = x1 x2 2 parameters to find the optimal solution against any changes in user
1 − 904

867 minimize 6.931 x3 x4 . (50) parameters. 905

⎩ Especially, when the problem is complex having many local
f2 (x) = max(x1 , x2 , x3 , x4 )
906

optima, it is crucially important to use the proper and efficient user 907

868 It is worth mentioning that all design variables in this problem parameters to boost the convergence speed and efficiency of the 908

869 have integer values that vary from 12 to 60. algorithms for finding global solution without getting trapped in 909

870 For performance metrics (i.e., the GD, S, and ) for this prob- local minima. 910

871 lem, there had not existed statistical reports in literature. Therefore, The problems considered are the CONSTR problem (Section 911

872 based on the results in literature, comparison for extreme solutions 4.1.1) and speed reducer design problem (Section 4.2.2) have been 912

Table 16
Comparison of the extreme solutions for objective functions (interval span) for the gear train design problem.

Algorithms Solutions Design variables Objective functions

x1 x2 x3 x4 f1 f2

NSGA- Min. F1 12 12 27 37 1.83E−08 37

II Min. F2 12 12 12 13 5.01E−01 13

MOWCA Min. F1 12 15 43 29 4.50E−09 43

Min. F2 12 12 12 12 7.32E−01 12

Table 17
Effects of the MOWCA parameters on the GD metric for the CONSTR problem.

Results Maximum Iteration = 100

Npop = 50 Nsr = 8 and dmax = 1e−5

dmax = 1E−5 Nsr = 8 Npop = 10 Npop = 30 Npop = 50

Nsr = 4 Nsr = 8 Nsr = 10 dmax = 1e−1 dmax = 1e−3 dmax = 1e−5 dmax = 1e−7

Best 8.22E−4 7.98E−4 8.34E−4 8.28E−4 7.98E−4 7.98E−4 8.31E−4 8.28E−4 8.89E−4 7.98E−4
Average 9.08E−4 9.01E−4 9.17E−4 9.12E−4 9.08E−4 9.01E−4 9.96E−4 9.36E−4 9.96E−3 9.01E−4
Worst 1.08E−4 1.04E−3 1.04E−4 1.06E−3 1.04E−3 1.04E−3 1.13E−3 1.14E−3 5.95E−2 1.04E−3
SD 5.86E−4 5.77E−5 5.21E−5 5.71E−5 5.77E−5 5.77E−5 6.51E−5 8.32E−5 1.22E−2 5.77E−5

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx 19

Table 18
Effects of the MOWCA parameters on metric of spacing for the CONSTR problem.

Results Maximum Iteration = 100

Npop = 50 Nsr = 8 and dmax = 1e−5

dmax = 1E−5 Nsr = 8 Npop = 10 Npop = 30 Npop = 50

Nsr = 4 Nsr = 8 Nsr = 10 dmax = 1e−1 dmax = 1e−3 dmax = 1e−5 dmax = 1e−7

Best 2.97E−2 2.73E−2 2.83E−2 2.78E−2 2.89E−2 2.73E−2 2.80E−2 2.82E−2 4.45E−2 2.73E−2
Average 3.52E−2 3.36E−2 3.39E−2 3.76E−2 3.39E−2 3.36E−2 3.54E−2 3.54E−2 3.41E−1 3.36E−2
Worst 5.78E−2 4.79E−2 6.19E−2 6.67E−2 4.54E−2 4.79E−2 6.23E−2 6.99E−2 1.23 4.79E−2
SD 5.03E−3 3.19E−3 7.05E−3 7.26E−3 3.42E−3 3.19E−3 6.46E−3 9.12E−3 2.62E−1 3.19E−3

Judging by three first columns of Tables 17 and 18, “Nsr = 8”, “dmax = 1e−5”, and “Npop = 50” have the best performance in terms of performance metrics in comparison with
other values. Similarly, the speed reducer design problem is investigated and the optimization results are tabulated in Tables 19 and 20 for different performance evaluators.

Table 19
Effects of the MOWCA parameters on the GD metric for the speed reducer problem.

