Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

J. Inst. Eng. India Ser.

B (February 2022) 103(1):107–117


https://doi.org/10.1007/s40031-021-00645-y

ORIGINAL CONTRIBUTION

Improved based Differential Evolution Algorithm using New


Environment Adaption Operator
Shailendra Pratap Singh1

Received: 20 November 2020 / Accepted: 10 June 2021 / Published online: 23 July 2021
 The Institution of Engineers (India) 2021

Abstract In this work, a novel operator-based differential Introduction


evolution (DE) algorithm has been proposed. The proposed
approach has been inspired by the internal adaption (en- Nature-inspired optimization is the field of such study for
vironment) of the search space. Therefore, maintaining planning, simulation, and execution of problems using
environment (vectors) for the search space can be achieved scientific methodologies which are inspired by nature.
by introducing the better fitness of candidate solution. In Meanwhile, practical applications are multidimensional
the proposed approach, candidate solutions are multiplied and nonlinear, and they require sophisticated optimization
with the different parameter values, which depend on the implementations to handle such a curse of dimensionality.
nature of the problem and available counterbalancing Nowadays, computer simulations are the vibrant imple-
resources. The proposed variant termed an internal adap- mentation for resolving such optimization complications
tion-based environment is considered in the existing through nature-inspired or meta-heuristic search
mutation and crossover operators to provide more diversity algorithms.
for selecting the effective mutant solutions. In the experi- Meta-heuristic is a class of algorithms that are used for
mental analysis, the proposed approach is compared with the optimization of candidate solutions to discover, pro-
the five modern DE variants and tested on benchmark duce, select an optimized candidate solution that may
function (f1 to f24) on 20, and 40 dimensions. In addition, deliver a suitably better solution to an optimization prob-
it is also verified by hypothesis testing in terms of the lem. Meta-heuristics may feasibly make reasonable, con-
minimum error. From the obtained results, it is observed siderable, justifiable assumptions about the optimization
that the proposed algorithm is found to be a better target problem being solved by the concept of nature. Many
value in terms of the minimum number of function eval- meta-heuristics algorithms are developed such as ant col-
uation and statistical functions. It also validates that the ony optimization (ACO), particle swarm optimization
proposed algorithm has achieved a reasonable convergence (PSO), artificial bee colony (ABC), and genetic algorithm
rate and diversity on 20 and 40 dimensions. GA). The GA is a standard in optimization due to its nature
of finding global optima. The differential evolution (DE) is
Keywords Adaptation  Optimization  an optimization algorithm like GA and is efficient in
Evolutionary algorithm finding the global minima of non-differentiable problems.
In this article, a novel mutation-based modified differential
evolution algorithm has been proposed. The original DE
[1] and [23] was anticipated by Rainer Storn and Kenneth
Price in 1995-96. The philosophy of DE was inspired by
& Shailendra Pratap Singh natural evolution.
shail2007singh@gmail.com DE is basically a vector-centered meta-heuristic
1 methodology, which is similar to genetic algorithms
Department of Computer Science and Engineering,
Bundelkhand Institute of Engineering and Technology, because both the methodologies use the same kinds of
Jhansi, Uttar Pradesh, India operations like crossover and mutation. It is a stochastic

123
108 J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117

search tendency to guide itself during the search process to


derive the optimized solution and it is worth mentioning
that it does not use the domain derivatives. In this, we can
say that it is a population-based and derivative-free
methodology. DE algorithm explain here:

Basic concepts of DE

The basic differential evolution algorithm [1, 2] is descri-


bed in algorithm 1. DE process goes through four steps,
i.e., initialization, mutation, crossover, and selection.

Initialization

In this process, all the population members are randomly


initialized between the lower and upper bound values. The
representation of target vector is shown in Eq. 1.
ai;G ¼ fa
~ ~1 ;i ;G ; ~ ~D ;i ;G g
a2 ;i ;G ; :::::::a ð1Þ
where NP is a population size of parameter vectors, i ¼ 1,
2, 3........NP, D is a dimension and G is a generation.

