Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
a r t i c l e i n f o a b s t r a c t
Article history: In this paper, modified schemes are proposed for preventing a grey wolf optimizer (GWO) from pre-
Received 26 February 2017 mature exploration and convergence on optimization problems. Three novel strategies are developed to
Revised 4 May 2017
improve the performance of existing GWO. The first strategy uses the concept of prey weight. The sec-
Accepted 14 May 2017
ond strategy uses the astrophysics concepts, which guide the grey wolves toward more promising areas
Available online 22 May 2017
of the search space. The beauty of this strategy is to let each grey wolf learn from not only movement
Keywords: of sun (symbolizes prey) in the search space but also the wolves are made to explore and exploit si-
Meta-heuristics multaneously. Third strategy combines the both, first and second strategies to take advantages of prey
Grey wolf optimizer weight and astrophysics strategies. The proposed improvements in GWO have been evaluated on thirteen
Astrophysics concept benchmark test functions. The performance of the proposed modifications has been compared with other
Function optimization five recently developed state-of-the-art techniques. The effects of scalability, noise, and control parame-
Constrained design problems
ter have also been investigated. The statistical tests have been performed to validate the significance of
modified variants. The proposed variants are also applied for seven well-known constrained engineering
design problems. The experimental results depict the supremacy of the proposed modified algorithm in
solving engineering design problems when compared with several existing techniques.
© 2017 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.advengsoft.2017.05.008
0965-9978/© 2017 Elsevier Ltd. All rights reserved.
232 V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254
solving optimization problems and the algorithm thus proposed is It is assumed thatα , β , and δ have better knowledge of the po-
called Modified Grey Wolf Algorithm-II (MGWO-II). This learning sition of the prey. Hence, these three solutions are saved for hunt-
strategy uses the elliptical orbit concept for wolves. The ellipti- ing mechanism. The position of ω is updated based on the posi-
cal orbit is responsible for both exploration and exploitation. This tions of α , β , and δ . These are mathematically modelled as follows
astrophysics-based learning strategy is incorporated in GWO to [3].
speed up the search and increase the efficiency of search process.
The third strategy combines the both above-mentioned strategies Distα = |C1 × Yα − Y | Distβ = C2 × Yβ − Y Distδ = |C3 × Yδ − Y |
and hence the algorithm called Modified Grey Wolf Algorithm- (5)
III (MGWO-III). The performance of the proposed modifications in
GWO has been evaluated on thirteen benchmark test functions.
These are compared with the original version of GWO and five Y1 = Yα − A1 × Distα Y2 = Yβ − A2 × Distβ Y3 = Yδ − A3 × Distδ
recently developed metaheuristic techniques. The effects of scal- (6)
ability and noise have also been investigated over three variants
Y1 + Y2 + Y3
of GWO. The different values of control parameters have also been Y3 = (7)
studied. The proposed approach is also employed for engineering 3
structural design problems. Fig. 1(a) shows the flowchart of GWO.
The remainder of this paper is organized as follows.
Section 2 describes the concepts of GWO and introduces an 2.2. Related works
overview of related works. The details of proposed modifications
are presented in Section 3. Section 4 presents the benchmark test In the past few years, it has been observed that grey wolf op-
functions, experimental setups and results obtained. Section 5 in- timizer (GWO) has been adopted by many researchers to solve the
troduces the application of proposed algorithms on structural optimization problems. A number of improvements in GWO were
design problems. Eventually, Section 6 draws the conclusions from also proposed from time to time to increase its performance.
the experimental results. Kamboj [9] developed a novel approach combining both PSO
and GWO. This approach suffers from premature convergence
2. Background problem. Jitkongcheun [8] used an invasion-based migration oper-
ation, which was integrated with GWO. This migration was respon-
In this section, we first describe the basics of GWO followed by sible for information exchange within population and produced
brief description of recently developed modifications in GWO. new individuals. However, it fails to converge in problems that
possess high dimensionality. Saremi et al. [23] utilized the evolu-
2.1. Classical Grey wolf optimizer tionary population dynamics (EPD) concept in GWO. EPD was used
to remove the poor search agents of GWO and reposition them to
Grey wolf optimizer algorithm (GWO) was developed by Mir- enhance the exploitation. However, it requires random positioning
jalili et al. [15]. It is a recently developed metaheuristic algorithm mechanism to avoid the local optima. Zhang and Zhou [28] pre-
that mimics the leadership hierarchy and hunting behavior of grey sented an extended GWO algorithm based on Powell local opti-
wolves in nature. The grey wolves are categorized into four groups mization method named as PGWO. The convergence property of
such as α , β , δ , and ω wolves [23]. These are involved in simu- Powell’s method was utilized to improve the performance of GWO.
lating the leadership hierarchy. The α wolves are leading wolves. Zhu et al. [29] presented a hybridization of GWO and differential
They are responsible for taking decisions during the hunting pro- evolution (DE). This approach is known as hybrid GWO (HGWO).
cess. The second level of dominated wolves, in the group, is β They have utilized the search capability of DE to remove the stag-
wolves. When α wolves die or become old, then they are upgraded nation problem and to improve the convergence speed.
to α wolves. The δ wolves control the ω wolves and provide the Kishor and Singh [12] provided an empirical study of classical
information to α and β wolves. The lowest level in wolves’ hierar- GWO. They proposed a modification in GWO, which improves the
chy is ω wolves. Besides social hierarchy, they have an interesting information sharing mechanism of GWO and also improves its per-
hunting behaviour. The main steps of hunting process are: search- formance. However, it suffers from premature convergence prob-
ing for prey, encircling the prey, and attacking the prey. The fol- lem. Mittal et al. [19] developed a modified GWO (mGWO) to im-
lowing two equations model the encircling behaviour of the grey prove the performance of GWO. They have introduced exponen-
wolves [15]. tial decay function for fine tuning of parameters. The decay func-
tion utilizes the coefficient parameter, A. However, it suffers from
Dist = |C × YP (tcur ) − Y (tcur )| (1) premature convergence problem. Emary et al. [4] proposed a bi-
nary GWO (BGWO) for feature selection. They used two approaches
for converting the classical GWO into binary form. They developed
Y (tcur + 1 ) = YP (tcur ) − A × Dist (2)
a classification-based fitness function to eliminate the redundant
where Dist represents the distance between the position of both data, which was optimized through BGWO.
the prey (YP ) and a grey wolf (Y). tcur is the current iteration. Wen et al. [25] presented a hybrid GWO which combines
YP (tcur ) represents the position vector of prey. A and C are the co- chaotic mapping and elite opposition based learning strategy to
efficient vectors calculated as follows: improve the performance. The chaotic sequence was used to ini-
tiate position of individuals. The elite opposition based learning
A = 2 × au × rand1 − au (3) strategy is applied to elite individuals for better exploration. Wen
[26] designed a nonlinear adjustment strategy of control parame-
C = 2 × rand2 (4) ter au in GWO, known as improved GWO (IGWO). The opposition-
based learning strategy is introduced to initialize the population.
The value of au is linearly decreased from 2 to 0 during the However, it fails to converge in high dimensionality problems.
iterations. The rand1 and rand2 are random variables and may have Emary et al. [5] proposed a new variant of GWO (EGWO) that
any value in the range of [0, 1]. The exploration and exploitation uses reinforcement learning principles with neural networks, to
decisions are made based on the value of vector A. enhance the performance. The reinforcement learning is used to
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 233
Fig. 1. Flowchart of Grey wolf algorithms: (a) GWO (b) MGWO-I (c) MGWO-II (d) MGWO-III.