Results Maximum Iteration = 100

Npop = 50 Nsr = 8 and dmax = 1e−5

dmax = 1e−5 Nsr = 8 Npop = 10 Npop = 30 Npop = 50

Nsr = 4 Nsr = 8 Nsr = 10 dmax = 1e−1 dmax = 1e−3 dmax = 1e−5 dmax = 1e−7

Best 0.94 0.87 0.91 0.91 0.88 0.87 0.99 1.24 0.92 0.87
Average 1.67 1.42 1.45 1.44 1.52 1.42 1.46 11.93 1.59 1.42
Worst 4.50 2.49 2.89 3.47 2.69 2.49 3.19 61.72 2.64 2.49
SD 0.81 0.49 0.55 0.56 0.52 0.49 0.57 15.92 0.50 0.49

Table 20
Effects of the MOWCA parameters on metric of spacing for the speed reducer problem.

Results Maximum Iteration = 100

Npop = 50 Nsr = 8 and dmax = 1e−5

dmax = 1E−5 Nsr = 8 Npop = 10 Npop = 30 Npop = 50

Nsr = 4 Nsr = 8 Nsr = 10 dmax = 1e−1 dmax = 1e−3 dmax = 1e−5 dmax = 1e−7

Best 11.77 11.05 12.67 15.12 11.85 11.05 12.44 13.55 12.95 11.05
Average 16.24 16.63 16.92 17.14 16.33 16.63 16.35 25.14 17.81 16.63
Worst 18.54 18.04 18.84 19.25 21.34 18.04 18.36 66.72 18.82 18.04
SD 1.79 0.86 3.42 1.93 2.12 0.86 1.61 17.66 2.98 0.86

913 investigated by changing the user parameters of MOWCA which The last parameter which can have effects on the performance 940

914 are Npop , Nsr , dmax , and the maximum iteration number. of WCA is the number of rivers and sea (Nsr ). It is worth mention- 941

915 In each problem, different optimization results for various per- ing that the Npop is considered as a common user parameter for 942

916 formance metrics (i.e., GD and S) using diverse user parameters are the most metaheuristic algorithms. In fact, the chosen value of Nsr 943

917 represented. Tables 17 and 18 represent the sensitivity of initial should be proportional with the chosen value of Npop . 944

918 parameters with respect to the values for different performance For instance, when values of Nsr and Npop are equal to 8 and 10, 945

919 metrics including the GD and S for the CONSTR problem, respec- respectively, it means that some rivers have been created without 946

920 tively. having any streams moving toward those rivers. These parame- 947

921 As the second user parameter, dmax is used for applying the evap- ters’ selection can cause the detrimental effects on the optimization 948

922 oration condition and raining process and affects the exploration results. By observing Tables 17–20, when Nsr is equal to 8, the worst 949

923 phase in the WCA. Generally, the large values for dmax reduce a and best performance corresponds to “Npop = 10” and “Npop = 50”, 950

924 suitable search around the best obtained point (i.e., sea). By set- respectively. 951

925 ting a large value for dmax (i.e., increasing the number of raining In conclusion, it can be inferred that the best value for dmax is 952

926 process), the WCA operates as a random search method instead of in the range of 1e−5 and 1e−3 and also, the value of Nsr should be 953

927 a being metaheuristic one. Hence, it may not be able to find the selected based on the value of Npop . Indeed, the best value of the 954

928 global optimum solution. Nsr may be chosen as one sixth of Npop . Therefore, for population 955

929 Likewise, the small values of dmax have a negative influence on size of 50, eight rivers and sea (Nsr ) may be considered as a proper 956

930 the final results as it can reduce the number of raining processes value. Note that if the number of iteration increases, the WCA or 957

931 (i.e., fewer explorations). Therefore, the WCA concentrates on the other metaheuristic algorithms will have more chances (i.e., com- 958

932 domain solution near to the best obtained point (sea) without pay- putational time) to find the best solution and explore the search 959

933 ing more attentions to other regions. As a consequence, for complex domain. 960