Mutation

DE generates new offspring by adding the weighted dif-


ference of two candidate solution trajectories to a third
trajectory. This process is termed as the mutation. This
operator is explained in the algorithm 1.
In this paper, modification of standard DE algorithm is
Crossover suggested named as improved based differential evolution
algorithm using environment adaption operator (IBDE). In
After mutation, the crossover binomial operator is applied. the IBDE algorithm, a novel new mutation operator is
The binomial crossover is explained in the algorithm 1. devised. The IBDE algorithm makes use of a new mutation
strategy. This variant derives itself from the process of
Selection environment factor which entirely invests in sustaining the
overall environment of the particular area against any
In this phase, the trial vector formed afterward mutation change in the external environment. This variant also aims
and crossover are estimated. Then, the effectiveness of the to give impetus to the convergence rate of the algorithm
trial and target vector is equated and the fittest vector or and thereby retains the same diversity generated by the
candidate is selected. If the trial vector produces an environment factor during the initial generation in the
insignificant fitness value, it may be optimized for the next mutation operator. Diversity is taken as the internal and
generations. external environment by proposed algorithms. The pro-
posed algorithm is tested for benchmark function (f1 to
f24) better than other variants of DE algorithms.

Related Works

The existing evolutionary approaches and advantage of


proposed approaches over existing are presented.
The experimental assessment of some DE variants to
solve optimization problems have been described [3]. The
main goal is to recognize which one of them is better to

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117 109

solve optimization, subject to features of the problem offspring by evolving 3 trial vector generations with the
domain. Eight different categories of DE variants are use of three different control parameter settings. The
executed on 13 benchmark problems. A set of statistical experimental outcome shows that CoDE provides the better
examinations are also accomplished to get assurance on the result.
soundness of the outcomes. As per the researches [16], new mutation strategy has
A new local search trigonometric mutation has been been proposed for the parent selection method of the
proposed for DE [4]. This alteration allows the algorithm to crossover in the DE. DE variation is achieved by incor-
achieve an improved trade-off between convergence and porating JDE and JADE variants, and their performance is
robustness. In the literature [5], author proposed a improved. The effectiveness of the algorithm [17] is
methodology for DE with the self-adaptive feature. This evaluated by different sets of optimization test problems.
method used a self-adaptive associated control parameter The outcomes determine that the projected process is better
for the algorithm and generated the trial vector. in terms of computational time over the different standard
The efficacy of some current population-direct search algorithms. In the literature [18], crossover rate repair
algorithm approaches such as Random Search, DE, and GA technique for the adaptive DE has been proposed. This
has been depicted [6]. Also, some improvement is made to crossover rate in DE is refurbished by using the average
DE and to GA to improve their effectiveness and are tested number of mutants. In earlier researches [19], a role
for different test problems. assignment (RA) method has been proposed. The RA
In the literature [7], a new version of the DE algorithm scheme uses fitness and positional information to dynam-
has been proposed for finding self-adaptive control ically divide as a role having its own mutation. In the
parameter settings. The outcomes demonstrate that the literature [20], a new DE variant has been proposed with an
algorithm with self-adaptive control parameter settings is attractive self-adaptive strategy and hybrid mutation
better as compared to the standard DE [8]. methodology. The proposed method [21] is equated with
As per the researches [9], authors have examined the different DE variants. The statistical outcomes show that
consequences of varying population size on the quality of the average performance of the proposed method is better
generated candidate solutions and evaluated on 5 problem as compared to other standard algorithms.
sets opted from CEC. Outcomes include the effects of the As per the researches [23], suggestion is given for a new
different populations with varying different dimensions, heuristic approach to diminish nonlinear and non-differ-
and it has shown the considerable effect of population size entiable continuous space functions and also for signifying
on the efficiency of DE. that the new strategy merges faster with impetus than
In earlier investigations [10], two new, improved vari- existing computational techniques. In the literature [24], an
ants of DE have been proposed. The effectiveness of the enhancement of optimization performance is suggested by
proposed methods is compared against original DE, the way of a new DE algorithm which adapts a new
canonical PSO, and two PSO-variants. In the literature mutation method ‘‘DE/current-to-pbest.’’
[11], the article’s authors investigated that how DE can The literature of the DE [29] and a survey of its many
work in parallel according to the ring topology and variants of DE are explained. In the researches [30], an
improved convergence rate with added performance to the adaptive version DE algorithm has been proposed and
technique. evaluated on the 24 mathematical functions taken Black-
An adaptive-based crossover local search (LS) operation Box Optimization benchmark function. A suggestion for a
for the DE algorithm has been proposed by the researchers novel hybrid DE variant is proposed in the literature [31]
[12]. LS is very beneficial in terms of planning an evolu- that contains two sub-variants: HMLSL and MLSL. This
tionary algorithm for the optimization problem. In this algorithm has a high success rate and also is much faster
paper, authors highlighted how the scheme can increase the when it comes to hybrid algorithms.
effectiveness of DE. The modified version is tested with the In the literature [38], authors empirically examine the
different benchmark functions. Performance evaluations selection strategy of solutions for the mutation (donor
with LS and with some well-known algorithms were also vector) in DE. This algorithm is performed on wide
presented. In the literature [13], authors proposed five new experimentation, and the authors suggest that some state-
mutation methods for the DE algorithm and are termed of-the-art algorithms provided the better optimum value.
MDE1, MDE2, MDE3, MDE4, and MDE5. In the earlier researches [39], a cumulative population
Various researchers [14] have studied different distribution-based DE method has been proposed. The
replacement schemes in the DE through an experiment on proposed method has been applied to different variants of
different standard test functions. In the literature [15], a DE for solving two sets of benchmark functions. In the
novel method composite DE (CoDE) has been proposed. literature [40], covariance matrix learning has been pro-
The methodology added an adaptive way to produce the posed and a bimodal distribution parameter setting