234 V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254
set the parameter’s value based on individual agent. The neural Table 1
Unimodal benchmark test functions.
network is used to map a set of agents’ states to corresponding
actions that influence the exploration rate. The main drawback of Benchmark Test Function Search Range Dim.
this technique is that it employs complex methodology.
D
F1 (Y ) = y2i [ − 100, 100] 30
Although these works maintain a balance between intensifi- i=1
cation and diversification, still these are far from yielding opti- D D
F2 (Y ) = |yi | + i=1 |yi | [ − 10, 10] 30
mal solution. Hence, there is a scope to think of some new con- i=1
2
D
i
cepts and/or strategies that may help provide a better balance be- F3 (Y ) = ( yj) [ − 100, 100] 30
i=1 j=1
tween the two (intensification and diversification). This motivates
F4 (Y ) = max{|yi |, 1 ≤ i ≤ D} [ − 100, 100] 30
us to develop a modified algorithm based on new proposed con- D
i
−1
2
F5 (Y ) = [100(yi+1 − ) + ( yi − 1 ) ]
2
cepts/strategies. y2i [ − 30, 30] 30
i=1
D
3.2. Proposed approach 1: prey weight strategy 3.2.2. Modified Grey wolf Algorithm-I
The modified grey wolf algorithm-I (MGWO-I) follows the basic
The value of the important control parameter C, responsible for steps of GWO as mentioned in Section 2.1. The above-mentioned
exploration process, indicates the weight of the prey in distance proposed equation is introduced in Step 2. The Eq. (4) is replaced
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 235
Table 2
Multimodal benchmark test functions.
D
F8 (Y ) = −yi sin( |yi | ) [ − 50 0, 50 0] 30
i=1
D
F9 (Y ) = [y2i − 10 cos(2π yi ) + 10] [ − 5.12, 5.12] 30
i=1
D
D
F10 (Y ) = −20 exp −0.2 1
D
y2i − exp 1
D
cos(2π yi ) + 20 + e [ − 32, 32] 30
i=1 i=1
D D
F11 (Y ) = 1
40 0 0
Yi2 − i=1 cos y
√i +1 [ − 60 0, 60 0] 30
i
i=1
D
−1
F12 (Y ) = πD {10 sin(π z1 ) + (zi − 1 )2 [1 + 10sin2 (π zi+1 )]}.....
i=1
D
..... + πD { (zD − 1 ) } + u(yi , 10, 100, 4)
2
i=1
[ − 50, 50] 30
D
..... + 0.1{ (yD − 1 ) [1 + sin (2π yD )]} + u(yi , 5, 100, 4)
2 2
i=1
with the proposed Eq. (8). The flowchart of MGWO-I is shown in Here, Ri, α indicates the Euclidean distance computed between
Fig. 1(b). The proposed approach of MGWO-I consists of following Yi and Yα . U( − 2, 2) represents uniformly distributed random num-
the steps. ber in the interval [ − 2, 2]. tcur is the current iteration. The reason
behind using U( − 2, 2) is that the proposed algorithm exploits if
Algorithm (MGWO-I).
U( − 2, 2) returns value near to 1, otherwise it explores.
Step 1. Initialize the algorithm parameters, such as number of gray The positions of K grey wolves (Yi , i = 1, ...., K) are updated only
wolves (or search agent), maximum number of iterations when their new positions are better than the original positions,
(MaxIteration), and the control coefficient (au). described as under:
Step 2. The positions of grey wolves are initialized randomly in the
Yi, new (tcur + 1 ), i f F it (Yi, new ) < F it (Yi )
search space. Yi (tcur + 1 ) = (11)
Yi (tcur ), otherwise
Step 3. Compute the fitness value of each search agent which rep-
resents the distance between wolf and prey.
Step 4. Repeat the following steps until the maximum number of 3.3.1. Justification of proposed astrophysics strategy
iterations is reached: The rationale behind the proposed astrophysics strategy is that
(a) Based on the fitness values, the best (Yα ), second best planets move round the Sun in elliptical orbit. Due to this, the
(Yβ ), and third best (Yδ )solutions are identified. planets have different positions with respect to Sun at different
(b) Compute control coefficients parameters i.e., A and C us- times. Some planets are much closer to the Sun than others mean-
ing Eqs. (3) and (8) respectively. ing thereby that they (the closer ones) are best candidates for ex-
(c) Modify the position of grey wolves using Eq. (7). ploitation. Those, far from the Sun, are in better position to explore
(d) Update the fitness value using the modified position of whole search space. The elliptical orbital shape, in itself, is respon-
the grey wolves. sible for simultaneous exploration and exploitation. Here, the Sun
Step 5. The best grey wolf (Yα ) will yield the optimal solution at is analogous to the prey and planets are analogous to grey wolves.
the final iteration. Fig. 4 depicts how the elliptical orbit concept is responsible for
both exploration and exploitation in GWO.
3.3. Proposed approach 2: astrophysics strategy
Fig. 4. Concept of astrophysics strategy (a) 2D position of solar system, (b) 2D position of grey wolves, (c) 3D position of solar system, (d) 3D position of grey wolves.
(d) Update the fitness values according to the modified posi- 3.4. Proposed approach 3: combined strategy
tions of the grey wolves.
(e) Apply astrophysics strategy, as mentioned in Fig. (3), on The above-mentioned strategies, i.e., prey weight and astro-
the positions of grey wolves. physics, are combined to take the advantages of both strategies.
The combined strategy provides balance between exploration and
exploitation. This combined strategy is known as Modified GWO-III
Step 5. The best grey wolf (Yα )will yield the optimal solution at (MGWO-III).
the final iteration.
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 237
The complexity analysis of MGWO-III is presented in this sub- 4.3. Experiment 1: performance comparison
section. Both time and space complexities of MGWO-III is de-
scribed below. To illustrate the performance of MGWO’s variants, its results are
compared with six above-mentioned techniques on benchmark test
3.5.1. Time complexity functions stated in Section 4.1. The dimension of search space is
1. Initialization of MGWO-III requires O(n × d) time where n indi- set to 30 for this comparison, as mentioned in literature.
cates the number of wolves and d is the dimension of the test Tables 4 and 5 represent the performance comparisons of the
function. proposed MGWO’s variants with above-mentioned existing tech-
2. Control parameter calculation and position update steps of niques for unimodal and multimodal benchmark test functions re-
MGWO-III requires O(n × d) each. spectively. The results have been compared in terms of ’mean’ and
3. Fitness evaluation requires O(n × d) time. ’standard deviation’ over 30 independent simulation runs for each
4. Computation of astrophysics-based strategy requires O(K × d) case. The results reveal that the proposed MGWO-III is able to de-
time where K indicates the number of selected wolves. termine the optimal solution for all unimodal test functions in al-
most every run exceptF6 . For F1 − F4 functions, it has been found
Therefore, summing up the complexities of all the above
that MGWO-II and MGWO-III outperform all the above-mentioned
steps and considering that K n, the total time complexity
techniques. For F5 function, MGWO-III provides better results than
becomes O(n × d) per generation. The total time complexity of
the other competitive techniques. GSA provides much better results
MGWO-III for maximum number of iterations is O(n × d × MaxItera-
than the other competitive techniques for F6 function. MGWO-III is
tion). Here MaxIterationindicates the maximum number of genera-
the third best algorithm after PSO. For F7 function, MGWO-III per-
tions/iterations.
forms better than GSA, PSO, MVO, MFO, SCA, GWO, MGWO-I and
3.5.2. Space complexity MGWO-II.
The space requirement of MGWO-III is due to its initialization It has been observed from results presented in Table 5 that
process. Thus, the total space complexity of MGWO-III is O(n × d). MGWO-III is able to find the optimal solution for almost all mul-
timodal test function in every run. For F8 function, MGWO-III
4. Experimental results and discussion outperforms the other techniques. For F9 , F10 , and F11 functions,
MGWO-III and MGWO-II perform better than the other competi-
This section compares the performance of MGWO’s variants tive techniques. For F12 function, MGWO-III provides better results
with six other competitive algorithms and validates over thirteen than the other techniques. GWO and MGWO-I provide comparable
well-known benchmark test functions. These benchmark test func- results. For F13 function, GSA is superior to other algorithms such
tions are as follows: as PSO, MVO, MFO, SCA, GWO and MGWO’s variants.