934 problems having many local optima, the WCA traps in local solu-
935 tions. 5. Conclusions 961
936 As it can be seen from Tables 17–20, dmax = 1e−1 and dmax = 1e−7
937 found the worst solutions with the largest SD values. Therefore, it This paper presented multi-objective water cycle algorithm 962
938 can be concluded that if an unsuitable value select for dmax , the (MOWCA), a novel multi-objective optimization technique for 963
939 moving toward the best solution probably misdirects. solving constrained multi-objective problems (CMOPs). The basic 964

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042
 G Model
ASOC 2618 1–20 ARTICLE IN PRESS
20 A. Sadollah et al. / Applied Soft Computing xxx (2014) xxx–xxx

965 concepts of the proposed MOWCA are inspired by the water cycle [17] X. Zhang, B. Lu, S. Gou, L. Jiao, Immune multiobjective optimization algorithm 1027

966 process in the real world. In this paper, the MOWCA was used to for unsupervised feature selection, in: Proceedings of Applications of Evolu- 1028
tionary Computing, vol. 39, Springer, LNCS, Budapest, 2006. 1029
967 solve a number of well-known benchmark and engineering MOPs. [18] L. Jiao, M. Gong, R. Shang, et al., Clonal selection with immune dominance 1030
968 The efficiency and performance of the MOWCA was demonstrated and energy based multiobjective optimization, in: Proceedings of Third Inter- 1031
969 using three popular criteria (i.e., generational distance, metric of national Conference on Evolutionary Multi-Criterion Optimization, vol. 3410, 1032
LNCS, Springer-Verlag, Guanajuato, 2005, pp. 474–489. 1033
970 spacing, and spread metric). The statistical optimization results [19] X.L. Wang, M. Mahfouf, ACSAMO: an adaptive multiobjective optimization 1034
971 obtained from performance metrics revealed that the proposed algorithm using the clonal selection principle, in: Proceedings of Second Euro- 1035
972 multi-objective optimizer was able to approach a full optimal pean Symposium on Nature-inspired Smart Information Systems, Puerto de la 1036
Cruz, 2006. 1037
973 Pareto front and provide a superior quality of solutions in compar-
[20] A. Kaveh, K. Laknejadi, A novel hybrid charge system search and particle swarm 1038
974 ison to a variety of state-of-the art algorithms considered in this optimization method for multi-objective optimization, Expert Syst. Appl. 38 1039
975 paper. In general, the MOWCA offers competitive solutions com- (12) (2011) 15475–15488. 1040
[21] M.R. Narimani, A.A. Rasoul, Z.M.S. Behrouz, K. Gholami, A novel approach 1041
976 pared with other population-based algorithms, according to the
to multi-objective optimal power flow by a new hybrid optimization algo- 1042
977 reported results and the numerical results in this research offer- rithm considering generator constraints and multi-fuel type, Energy 49 (2013) 1043
978 ing a wider range of non-dominated solutions. In fact, although 119–120. 1044

979 the robustness and exploratory capability of the MOWCA depend [22] V.S.Q. Luis, G.H.D. Alfredo, M. Julián, C.A.C. Coello, DEMORS: a hybrid multi- 1045
objective optimization algorithm using differential evolution and rough set 1046
980 on the nature and complexity of the problems, the optimization theory for constrained problems, Comput. Oper. Res. 37 (2010) 470–480. 1047
981 results obtained show that the MOWCA might be a suitable alter- [23] H. Eskandar, A. Sadollah, A. Bahreininejad, M. Hamdi, Water cycle algorithm – a 1048