123
110 J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117

approach is proposed with the aim of balancing the Environment Adaption-based Mutation Operator
exploitation or exploration abilities of DE. The effective-
ness of the method is checked on twenty-five benchmark In this section, the proposed approach focuses on the
functions. The investigational outcomes establish the effi- maintaining environment (vectors) for the search space that
ciency of the approach. can be achieved by introducing the better fitness of the
candidate solution. The objective of the proposed approach
Advantages of Proposed Over Existing Approaches is achieved by environment vectors (E1 and E2) which
maintain a better environment of search space. The envi-
List of advantages over the exiting approaches are shown ronment of DE can be preserved based on the environment
as follows [20, 39–44]: through these two equations Eq(4) and Eq(5).
• Proposed algorithm enhances the convergence rate of E1 ¼ðvectorÞ  feasibleenvironment ð4Þ
global search space, in contrast to the existing
E2 ¼ðvectorÞ  infeasibleenvironment  EF ð5Þ
approaches [37] where convergence rate is low on
tested benchmark functions. where E1 and E2 are feasible and infeasible environment
• This algorithm provides the balance in the environment vectors of global search space. The vector denotes the
from local to the global area for different applications, current population of search space, and EF denotes the
whereas, existing approaches not able to balance on environment factor. This factor defined according to the
some of the tested benchmark functions [40]. internal environment vector, which is multiplied by EF
• Existing methods [29–40] facing the problem of value (0.1 to 0.01) which is set between the interval 0.1 to
stagnation after a certain number of iterations. For 0.01. This value provides sufficient diversity to the current
resolving the issue, a new environment adaption vector and also improves the performance of convergence
mutation operator for the DE algorithm is devised. speed. Therefore, when it is implicated in the local optimal
• Proposed algorithm also reduces the time complexity problem, sufficient required variation is obtained by the EF
and space complexity, therefore, it outperforms the value. The detailed description of the generation of new
other tested optimization algorithms on standard bench- donor vector methodology is as follows.
mark functions [33–35].
Most of the above literature survey are facing the problem Generation of New Donor Vector(DV)
of diversity, tuning parameter, premature convergence, and
stagnation after a certain number of iterations. Also, the In this process, environmental optimization is used on the
decrease convergence speed of these approaches is quite current candidate solution in the search space, to generate
slow in terms of the best optimum value. For resolving all an environment optimization-based mutation scheme of
these issues, we have introduced a new environment DE / BEST / 1. We define the common mutation operator
adaption operator for DE Algorithm, which is discussed in in equation (6) and feasible environments (E1 and E2) in
the next section. equation (7) and (8), respectively.
~ i;G ¼ ~
DV ar1i ;G þ d1  ða
~r2i ;G  ~
ar3i ;G Þ ð6Þ