238 V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254
Table 4
Results on unimodal benchmark test functions.
F1 1.05E−16 3.92E−10 2.78E−01 1.02E−04 1.33E−03 8.01E−59 1.54E−59 0.0 0E+ 0 0 0.0 0E+ 0 0
(6.10E−17) (1.40E−08) (1.11E−01) (5.99E−04) (1.06E−01) (1.85E−58) (4.77E−58) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F2 5.31E−08 1.03E−04 3.44E−01 4.00E+ 01 3.53E−06 1.23E−34 8.58E−35 0.0 0E+ 0 0 0.0 0E+ 0 0
(9.29E−01) (1.84E−03) (1.41E−01) (2.16E+ 01) (8.57E−05) (2.80E−34) (6.69E−35) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F3 3.96E+ 02 1.17E+ 01 4.26E+ 01 3.07E+ 03 4.07E+ 03 9.22E−15 8.06E−15 0.0 0E+ 0 0 0.0 0E+ 0 0
(1.56E+ 02) (7.13E+ 00) (8.97E+ 00) (3.71E+ 03) (8.21E+ 00) (3.29E−14) (1.92E−14) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F4 8.85E−01 5.79E−01 9.35E−01 6.76E+ 01 1.71E+ 01 1.65E−14 1.51E−14 0.0 0E+ 0 0 0.0 0E+ 0 0
(9.89E−01) (1.72E−01) (2.50E−01) (1.06E+ 01) (8.21E+ 00) (2.02E−14) (2.89E−14) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F5 2.75E+ 01 2.80E+ 01 3.26E+ 01 6.28E+ 02 5.32E+ 01 2.70E+ 01 2.70E+ 01 1.79E+ 01 1.63E+ 01
(3.47E+ 01) (3.89E+ 01) (1.43E+ 02) (3.98E+ 03) (1.98E+ 03) (7.84E−01) (6.63E−01) (1.20E+ 01) (1.25E+ 01)
F6 9.60E−17 1.27E−09 2.98E−02 1.02E−04 4.62E+ 00 5.97E−01 5.65E−01 2.09E−05 2.04E−05
(4.00E−17) (1.78E−08) (9.98E−02) (2.01E−04) (9.75E−01) (3.67E−01) (3.81E−01) (9.16E−06) (6.76E−06)
F7 5.23E−02 6.40E−02 1.92E−02 1.89E−01 2.04E−02 8.38E−04 7.30E−04 3.16E−05 3.10E−05
(2.77E+ 00) (2.87E−02) (7.43E−03) (2.93E−01) (5.79E−02) (5.87E−04) (4.02E−04) (2.72E−05) (2.76E-05)
Table 5
Results on multimodal benchmark test functions.
4.4. Experiment 2: statistical analysis most all benchmark test functions. Therefore, MGWO’s variants are
statistically significant from the existing algorithms for both uni-
Besides basic statistical analysis (i.e., mean and standard devia- modal and multimodal benchmark test functions.
tion), ANOVA test has been conducted for comparison of GSA, PSO,
MVO, MFO, SCA, GWO and MGWO’s variants. It is used to deter- 4.5. Experiment 3: scalability study
mine whether the proposed MGWO’s variants are statistical signifi-
cant or not. A p-value determines the significance level of MGWO’s The next experiment is carried out to perceive the effect of scal-
variants. An algorithm is statistically significant if and only if the ability on above-mentioned algorithms for all benchmark test func-
p-value is less than 0.05. The result analysis of the ANOVA test for tions. The dimensionality of benchmark test functions is made to
the benchmark test functions is tabulated in Table 6. vary as 30, 50, 70, 90 and 100. Tables 7 and 8 depict the perfor-
The proposed variants are statistical different from the above- mance of algorithms on scalable unimodal and multimodal bench-
mentioned algorithms for F1 function. The p-value obtained from mark test functions. For F1 − F4 functions, MGWO-II and MGWO-III
MGWO’s variants is less than 0.05. For F2 function, MGWO’s vari- provide same results over all dimensions. This may be attributed
ants are statistically significant as compared to other algorithms. to the fact that they have better exploration and exploitation ca-
For F3 − F6 functions, the MGWO’s variants are statistically signifi- pability owing to the proposed approaches. For F5 − F7 functions,
cant than the existing algorithms. For F7 function, the p-value ob- the performance of MGWO’s variants decreases with increase in
tained from MGWO’s variants is smaller than 0.05. MGWO-II and dimension. The results also show that the degradation in the per-
MGWO-III are statistically different from the existing algorithms. formance of GSA, PSO, MVO, MFO, SCA, and GWO techniques with
MGWO-II and MGWO-III are statistically similar for F1 − F7 func- increase in dimensionality of search space for all unimodal bench-
tions. mark test functions.
For F8 function, MGWO-III is significantly different from the The results mentioned in Table 8 reveal that all the above-
existing algorithms (i.e., GSA, PSO, MVO, MFO, SCA, GWO, and mentioned algorithms namely GSA, PSO, MVO, MFO, SCA, GWO,
MGWO-I). MGWO-II is not statistically different from MGWO-I and and MGWO’s variants improve their performance with increase
MGWO-III. For F9 and F10 functions, the p-value obtained from in dimensionality forF8 . However, MGWO-III provides better re-
MGWO-II and MGWO-III is 0.00. Hence, MGWO-II and MGWO- sults than the other competitive algorithms. The performance of
III are statistically significant from other existing algorithms. For MGWO-II and MGWO-III is not influenced by scalable search spaces
F11 function, MGWO’s variants are statistically different from ofF9 , F10 and F11 . It has also been noticed that the performance
the above-mentioned algorithms except MFO. For F12 function, of GSA, PSO, MVO, MFO, SCA, and GWO degrades with increase
MGWO’s variants are statistically significant from the all above- in dimension. For these functions, MGWO-I and GWO are the sec-
mentioned algorithms except SCA. MGWO’s variants are statisti- ond and third best algorithms among all used for comparison. For
cally similar for F13 function. F12 and F13 functions, the performance of GSA, PSO, MVO, MFO,
The results demonstrate that the p-value, obtained using SCA, GWO and MGWO’s variants reduces with increase in dimen-
MGWO’s variants (i.e., MGWO-I, MGWO-II and MGWO-III) with re- sionality of search space.
spect to the existing algorithms, is much smaller than 0.05 for al- The results illustrated that the performance of MGWO’s vari-
ants is not much affected with increase in dimensionality of search
Table 6
Anova test results on benchmark test functions.
p-value Algorithms for comparison
GSA PSO MVO MFO SCA GWO MGWO-I MGWO-II MGWO-III
F1 0.0 0E+ 0 0 MVO, MFO, SCA, MVO, MFO, SCA, GSA, PSO, MFO, SCA, GSA, PSO, MVO, SCA, GSA, PSO, MVO, GSA, PSO, MVO, GSA, PSO, MVO, GSA, PSO, MVO, GSA, PSO, MVO,
GWO, MGWO-I, GWO, MGWO-I, GWO, MGWO-I, GWO, MGWO-I, MFO, GWO, MGWO-I, MFO, SCA, MGWO-II, MFO, SCA, MGWO-II, MFO, SCA,GWO, MFO, SCA,GWO,
MGWO-II, MGWO-III MGWO-II, MGWO-III MGWO-II, MGWO-III MGWO-II, MGWO-III MGWO-II, MGWO-III MGWO-III MGWO-III MGWO-I MGWO-I
F2 0.0 0E+ 0 0 MFO MVO, MFO, SCA, PSO, MFO, SCA, GSA, PSO, MVO, SCA, PSO, MVO, MFO, PSO, MVO, MFO, PSO, MVO, MFO, PSO, MVO, MFO, PSO, MVO, MFO,
239
240 V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254
Table 7
Performance of algorithms on high dimensional unimodal benchmark functions.