982 native method having an acceptable degree of accuracy for finding novel metaheuristic optimization method for solving constrained engineering 1049
optimization problems, Comput. Struct. 110–111 (2012) 151–160. 1050
983
Pareto optimal solutions for different scales of MOPs. [24] R.L. Haupt, S.E. Haupt, Practical Genetic Algorithms, 2nd ed., John Wiley & Sons, 1051
Inc., USA, 2004. 1052
[25] C.A.C. Coello, An updated survey of GA-based multi-objective optimization 1053
984 Acknowledgment techniques, ACM Comput. Surv. 32 (2) (2000) 109–110. 1054
[26] D.A.V. Veldhuizen, G.B. Lamont, Multiobjective Evolutionary Algorithm 1055
Q4
985 This work was supported by the National Research Foundation Research: A History and Analysis, Technical Report TR-98-03, Department of 1056
Electrical and Computer Engineering, Graduate School of Engineering, Air Force 1057
986 of Korea (NRF) grant funded by the Korean government (MSIP)
Institute of Technology, Wright-Patterson AFB, OH, 1998. 1058

987
(NRF-2013R1A2A1A01013886). [27] C.A.C. Coello, G.T. Pulido, Multiobjective structural optimization using a micro 1059
genetic algorithm, Struct. Multidiscip. Optim. 30 (5) (2005) 388–390. 1060
[28] W. Gong, Z. Cai, L. Zhu, An efficient multiobjective differential evolution algo- 1061
988 References rithm for engineering design, Struct. Multidiscip. Optim. 38 (2) (2009) 137–140. 1062
[29] J.R. Schott, Fault Tolerant Design Using Single and Multicriteria Genetic 1063
989 [1] T. Ray, K.S. Won, An evolutionary algorithm for constrained bi-objective opti- Algorithm Optimization (Master’s thesis), Department of Aeronautics and 1064
990 mization using radial slots, Lect. Notes Comput. Sci. 3684 (2005) 49–50. Astronautics, Massachusetts Institute of Technology, Cambridge, MA, 1995. 1065
991 [2] Z. Zhang, Immune optimization algorithm for constrained nonlinear multiob- [30] K. Deb, Multi-objective Optimization Using Evolutionary Algorithms, Wiley, 1066
992 jective optimization problems, Appl. Soft. Comput. 7 (3) (2007) 840–850. Chichester, 2001. 1067
993 [3] Y. Gao, M. Qu, Constrained multi-objective particle swarm optimization algo- [31] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proceedings of the 1068
994 rithm, in: Emerging Intelligent Computing Technology and Applications, vol. IEEE International Conference on Neural Networks, Perth, Australia, 1995, pp. 1069
995 304, 2012, pp. 47–50. 1942–1948. 1070
996 [4] B. Suman, Study of self-stopping PDMOSA and performance measure in multi- [32] M. Tanaka, GA-based decision support system for multicriteria optimization, in: 1071
997 objective optimization, Comput. Chem. Eng. 29 (2005) 1131–1147. IEEE International Conference on Systems, Man and Cybernetics, vol. 2, 1995, 1072
998 [5] E. Zitzler, L. Thiele, Multi objective evolutionary algorithms: a comparative case pp. 1556–1561. 1073
999 study and the strength Pareto approach, IEEE Trans. Evol. Comput. 3 (4) (1999) [33] H. Kita, Y. Yabumoto, N. Mori, Y. Nishikawa, Multi-objective optimization by 1074
1000 257–260. means of the thermodynamical genetic algorithm, in: H.M. Voigt, W. Ebeling, 1075
1001 [6] E. Zitzler, M. Laumanns, L. Thiele, SPEA2: Improving the Strength Pareto Evo- I. Rechenberg, H.P. Schwefel (Eds.), Parallel Problem Solving From Nature- 1076
1002 lutionary Algorithm, Swiss Federal Institute Technology, Zurich, Switzerland, PPSN IV, Lecture Notes in Computer Science, Springer-Verlag, Berlin, 1996, pp. 1077
1003 2001. 504–510. 1078
1004 [7] J.D. Knowles, D.W. Corne, Approximating the nondominated front using the [34] P.M. Pradhan, G. Panda, Solving multiobjective problems using cat swarm opti- 1079