Proposed Approach: Improved based DE Update the new mutation operator:


Algorithm using new Environment Adaption ~ i;G ¼a
DV ~ ri ;G  E1
~ri ;G þ d1  ðE1 ~ ri ;G Þ ð7Þ
Operator (IBDE) 1 2 3

~ i;G ¼a
DV ~ ri ;G  E2
~r1i ;G þ d1  ðE2 ~ ri ;G Þ ð8Þ
2 3
For the driving of the DE algorithm, the role of mutation
and crossover is quite important. Therefore, this paper uses where DV ~ i;G denote the mutant vector or donor vector,
a mutation operator that dynamically updates the progress i ¼ 1; 2; ::::p is population(P), and G is generation. ~ ar1i ;G
of the algorithm. Convergence and diversity are an denotes the best random vector of current population, and
important factor in determining a candidate solution from a d1 denote the scale factor of mutation operator. ~ ar2i ;G and
global search space. We cannot think of a better solution ~
ar3i ;G will be generated from the entire search space. E1 ~ ri ;G ,
2
unless the entire search location is initially searched cor- ~ ri ;G , E2
E1 ~ ri ;G , and E2
~ ri ;G are internal adaption environment
3 2 3
rectly or at the later stages of evaluation. In DE, the key vectors of global search. These mutation strategies (equa-
feature of diversification maintenance goes to the mutant tion 7 & 8) perform better to enhance the bandwidth of
operator, and in DE, diversity maintenance is performed by diversity from search space when stagnation condition
a mutation operator as a new environment adaption. This arises with the adaption technique introduced through these
process is explained in the next section. environments (E1 and E2).

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117 111

Crossover Description of Test-Bed and Performance Indicators

In this process, a conventional crossover operator is The Comparing Continuous Optimizers (COCO) is a
adopted to generate the trail vector. In this operator, firstly, mathematically model platform. It performs searching in a
the random number (rand) is generated between 0 and 1. If large space and compares the results with respect to global
rand is less than CR, then crossover is performed between parameters. The COCO benchmark function test-bed pro-
target vector and donor vector, otherwise, we will continue vides 24 noise-less test functions [22, 25–28]. The obtained
with the target vector. This operator mainly focuses on the results of all these functions are verified in the COCO
tuning of parameters according to the internal and external framework based on the below-mentioned criteria:
environment. In this paper, the crossover rate is considered
1. Wilcoxon Rank-Sum Test
between 0.2 to 0.7.
2. ERT-Expected is running time Loss Ratio.
3. ECDF-Empirical cumulative distribution functions.
Selection
In the comparative analysis, the obtained results of the
After completion of the crossover operation, the selection proposed method (IBDE Algorithm) on 24 benchmark
operator is applied to choose the solution based on the sur- functions(f1-f24) are compared with the existing DE
vival of the fittest strategy. In this process, three vectors are variants such as CPI-DE [33], TSDE [35], ToPDE [34],
taken into consideration, namely target vector, donor vector, JADEcr [18] and MPEDE [32] for all the dimensions 20-D
and trail vector. Afterward, the best fitness valued vector & 40-D. The control parameters considered by the
among them is considered for the further evolution process. proposed algorithm are mentioned in Table 1.