Dim GSA PSO MVO MFO SCA GWO MGWO-I MGWO-II MGWO-III
F1 30 1.05E-16 3.92E-10 9.21E−01 1.02E−04 1.33E−03 8.01E−59 1.54E−59 0.0 0E+ 0 0 0.0 0E+ 0 0
(6.10E-17) (1.40E−08) (1.27E−01) (5.99E−04) (1.06E−01) (1.85E−58) (4.77E−58) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
50 1.06E−04 9.12E−04 6.81E+ 00 9.33E+ 03 8.04E+ 01 1.05E−43 1.18E−58 0.0 0E+ 0 0 0.0 0E+ 0 0
(5.86E−05) (9.39E−04) (1.26E+ 00) (1.11E+ 04) (1.36E+ 02) (1.46E−43) (2.26E−58) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
70 3.32E−04 1.17E−01 2.79E+ 01 1.35E+ 04 1.19E+ 03 6.55E−36 6.45E−49 0.0 0E+ 0 0 0.0 0E+ 0 0
(1.34E−04) (8.98E−02) (5.22E+ 00) (1.07E+ 04) (1.14E+ 03) (1.31E−35) (1.46E−48) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
90 1.00E+ 01 1.21E+ 00 6.83E+ 01 2.16E+ 04 4.05 + 03 3.88E−31 7.18E−43 0.0 0E+ 0 0 0.0 0E+ 0 0
(4.52E+ 01) (7.61E−01) (1.24E+ 01) (1.04E+ 04) (3.56E+ 03) (4.77E−31) (1.03E−42) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
100 5.76E+ 02 2.97E+ 00 1.19E+ 02 3.51E+ 04 6.10E+ 03 2.09E−29 1.06E−40 0.0 0E+ 0 0 0.0 0E+ 0 0
(3.05E+ 02) (1.69E+ 00) (2.28E+ 01) (1.22E+ 04) (4.20E+ 03) (2.79E−29) (1.64E−40) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F2 30 5.31E−08 1.03E−04 9.95E−01 4.00E+ 01 3.53E−06 1.23E−34 8.58E−35 0.0 0E+ 0 0 0.0 0E+ 0 0
(9.29E−01) (1.84E−03) (1.27E+ 00) (2.16E+ 01) (8.57E−05) (2.80E−34) (6.69E−35) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
50 6.23E−02 1.50E−01 4.26E+ 02 7.22E+ 01 1.27E−02 5.30E−26 2.18E−34 0.0 0E+ 0 0 0.0 0E+ 0 0
(8.66E−03) (1.17E−01) (1.39E+ 03) (3.29E+ 01) (1.63E−02) (4.99E−26) (2.46E−34) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
70 1.38E−01 1.66E+ 00 4.25E+ 09 1.16E+ 02 3.21E−01 7.53E−22 4.50E−29 0.0 0E+ 0 0 0.0 0E+ 0 0
(3.11E−02) (8.52E−01) (1.70E+ 10) (3.88E+ 01) (5.19E−01) (4.48E−22) (5.45E−29) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
90 2.45E−01 6.29E+ 00 6.02E+ 17 1.52E+ 02 9.55E−01 5.81E−19 8.71E−26 0.0 0E+ 0 0 0.0 0E+ 0 0
(4.67E−02) (2.41E+ 00) (3.28E+ 18) (5.34E+ 01) (9.71E−01) (3.68E−19) (7.29E−26) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
100 9.90E+ 00 1.18E+ 01 1.47E+ 23 1.67E+ 02 1.46E+ 00 5.29E−18 1.50E−24 0.0 0E+ 0 0 0.0 0E+ 0 0
(4.76E+ 00) (3.61E+ 00) (5.62E+ 23) (4.77E+ 01) (1.84E+ 00) (3.15E−18) (1.16E−24) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F3 30 3.96E+ 02 1.17E+ 01 1.38E+ 02 3.07E+ 03 4.07E+ 03 9.22E−15 8.06E−15 0.0 0E+ 0 0 0.0 0E+ 0 0
(1.56E+ 02) (7.13E+ 00) (4.85E+ 01) (3.71E+ 03) (8.21E+ 00) (3.29E−14) (1.92E−14) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
50 1.19E+ 02 6.26E+ 02 4.67E+ 03 4.80E+ 04 3.27E+ 04 3.73E−06 2.88E−09 0.0 0E+ 0 0 0.0 0E+ 0 0
(5.42E+ 01) (1.69E+ 02) (1.29E+ 03) (1.59E+ 04) (1.19E+ 04) (1.32E−05) (1.52E−08) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
70 7.16E+ 02 2.89E+ 03 1.92E+ 04 9.97E+ 04 8.79E+ 04 9.73E−03 2.63E−04 0.0 0E+ 0 0 0.0 0E+ 0 0
(2.35E+ 02) (7.80E+ 02) (3.33E+ 03) (3.99E+ 04) (2.27E+ 04) (2.26E−02) (1.13E−03) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
90 1.66E+ 03 7.76E+ 03 4.64E+ 04 1.61E+ 05 1.63E+ 05 2.58E+ 00 3.61E−03 0.0 0E+ 0 0 0.0 0E+ 0 0
(4.36E+ 02) (1.54E+ 03) (5.90E+ 03) (5.00E+ 04) (2.95E+ 04) (6.37E+ 00) (1.15E−02) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
100 8.77E+ 03 1.12E+ 04 6.14E+ 04 1.76E+ 05 2.04E+ 05 6.92E+ 00 3.97E−01 0.0 0E+ 0 0 0.0 0E+ 0 0
(2.00E+ 03) (3.16E+ 03) (8.42E+ 03) (4.92E+ 04) (4.79E+ 04) (1.21E+ 01) (1.73E+ 00) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F4 30 8.85E−01 5.79E−01 1.59E+ 00 6.76E+ 01 1.71E+ 01 1.65E−14 1.51E−14 0.0 0E+ 0 0 0.0 0E+ 0 0
(9.89E−01) (1.72E−01) (4.56E−01) (1.06E+ 01) (8.21E+ 00) (2.02E−14) (2.89E−14) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
50 7.21E+ 00 2.26E+ 00 1.52E+ 01 8.38E+ 01 5.73E+ 01 2.21E−09 3.11E−15 0.0 0E+ 0 0 0.0 0E+ 0 0
(1.51E+ 00) (3.31E−01) (5.77E+ 00) (4.11E+ 00) (8.0 0E+ 0 0) (3.57E−09) (1.08E−14) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
70 1.19E+ 01 4.52E+ 00 3.70E+ 01 9.02E+ 01 7.52E+ 01 1.91E−06 1.35E−11 0.0 0E+ 0 0 0.0 0E+ 0 0