1005 Pareto archived evolution strategy, Evol. Comput. 8 (2) (2000) 149–150. mization, Expert Syst. Appl. 39 (2012) 2956–2964. 1080
1006 [8] C.A.C. Coello, G.T. Pulido, Multiobjective optimization using a micro-genetic [35] N. Srinivas, K. Deb, Multi objective function optimization using nondominated 1081
1007 algorithm, in: L. Spector, E.D. Goodman, A. Wu, W.B. Langdon, H.M. Voigt, M. sorting genetic algorithms, Evol. Comput. 2 (3) (1995) 221–230. 1082
1008 Gen, S. Sen, M. Dorigo, S. Pezeshk, M.H. Garzon, E. Burke (Eds.), San Francisco: [36] T.T. Binh, U. Korn, MOBES: a multiobjective evolution strategy for constrained 1083
1009 Proc. Genetic and Evolutionary Computation Conf. (GECCO’2001), 2001, pp. optimization problems, in: 3rd International Conference on Genetic Algo- 1084
1010 274–282. rithms, 1997, pp. 176–182. 1085
1011 [9] K. Deb, Multi-objective genetic algorithms: problem difficulties and construc- [37] A. Osyczka, S. Kundu, A new method to solve generalized multicriteria opti- 1086
1012 tion of test problems, Evol. Comput. 7 (2002) 205–210. mization problems using the simple genetic algorithm, Struct. Multidiscip. 1087
1013 [10] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multi objective Optim. 10 (1995) 94–100. 1088
1014 genetic algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2) (2002) 182–190. [38] A. Farhang-Mehr, S. Azarm, Entropy-based multi-objective genetic algorithm 1089
1015 [11] C.A.C. Coello, G.T. Pulido, M.S. Lechuga, Handling multiple objectives with par- for design optimization, Struct. Multidiscip. Optim. 24 (2002) 351–360. 1090
1016 ticle swarm optimization, IEEE Trans. Evol. Comput. 8 (3) (2004) 256–260. [39] A. Kurpati, S. Azarm, J. Wu, Constraint handling improvements for multiobjec- 1091
1017 [12] F. Freschi, M. Repetto, VIS: an artificial immune network for multi-objective tive genetic algorithms, Struct. Multidiscip. Optim. 23 (2002) 204–210. 1092
1018 optimization, Eng. Optim. 38 (8) (2006) 975–980. [40] L.V. Santana-Quintero, C.A.C. Coello, An algorithm based on differential evolu- 1093
1019 [13] M.J. Reddy, D.N. Kumar, Multi-objective particle swarm optimization for gen- tion for multi-objective problems, Int. J. Comput. Intell. Res. 1 (2005) 151–160. 1094
1020 erating optimal trade-offs in reservoir operation, Hydrol. Process. 21 (2007) [41] J. Wu, Quality Assisted Multiobjective and Multidisciplinary Genetic Algo- 1095
1021 2897–2909. rithms (PhD thesis), Department of Mechanical Engineering, University of 1096
1022 [14] J.Q. Gao, L. Wang, WBMOAIS: a novel artificial immune system for multiobjec- Maryland at College Park, College Park, MD, 2001. 1097
1023 tive, Comput. Oper. Res. 37 (1) (2010) 50–60. [42] T. Ray, K.M. Liew, A swarm metaphor for multiobjective design optimization, 1098
1024 [15] B. Suman, N. Hoda, S. Jha, Orthogonal simulated annealing for multiobjective Eng. Optim. 34 (2002) 141–150. 1099
1025 optimization, Comput. Chem. Eng. 34 (2010) 1618–1631. [43] K. Deb, A. Pratap, S. Moitra, Mechanical component design for multiple objec- 1100
1026 [16] J. Gao, J. Wang, A hybrid quantum-inspired immune algorithm for multiobjec- tives using elitist non-dominated sorting GA, Parallel Probl. Solv. Nat. PPSN VI 1101
tive optimization, Appl. Math. Comput. 217 (2011) 4754–4770. 1917 (2000) 859–868. 1102

Please cite this article in press as: A. Sadollah, et al., Water cycle algorithm for solving constrained multi-objective optimization problems,
Appl. Soft Comput. J. (2014), http://dx.doi.org/10.1016/j.asoc.2014.10.042