IBDE Algorithm Simulation Framework

The pseudocode of proposed IBDE algorithm is shown in The description of simulation framework considered to
algorithm 2. execute the proposed and existing algorithms is as follows:
• The experiments were executed with Intel Core i7-
8850H Processor, 8 GB RAM and operating system
Windows 10 Pro 64 bit.
• The IBDE algorithm is tested on twenty-four bench-
mark functions (f1-f24) using the COCO platform.
These functions tested in given search space, i.e., [5, -
5]D , where D denotes the dimension.
• The proposed algorithm is executed 20 times, and
average results are taken into consideration for the
comparative analysis.
• Initialize population of all the test functions is 100. The
exploration and exploitation steps are performed over it
either for the maximization and minimization problem.
• The minimization problem has been considered accord-
ing to the benchmark functions (f1-f24), as given in
Table 2.

Table 1 Control parameters of algorithm 2


Sr. No. Parameter Type
Experimental Framework and Analysis
1 Population size(NP) 100
2 Mutation(scale factor) d[0-2]
In this section, a description of the test-bed, performance
3 Crossover rate Cr[0-1]
indicators, and results obtained with respect to the perfor-
4 Dimension [20, & 40]
mance indicators are mentioned.
5 FEs Function evaluations
6 Environment factor (EF) Rand[0.1–0.01]

123
112 J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117

Table 2 Benchmark Function


Sr. No. Benchmark functions Features Group-based functions

1 f1 f1 is related to the sphere function Group (f1-f5) to separable


2 f2 f2 is related to the Ellipsoidal function
3 f3 f3 is related to the Rastrigin function
4 f4 f4 is related to the Buche-Rastrigin function
5 f5 f5 is related to the linear slope function
6 f6 f6 is related to the attractive sector function Group (f6-f9) to low or moderate conditioning
7 f7 f7 is related to the step ellipsoidal function
8 f8 f8 is related to the Rosenbrock, original function
9 f9 f9 is related to the Rosenbrock, rotated function
10 f10 f10 is related to the Ellipsoidal function Group (f10-f14) to unimodal with high conditioning
11 f11 f11 is related to the Discus function
12 f12 f12 is related to the Bent Cigar function
13 f13 f13 is related to the sharp Ridge function
14 f14 f14 is related to the different powers function
15 f15 f15 is related to the Rastrigin function Group (f15-f19) to multi-modal with adequate structure
16 f16 f16 is related to the Weierstrass function
17 f17 f17 is related to the Schaffers f7 function
18 f18 f18 is related to the Schaffers f7, moderately function
19 f19 f19 is related to the composite (f2 and f8) function
20 f20 f20 is related to the Schwefel function Group (f20-f24) to multi-modal global structure
21 f21 f21 is related to the Gallagher’s function
22 f22 f22 is related to the Gallagher’s Gaussian 21 function
23 f23 f23 is related to the Katsuura function
24 f24 f24 is related to the Lunacek bi-Rastrigin function

Performance Measurement in terms of Wilcoxon Performance Measurement in terms of ERT Test


Rank-Sum Test on 20 Dimension

The proposed method is verified by the Wilcoxon rank-sum In the section, obtained performance with respect to the
test on 20 dimensions and presented in Figs. 2, and 3. ERT test on 20 dimensions is discussed and graphically
These tests have used the comparison of two or more presented in Fig. 1.
groups of nonparametric on the benchmark functions(f1- Illustration of Fig. 1: It shows the ERT (Expected run-
f24). It is observed that the proposed method finds the best ning time) loss ratio for 20 dimension. It displays plotted
target f opt þ 108 in terms of the minimum number of log10 of the given function’s ERT loss ratio
function evaluation and statistical functions. (FEvals ¼ FEs). This function demonstrates the function
Wilcoxon rank-sum test is important for checking the evaluation of each group taken from the benchmark func-
higher ranking (higher to lower) according to the distri- tion set based on the ERT loss ratio like f1-f5, f6-f9, f10-
butions of observations. This observation takes the sum of f14, f15-f19, f20-f24, and f1-f24. Each group of functions
the ranks of individuals groups like f1-f5, f6-f9, f10-f14, uses the Box-Whisker plot to demonstrate the error loss of
f15-f19, f20-f24, and f1-f24. These functions are compar- diversity. The minimum point is shown in the box means
ing the two separate groups on nonparametric tests the error ratio is low in the proposed algorithm. Hence, the
according to statistical probability (p ¼ 0:05). It is proposed algorithm shows significant performance as given
observed that the proposed technique achieves the best in Fig. 1.
target(108 ) with the minimum number of iterations as
compare to existing DE algorithms. The obtained results of
ERT and ECDF are mentioned in the next sections.