(1.77E+ 00) (7.45E−01) (7.67E+ 00) (2.81E+ 00) (6.19E+ 00) (2.68E−06) (2.74E−11) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
90 1.42E+ 01 8.29E+ 00 5.05E+ 01 9.23E+ 01 8.43E+ 01 2.82E−03 9.05E−05 0.0 0E+ 0 0 0.0 0E+ 0 0
(1.25E+ 00) (1.41E+ 00) (6.13E+ 00) (2.36E+ 00) (3.76E+ 00) (1.27E−02) (5.18E−04) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
100 1.55E+ 01 9.06E+ 00 5.71E+ 01 9.44E+ 01 8.60E+ 01 2.15E−03 6.34E−06 0.0 0E+ 0 0 0.0 0E+ 0 0
(1.55E+ 00) (1.61E+ 00) (6.86E+ 00) (1.80E+ 00) (2.71E+ 00) (3.32E−03) (2.71E−05) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F5 30 2.75E+ 01 2.80E+ 01 2.91E+ 02 6.28E+ 02 5.32E+ 01 2.70E+ 01 2.70E+ 01 1.79E+ 01 1.63E+ 01
(3.47E+ 01) (3.89E+ 01) (6.02E+ 02) (3.98E+ 03) (1.98E+ 03) (7.84E−01) (6.63E−01) (1.20E+ 01) (1.25E+ 01)
50 4.73E+ 01 1.62E+ 02 5.58E+ 02 5.36E+ 06 1.31E+ 06 4.73E+ 01 4.76E+ 01 3.00E+ 01 3.50E+ 01
(1.01E+ 01) (7.20E+ 01) (4.75E+ 02) (2.03E+ 07) (1.70E+ 06) (8.84E−01) (7.13E−01) (2.16E+ 01) (1.92E+ 01)
70 8.47E+ 01 5.87E+ 02 2.00E+ 03 2.25E+ 07 8.69E+ 06 6.74E+ 01 6.79E+ 01 5.17E+ 01 5.14E+ 01
(3.35E+ 01) (5.58E+ 02) (1.51E+ 03) (4.13E+ 07) (7.57E+ 06) (8.93E−01) (6.45E−01) (2.65E+ 01) (2.57E+ 01)
90 1.39E+ 02 1.46E+ 03 5.58E+ 03 2.42E+ 07 3.58E+ 07 8.76E+ 01 8.81E+ 01 6.13E+ 01 6.73E+ 01
(4.33E+ 01) (9.93E+ 02) (5.40E+ 03) (5.25E+ 07) (2.58E+ 07) (7.88E−01) (5.41E−01) (3.66E+ 01) (3.13E+ 01)
100 1.12E+ 04 1.93E+ 03 6.00E+ 03 2.74E+ 07 6.65E+ 07 9.75E+ 01 9.80E+ 01 8.21E+ 01 6.61E+ 01
(7.41E+ 03) (8.16E+ 02) (4.40E+ 03) (5.46E+ 07) (3.48E+ 07) (7.11E−01) (5.19E+ 01) (3.07E+ 01) (4.18E+ 01)
F6 30 9.60E−17 1.27E−09 9.30E−01 1.02E−04 4.62E+ 00 5.97E−01 5.65E−01 2.09E−05 2.04E−05
(4.00E−17) (1.78E−08) (2.31E−01) (2.01E−04) (9.75E−01) (3.67E−01) (3.81E−01) (9.16E−06) (6.76E−06)
50 1.15E−04 9.40E−04 6.57E+ 00 9.38E+ 03 7.73E+ 01 2.24E+ 00 4.50E+ 00 2.15E+ 00 2.65E+ 00
(4.75E−05) (1.30E−03) (1.43E+ 00) (9.78E+ 03) (9.80E+ 01) (6.85E−01) (6.69E−01) (2.11E+ 00) (2.91E+ 00)
70 3.66E−04 1.11E−01 2.78E+ 01 1.41E+ 04 1.45E+ 03 5.04E+ 00 7.81E+ 00 3.39E+ 00 3.27E+ 00
(1.70E−04) (7.50E−02) (4.97E+ 00) (9.83E+ 03) (1.95E+ 03) (8.55E−01) (6.69E−01) (3.39E+ 00) (4.21E+ 00)
90 5.09E−01 1.09E+ 00 7.08E+ 01 2.47E+ 04 3.40E+ 03 7.68E+ 00 1.14E+ 01 3.54E+ 00 7.22E+ 00
(1.68E+ 00) (6.13E−01) (1.14E+ 01) (1.54E+ 04) (2.21E+ 03) (8.37E−01) (6.76E−01) (4.88E+ 00) (6.74E+ 00)
100 6.03E+ 02 2.73E+ 00 1.12E+ 02 3.45E+ 04 3.99E+ 03 9.13E+ 00 1.34E+ 01 7.42E+ 00 6.91E+ 00
(2.83E+ 02) (1.62E+ 00) (1.75E+ 01) (1.34E+ 04) (3.27E+ 03) (9.94E−01) (1.03E+ 00) (6.28E+ 00) (7.81E+ 00)
F7 30 5.23E−02 6.40E−02 2.88E−02 1.89E−01 2.04E−02 8.38E−04 7.30E−04 3.16E−05 3.10E−05
(2.77E+ 00) (2.87E−02) (1.26E−02) (2.93E−01) (5.79E−02) (5.87E−04) (4.02E−04) (2.72E−05) (2.76E−05)
50 1.85E−01 4.16E−01 1.08E−01 2.24E+ 01 7.18E−01 1.45E−03 4.76E−04 4.13E−05 3.88E−05
(4.78E−02) (1.39E−01) (4.24E−02) (2.77E+ 01) (8.97E−01) (5.46E−04) (2.64E−04) (4.20E−05) (3.12E−05)
70 3.42E−01 6.65E+ 01 2.10E−01 5.11E+ 01 6.24E+ 00 2.04E−03 7.33E−04 3.37E−05 3.35E−05
(7.90E−02) (1.23E+ 02) (5.81E−02) (4.49E+ 01) (6.34E+ 00) (9.12E−04) (3.24E−04) (2.85E−05) (2.58E−05)
90 6.33E−01 1.08E+ 03 4.28E−01 1.15E+ 02 3.76E+ 01 2.23E−03 8.98E−04 3.72E−05 3.54E−05
(2.04E−01) (2.51E+ 02) (1.17E−01) (1.08E+ 02) (3.50E+ 01) (1.18E−03) (6.06E−04) (3.74E−05) (2.78E−05)
100 1.47E+ 01 1.37E+ 03 5.54E−01 1.30E+ 02 5.57E+ 01 2.38E−03 9.83E−04 3.85E−05 4.06E−05
(2.04E+ 01) (3.60E+ 02) (1.11E−01) (9.78E+ 01) (5.24E+ 01) (1.07E−03) (4.15E−04) (3.75E−05) (3.44E−05)
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 241
Table 8
Performance of algorithms on high dimensional multimodal benchmark functions.
Dim GSA PSO MVO MFO SCA GWO MGWO-I MGWO-II MGWO-III
Table 9
Results on noisy and high−dimensional unimodal benchmark test functions.
Table 10
Results on noisy and high−dimensional multimodal benchmark test functions.
of all the above-mentioned algorithms. The results obtained from 4.8. Experiment 6: sensitivity analysis
MGWO-II are far better than other four algorithms for F8 function.
For F9 − F13 functions, MGWO-III shows best results followed by The effect of au on the performance of MGWO’s variants has
MGWO-II, MGWO-I and GWO. The performance of GSA, PSO, MVO, also been investigated. The values of auused in experimentation
MFO, and SCA is affected to a greater extent with noisy benchmark are set to six different intervals such as [2, 0], [2.5, 0], [3, 0], [3.5,
test functions. 0], [4, 0], and [4.5, 0] while keeping the dimensionality of search
It has also been noticed that performance of MGWO-II and space as 30. Fig. 5 shows the performance of MGWO’s variants over
MGWO-III is least affected as compared to other techniques for different values of au.
noisy benchmark functions. This is attributed to the fact that the The result reveal that the MGWO-II and MGWO-III does
proposed approach has better exploration capability due to in- not affected from value of au for F1 , F2 , F3 , F4 , F9 , F10 and F11
corporation of the astrophysics concept. Due to this feature, it is functions. Therefore, a straight line is observed for the above-
able to determine the optimal value of fitness notwithstanding the mentioned functions. For F1 , F2 , F3 , F4 , F10 , and F11 functions,
noisy benchmark functions. performance of MGWO-I’s performance is greatly reduce with in-
crease in the value of au. The optimum value of these functions
4.7. Experiment 5: performance comparison with existing variants of for MGWO-I is obtained at au = 2. For F9 function, the value of fit-
GWO ness produced from MGWO-I is first decrease, then increase with
increase in the value of au. The optimum value of F9 is obtained at
In order to prove the supremacy of MGWO’s variants, these are au = 2.5.
compared with the five well-known existing algorithms. These are For F5 , F6 , F7 , F8 , and F13 functions, the performance of
HGWO [29], PGWO [28], IGWO [26], mGWO [19], and EGWO [5]. MGWO’s variants greatly reduce with increase in the value of au.
Tables 11 and 12 show the performance comparison of MGWO’s The optimum value of these functions is obtained at au = 2. For
variants with above-mentioned algorithms. F13 function, the value of this function produced from MGWO-I
For F1 − F4 functions, MGWO-II and MGWO-III provide better increases with increase in the value of au. The optimum value of
solution than the other competitive algorithms. EGWO performs this function is obtained at au = 2. For MGWO-II and MGWO-III,
better than MGWO-I and other existing variants of GWO. The re- the value of this function is first decreases until the value of au
sults obtained from MGWO-II and MGWO-III are far better than reaches at 2.5. Thereafter, its value increases with increase in value
other variants of GWO. For F8 function, EGWO provides best fit- of au. The optimum value of this function is obtained at au = 2.5.
ness value as compared to other techniques. For F9 − F11 functions, The results obtained from Fig. 5 reveal that the performance of
MGWO-II and MGWO-III perform much better than other tech- MGWO-II and MGWO-III is least affected as compared to MGWO-I
niques. EGWO is the third best algorithm that provides better so- for different values of au. The reason behind is that the MGWO-II
lution than the other competitive algorithms. MGWO-III is the best and MGWO-III have better exploration and exploitation capability
algorithm for solving F12 and F13 functions. MGWO-II is the second due to incorporate the astrophysics concept. Due to this feature, it
best algorithm for these functions.
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 243
Fig. 5. Continued
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 245
Fig. 5. Continued
246 V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254
Fig. 5. Continued
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 247
Fig. 5. Continued
Table 11
Performance comparison with existing variants of GWO on unimodal functions.
F1 2.92E−57 2.21E−58 2.42E−59 2.54E−58 1.12E−59 1.54E−59 0.0 0E+ 0 0 0.0 0E+ 0 0
(2.32E−57) (1.87E−58) (0.0 0E+ 0 0) (7.98E−53) (3.58E−59) (4.77E−58) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F2 9.33E−34 7.83E−35 8.84E−35 5.93E−34 7.41E−35 8.58E−35 0.0 0E+ 0 0 0.0 0E+ 0 0
(5.03E−33) (6.92E−32) (4.96E−35) (1.56E−32) (1.05E−34) (6.69E−35) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F3 5.68E−15 3.98E−15 3.74E−15 4.44E−15 2.93E−15 8.06E−15 0.0 0E+ 0 0 0.0 0E+ 0 0
(8.71E−14) (1.32E−13) (1.91E−15) (0.0 0E+ 0 0) (1.10E−16) (1.92E−14) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F4 2.27E−14 1.99E−14 1.69E−14 1.16E−14 1.00E−14 1.51E−14 0.0 0E+ 0 0 0.0 0E+ 0 0
(4.76E−14) (3.98E−14) (4.52E−14) (2.84E−14) (1.15E−14) (2.89E−14) (0.0 0E+ 0 0) (0.0 0E+ 0 0)
F5 2.98E+ 01 2.77E+ 01 2.69E+ 01 2.62E+ 01 1.98E+ 01 2.70E+ 01 1.79E+ 01 1.63E+ 01
(7.16E−01) (8.68E−01) (8.52E−01) (5.59E−01) (4.27E−01) (6.63E−01) (1.20E+ 01) (1.25E+ 01)
F6 8.69E−01 7.27E−01 7.86E−01 5.61E−01 2.21E−03 5.65E−01 2.09E−05 2.04E−05
(2.21E−01) (2.10E−01) (2.44E−01) (1.59E−01) (1.37E−02) (3.81E−01) (9.16E−06) (6.76E−06)
F7 6.79E−03 6.64E−03 6.09E−04 6.83E−04 4.36E−04 7.30E−04 3.16E−05 3.10E−05
(2.27E−02) (3.12E−02) (1.76E−04) (3.42E−04) (3.12E−04) (4.02E−04) (2.72E−05) (2.76E−05)
Table 12
Performance comparison with existing variants of GWO on unimodal functions.
is capable to determine the optimal value of fitness functions, even is being oval shaped. The prey weight control helps the approach
the value of au is significantly varied. The optimum value for most jump to more promising regions quickly, especially in complex
of the benchmark test functions is obtained at au = 2. search space. The results mentioned in Tables 4 and 5 demonstrate
that MGWO-III has almost zero standard deviation, which ensures
4.9. Result and discussion the robustness of the approach regardless of random factor used.
The performance on noisy and high dimensional test data also de-
The experimental results reveal that MGWO-III has better fit- picts that MGWO-III is a better performer than other algorithms.
ness function value that makes the approach to converge to better The MGWO-III, due to incorporation of the astrophysics concept,
optima. The improved performance is attributed to the fact that improvises superior balance between exploration and exploitation.
MGWO-III can exploit and explore well because of the search path
248 V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254
Table 14
Statistical results for different algorithms for three-bar truss design problem.
Table 13
Comparison of best solution obtained from different algorithms
for three−bar truss.
DW DC
Table 15 Table 18
Comparison of best solution obtained from different algorithms for welded beam. Statistical results for different algorithms for pressure vessel design problem.
Algorithm Optimal values for variables Optimal Cost Algorithm Best Mean Worst Std. Dev. Median
Table 16
Statistical results for different algorithms for welded beam problem.
Table 19
Comparison of best solution obtained from different algorithms for
spring problem.
DW DC NAC
TSh THd IR L observed from Table 18 that MGWO-III surpassed other algorithms
for offering the best solution in terms of standard deviation, mean,
MGWO-III 0.778197 0.384675 40.315493 199.95932 5884.0616
worst and median. MGWO-II and MGWO-I are the second and third
MGWO-II 0.778403 0.384823 40.331283 199.74989 5884.0810
MGWO-I 0.778181 0.384821 40.318142 199.93194 5884.2166 best algorithm respectively.