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117 113

(a) (b) (c)

(d) (e) (f)

Fig. 1 ERT of IBDE algorithms in 20-D

Performance Measurement in terms of ECDF Test Comparative Analysis with other existing DE
on 20 Dimension Variants

This section describes the obtained result with respect to In this subsection, the performance of the proposed opti-
the ECDF test on 20 dimensions and graphically presented mization technique is analyzed with the existing DE vari-
in Fig. 2. ants. These techniques are tested on 24 benchmark function
Illustration of Fig. 2: In the figure, the x-axis denotes the and applied on different dimension, i.e., 20-D, & 40-D for
log10 with funEvl/DIM, and the y-axis represents the the rigorous testing. The obtained results concerning the
proportional trails from 20 Dimension. It describes the function evaluations (FEs) and Best Search (target function
number of times the proposed algorithm runs to tries to 108 ) are presented in Table 3. In addition, the proposed
achieve the best target with respect to the function evolu- methodology and the existing variants of DE are also tested
tion. Functions are taken in groups similar to ECDF test on statistics functions which are depicted in Figs. 2, 3, and
like f1-f5, f6-f9, f10-f14, f15-f19, f20-f24, and f1-f24. It is 4.
observed that the proposed algorithm achieves the target
value with minimum function evaluations in the ECDF test. Comparative Analysis with Respect to 20 Dimension
In figure, ECDF test represented by the red line on 20
dimension for achieving the difficult target function value Obtained results on Statistics Functions: The proposed
f opt þ 108 [22]. The trails (lines) represent the four values algorithm achieved the optimal value of the statistics
that achieved the target according to hard trails, i.e., 1, -1, function in the minimum number of iterations as compared
-4, -8, or represent the color. The proposed algorithm to the other existing DE variants as shown in Fig. 3. From
achieved the target value of each benchmark function (f1- Fig. 3, it was observed that the IBDE algorithm produces
f24). These functions achieved to trials values, i.e., 1 rep- better target values as compared to the CPI-DE, TSDE,
resent obtained 23 optimum value, -1 represent obtained MPEDE, JADEcr, and ToPDE for all the standard func-
18 optimum value, -4 represent obtained 16 optimum tions (f1-f24). Hence, the proposed technique achieved a
value, and -8 represent obtained 16 optimum value is reasonable convergence rate and diversity on 20
shown in Fig. 2(f). dimension.

123
114 J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117

(a) (b) (c)

(d) (e) (f)

Fig. 2 ECDF of IBDE algorithm on 20-D

Table 3 Proposed (IBDE) algorithm comparing with other standard DE Algorithms


Sr. No. Benchmark function Performance algorithms on 20-D Performance algorithms on 40-D

1 Group (f1–f5) to separable IBDE IBDE


2 Group (f6–f9) to low or moderate conditioning IBDE IBDE
3 Group (f10–f14) to unimodal with high conditioning MPEDE CPI-DE
4 Group (f15–f19) to multi-modal with adequate structure CPI-DE IBDE
5 Group (f20–f24) to multi-modal global structure IBDE IBDE
6 Group (f1–f24) to all benchmark function IBDE IBDE