GWO 0.778535 0.384960 40.337793 199.65029 5884.3689
SCA 0.817577 0.417932 41.74939 183.57270 6137.3724
MVO 0.845719 0.418564 43.816270 156.38164 6011.5148 5.4. Constrained problem IV: tension/compression spring design
PSO 0.778961 0.384683 40.320913 20 0.0 0 0 0 0 5891.3879
GSA 1.085800 0.949614 49.345231 169.48741 11,550.2976 The tension/compression spring design problem (see
Appendix A.4), as shown in Fig. 9, is a minimization constrained
problem. The objective is to design a tension/compression spring
having minimum weight and satisfying constraints such as deflec-
shell (TSh ), thickness of the head (THd ), inner radius (IR), and length tion, shear stress, surge frequency, and limits on outer diameter
of cylindrical section of the vessel (L). There are four design vari- and design variables [1]. There are three design variables such
ables such as TSh ( = y1 ), THd ( = y2 ), IR( = y3 ), and L( = y4 ), as shown as the wire diameter Dw ( = y1 ), mean coil diameter DC ( = y2 ), and
in Fig. 8. number of active coils NAC ( = y3 ).
Table 17 represents the comparisons of different algorithms Table 19 shows the comparison of the best solutions for the six
for the best obtained solutions. As seen from Table 17, it is ob- optimization algorithms and proposed MGWO’s variants in terms
served that the proposed MGWO-III provides optimal solution at of design variables and function values. The statistical optimiza-
y∗ = (0.778197, 0.384675, 40.315493, 199.95932) with correspond- tion results for the reported algorithms are given in Table 20. From
ing fitness value equal to F(y∗ ) = 5884.0616. The simulation results Table 20, it is evident that the proposed MGWO-III produced the
reveal that MGWO-III provides better optimization than the other best solution with considerable improvement with respect to other
existing techniques. The statistical results thus obtained for pres- algorithms. The statistical results of MGWO-III in terms of mean,
sure vessel problem were compared and shown in Table 18. It is worst, median, and standard deviation were also best among all.
250 V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254
Table 20
Statistical results for different algorithms for spring design problem.
Table 21
Comparison of best solution obtained from different algorithms for Speed Reducer.
W M P L1 L2 D1 D2
The speed reducer (see Appendix A.4) is a gear box between Algorithm Best Mean Worst Std. Dev. Median
engine and propeller of airplane to permit each to rotate at ef- MGWO-III 2996.798 2997.978 2999.825 0.97880 2997.021
ficient speed, as demonstrated in Fig. 10. The aim of this design MGWO-II 2996.905 2998.517 30 0 0.739 1.10299 2997.223
problem is to minimize the weight of speed reducer subject to MGWO-I 2997.007 2998.927 3002.242 1.81536 2997.980
constraints on bending stress of gear teeth, transverse deflections GWO 2997.288 2999.640 3003.889 1.93193 2999.187
SCA 3030.563 3065.917 3104.779 18.0742 3065.609
of the shafts, surface stress, and stresses in the shafts [6]. There
MVO 3002.928 3028.841 3060.958 13.0186 3027.031
are seven design variables such as face width W( = y1 ), module of PSO 3005.763 3105.252 3211.174 79.6381 3105.252
teeth M( = y2 ), number of teeth on pinion P( = y3 ), length of the GSA 3051.120 3170.334 3363.873 92.5726 3156.752
first shaft between bearings L1 ( = y4 ), length of second shaft be-
tween bearings L2 ( = y5 ), diameter of first shaft D1 ( = y6 ) and di-
ameter of second shaft D2 ( = y7 ). The complexity of this problem is
sults, MGWO-III surpassed other reported algorithms for offering
high as compared to already mentioned problems. This is due to
the best solution.
the fact that it as higher number of constraints.
The optimization algorithms such as GSA, PSO, MVO, MFO, SCA,
GWO, and MGWO variants were applied to this problem. The com- 5.6. Constrained problem VI: rolling element bearing design
parison for the best solution given by such algorithms is presented
in Table 21. The best solution was obtained by MGWO-III at de- The aim of this design problem (See Appendix A.6) is to maxi-
sign variables y∗ = (3.50010, 0.7, 17, 7.300568, 7.806687, 3.35068, mize the dynamic load carrying capacity of a rolling element bear-
5.286903) with an objective function value of F(y∗ ) = 2996.798. The ing as demonstrated in Fig. 11. There are 10 decision variables
obtained statistical results for Speed Reducer design problem were such as pitch diameter Dm ( = y1 ), ball diameter Db ( = y2 ), number
compared and tabulated in Table 22. In terms of statistical re- of ballsZ( = y3 ), inner raceway curvature coefficient fi ( = y4 ), outer
raceway curvature coefficients fo ( = y5 ), KDmin ( = y6 ), KDmax ( = y7 ),
V. Kumar, D. Kumar / Advances in Engineering Software 112 (2017) 231–254 251
Table 23
Comparison of best solution obtained from different algorithms for Rolling Element Bearing.
Dm Db Z fi fo KDmin KDmax ε e ζ
MGWO-III 125.7051 21.42294 10.99905 0.515 0.515 0.4872 0.63834 0.3 0.07605 0.67328 85,525.305
MGWO-II 125.6730 21.42199 10.99759 0.515 0.515 0.4641 0.67631 0.300282 0.034802 0.66995 85,510.338
MGWO-I 125.6959 21.42125 10.99907 0.515 0.515 0.4289 0.64169 0.30 0 022 0.03578 0.63704 85,513.323
GWO 125.6999 21.42129 10.99781 0.515 0.515 0.4687 0.68807 0.300151 0.04254 0.61701 85,507.111
SCA 125 21.14834 10.96928 0.515 0.515 0.5 0.7 0.3 0.02778 0.62912 83,431.117
MVO 125.6402 21.42250 10.99338 0.515 0.5150 0 0 0.4998 0.68782 0.301348 0.04617 0.60061 85,491.266
PSO 125 20.75388 11.17342 0.515 0.5150 0 0 0.5 0.61503 0.30 0 0 0 0 0.05161 0.60 0 0 0 81,691.202
GSA 125 20.85417 11.14989 0.515 0.517746 0.5 0.61827 0.304068 0.020 0 0 0.624638 82,276.941
Table 24
Statistical results for different algorithms for Rolling Element Bearing design prob-
lem.
Table 25
Comparison of best solution obtained from different algorithms for Multiple Disk Clutch Brake
Design.
Ri Ro T F Z
Table 26
Statistical results for different algorithms for Multiple Disk Clutch Brake design.
Sub ject to g1 (y ) = τ (y ) − τmax ≤ 0,
Algorithm Best Mean Worst Std. Dev. Median g2 (y ) = σ (y ) − σmax ≤ 0,
MGWO-III 0.259771 0.259840 0.259915 0.0 0 0 031 0.259835 g3 ( y ) = y1 − y4 ≤ 0 ,
MGWO-II 0.259775 0.259837 0.259913 0.0 0 0 031 0.259833 g4 (y ) = δ (y ) − δmax ≤ 0
MGWO-I 0.259778 0.259817 0.259879 0.0 0 0 031 0.259778 g5 (y ) = P − Pc (y ) ≤ 0
GWO 0.259787 0.259838 0.259891 0.0 0 0 033 0.259835
SCA 0.262286 0.267783 0.272872 0.002719 0.268412
g6 (y ) = 0.125 − y1 ≤ 0
MVO 0.259790 0.260426 0.267507 0.001712 0.259887 g7 (y ) = 1.10471y21 + 0.04811y3 y4 (14.0 + y2 )
PSO 0.272782 0.324099 0.371779 0.029831 0.327459 −5.0 ≤ 0
GSA 0.320926 0.343254 0.393482 0.020212 0.337767 Range 0.1 ≤ y1 ≤ 2, 0.1 ≤ y2 ≤ 10, 0.1 ≤ y3 ≤ 10,
0.1 ≤ y4 ≤ 2
where
best results for high-dimensional and noisy benchmark problems. y
6P L 6P L3
σ (y ) = ; δ (y ) =
√
y4 y23 Ey33 y4
Minimize F ( y ) = 2 2y 1 + y 2
P = 60 0 0 lb, L = 15 in, G = 12 × 106 psi, E = 30 × 106 psi
δmax = 0.25 in, τmax = 13600 psi, σmax = 30 0 0 0 psi
√
2y 1 + y 2
Sub ject to g1 (y ) = √ P − σ ≤ 0, A.3. Pressure vessel design problem
2y21 + 2y1 y2
y2
g2 ( y ) = √ P − σ ≤ 0,
2y21 + 2y1 y2
Minimize F (y ) = 0.6224y1 y2 y3 y4 + 1.7781y2 y23
1
g3 ( y ) = √ P − σ ≤ 0, +3.1661y21 y4 + 19.84y31 y3
2y 2 + y 1
Range 0 ≤ y1 ≤ 1, 0 ≤ y2 ≤ 1 Sub ject to g1 (y ) = −y1 + 0.0193y3 ≤ 0,
L = 100 cm, P = 2 kN/cm , 2
σ = 2kN/cm 2 g2 (y ) = −y2 + 0.00954y3 ≤ 0,
g3 (y ) = −π y23 y4 − 43 π y33 + 12960 0 0 ≤ 0,
g4 (y ) = y4 − 240 ≤ 0
where
y32 y3 ⎡ ⎤−0.3
Sub ject to g1 ( y ) = 1 − ≤ 0, 1.72 0.41 10/3
7178y41 1−γ f i (2 f 0 − 1 )
fc = 37.91⎣1 + 1.04 ⎦
4y22 − y1 y2 1 1+γ f 0 (2 f i − 1 )
g2 ( y ) =
+ − 1 ≤ 0,
12566 y2 y31 − y41 5108y21 0.41
γ 0.3 (1 − γ )1.39 2 fi
140.45y1 ×
g3 ( y ) = 1 − ≤ 0, (1 + γ )1/3 2 fi − 1
y22 y3
y2 + y1 φ0 = 2π − 2cos−1
g4 ( y ) = −1≤0
1.5
{(D − d )/2 − 3(T /4)}2 + {D/2 − T /4 − Db}2 − {d/2 + T /4}2
2{(D − d )/2 − 3(T /4 )}{D/2 − T /4 − Db}
Db ri r0
A.5. Speed reducer design problem γ= , fi = , f0 = , T = D − d − 2 Db
Dm Db Db
D = 160, d = 90, Bw = 30, ri = r0 = 11.033
Range 0.4 ≤ KDmin ≤ 0.5, 0.6 ≤ KDmax ≤ 0.7, 0.3 ≤ e ≤ 0.4,
Minimize F (Y ) = 0.785y1 y22 3.3333y23 + 14.9334y3 − 43.0934
0.02 ≤ e ≤ 0.1, 0.6 ≤ ζ ≤ 0.85
− 1.508y1 y26 + y27 + · · · 7.4777 y36 + y37 0.5 ( D + d ) ≤ Dm ≤ 0.6 ( D + d ),
+ 0.78054 y4 y26 + y5 y27 0.15(D − d ) ≤ Db ≤ 0.45(D − d ),
4 ≤ Z ≤ 50, 0.515 ≤ fi , f0 ≤ 0.6
[6] Eskandar H, Sadollah A, Bahreininejad A, Hamdi M. Water cycle algorithm: a [19] Mittal N, Singh U, Sohi BS. Modified grey wolf optimizer for global engineering
novel metaheuristic method for solving constrained engineering optimization optimization. Appl Comput Intell Soft Comput 2016:1–16.
problems. Comput Struct 2012;110:151–66. [20] Rashedi E, Nezamabadi-pour H, Saryazdi S. GSA: a gravitational search algo-
[7] Goldberg D. Genetic algorithms in Search, optimization and machine learning. rithm. Inf Sci 2009;179(13):2232–48.
MA: Addison-Wesley Professional; 1989. [21] Ray T, Liew KM. Society and civilization: an optimization algorithm based on
[8] Jitkongchuen, D., 2015, A Hybrid differential evolution with grey wolf opti- the simulation of social behavior. IEEE Trans Evol Comput 2003;7:386–96.
mizer for continuous global optimization, International Conference on Infor- [22] Sadollah A, Bahreininejad A, Eskandar H, Hamdi M. Mine blast algorithm: a
mation Technology and Electrical Engineering, pp. 51–54, Chiang Mai. new population based algorithm for solving constrained engineering optimiza-
[9] Kamboj VK. A novel hybrid PSO-GWO approach for unit commitment problem. tion problems. Appl Soft Comput 2013;13:2592–612.
Neural Comput Appl 2015:1–13 2015. [23] Saremi S, Mirjalili SZ, Mirjalili SM. Evolutionary population dynamics and grey
[10] Kannan BK, Kramer SN. An augmented lagrange multiplier based method for wolf optimizer. Neural Comput Appl 2015;26:1257–63.
mixed integer discrete continuous optimization and its applications to me- [24] Sarafrazi S, Nezamaadi-pour H, Seydnejad SR. A novel hyrid algorithm of
chanical design. J Mech Des. 1994;116:405–11. GSA with kepler algorithm for numerical optimization. Comput Inf Sci
[11] Kennedy J, Eberhart R. Particle swarm optimization. In: IEEE International Con- 2015;27:288–96.
ference on Neural Networks, 4; 1995. p. 1942–8. [25] Wen L, Shao-Hong C, Jian-Jun J, Wen-Zhuan Z, Ming-Zhu T. Hybrid grey
[12] Kishor A, Singh PK. Empirical study of grey wolf optimizer. In: International wolf optimization algorithm for high dimensional optimization. Control Decis
Conference on Soft Computing for Problem Solving; 2016. p. 1037–49. 2016;31(11):1991–7.
[13] Korayem L, Khorsid M, Kassem SS. Using grey wolf algorithm to solve the ca- [26] Wen L. Grey wolf optimizer based on nonlinear adjustment control parameter.
pacitated vehicle routing problem. Mater Sci Eng 2015;83:1–10. In: International Conference on Sensors, Mechatronics and Automation; 2016.
[14] Kumar V, Chhabra JK, Kumar D. Variance-based harmony search algorithm for p. 643–8.
unimodal and multimodal optimization problems with application to cluster- [27] Yao X, Liu Y, Lin G. Evolutionary programming made faster. IEEE Trans Evol
ing. Cybern Syst 2014;45:486–511. Comput 1999;3:82–102.
[15] Mirjalili S, Mirjalili SM, Lewis A. Grey wolf optimizer. Adv Eng Softw [28] Zhang S, Zhou YQ. Grey wolf optimizer based on powell local optimization
2014;69:46–61. method for clustering analysis. Discrete Dyn Nature Soc 2015:1–17.
[16] Mirjalili S. Moth-flame optimization algorithm: a novel nature-inspired heuris- [29] Zhu A, Xu C, Li Z, Wu J, Liu Z. Hybridizing grey wolf optimization with dif-
tic paradigm. Knowledge-based Syst 2015;89:228–49. ferential evolution for global optimization and test scheduling for 3D stacked
[17] Mirjalili S, Mirjalili SM, Hatamlou A. Multi-Verse optimizer: a nature-inspired SoC. J Syst Eng Electron 2015;26:317–28.
algorithm for global optimization. Neural Comput Appl 2016;27:495–513.
[18] Mirjalili S. SCA: a sine cosine algorithm for solving optimization problems.
Knowledge-based Syst 2016;96:120–33.