Comparative Analysis with Respect to 40 Dimension even when faced with the local optimum problem. The
requisite threshold diversity is gained by applying the value
Obtained results on Statistics Functions: The proposed of EF from (0.1 to 0.01) and (0.1 to 1). EF is ample for
algorithm achieved the optimal value of the statistics encapsulating targeted solutions in all functions owing to
function in the fewer number of iterations as given in the design of the adaption operator. It has been considered
Fig. 4. This figure clearly demonstrates that the IBDE the fact that large size may lead to exploitation and
algorithm produces better target values as compared to the exploration around many areas. It can be seen an upsurge
TSDE, MPEDE, JADEcr, ToPDE, and CPI-DE for all the in diversity and convergence rate thereby improving the
standard functions (f1-f24). Hence, a reasonable conver- overall result as shown in Table 3, and Figs. 3, and 4.
gence rate and diversity were achieved on 40 dimensions. Effect of various parameters on the performance of
algorithm: NP (Population size) is an important parameter
Effect of Environment Adaption-Based Operator for any DE algorithm. The size of the population is
on the Performance of Algorithm dependent on the introduced dimensions that are smaller to
higher, and these dimensions are used in capturing the
The crucial operator for this algorithm is the Environment target value for a given function. This is due to the design
adaption-based operator. This operator retains the diversity of the adaptation operator. If we take large size we may do

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117 115

(a) (b) (c)

(d) (e) (f)

Fig. 3 IBDE algorithm comparing with standard algorithms on 20-D

(a) (b) (c)

(d) (e) (f)

Fig. 4 IBDE algorithm comparing with standard algorithms on 40-D

123
116 J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117

exploration and exploitation around many local areas. The optimization, in Proceedings of the 8th Annual Conference on
number of generations is important for the DE algorithm. Genetic and Evolutionary Computation (2006), pp. 485–492
4. H.-Y. Fan, J. Lampinen, A trigonometric mutation operation to
For this proposal, 150 generations may work efficiently differential evolution. J. Global Opt. 27(1), 105–129 (2003)
with any population. 5. A.K. Qin, V.L. Huang, P.N. Suganthan, Differential evolution
algorithm with strategy adaptation for global numerical opti-
Complexity Analysis for the Proposed Algorithm mization. IEEE Trans. Evol. Comput. 13(2), 398–417 (2009)
6. M.M. Ali, A. Torn, Population set based global optimization
algorithms: some modifications and numerical studies. Comput.
In the literature analysis, I have observed an increase in the Oper. Res. 31(10), 1703–1725 (2004)
time and space complexity, when stuck in local optima. 7. J. Brest, S. Greiner, B. Boskovic, M. Mernik, V. Zumer, Self
Therefore, the proposed algorithm provides the proper adapting control parameters in differential evolution: a compar-
ative study on numerical benchmark problems. IEEE Trans. Evol.
balance with the consideration of both internal and external Comput. 10, 646–657 (2006)
environmental factors, whereas existing approaches 8. V.L. Huang, S.Z. Zhao, R. Mallipeddi, P.N. Suganthan, Multi-
directly involved the solution vector for the next population objective optimization using self-adaptive differential evolution
generation that loses the diversity. Therefore, the proposed algorithm, in Proceedings of the 11th Conference on Evolution-
ary Computation,Trondheim, Norway, pp. 190-194, 18–21 May
algorithm’s initial generation has not stagnated most of the 2009
benchmark functions (f1-f24). Hence, the proposed algo- 9. R. Mallipeddi, P.N. Suganthan, Empirical study on the effect of
rithm is achieved a reasonable convergence rate and population size on differential evolution algorithm, in Proceed-
diversity on 20 and 40 dimensions. ings of the IEEE Congress on Evolutionary Computation, Hong
Kong, pp. 3663-3670, 1–6 June 2008
10. S. Das, A. Konar, U.K. Chakraborty, wo improved differential
evolution schemes for faster global search, in Proceedings of the
Conclusion Genetic Evolution Computing Conference, Washington DC, USA
(2005), pp. 991-998
11. D.K. Tasoulis, N.G. Pavlidis, V.P. Plagianakos, M.N. Vrahatis,
In this paper, the environment adaption-based operator Parallel differential evolution, in Proceedings of the 2004 Con-
maintains the diversity, and also improved the convergence gress on Evolutionary Computation (2004), pp. 2023–2029
speed on high dimensions in the global search. The DE 12. N. Noman, H. Iba, Accelerating differential evolution using an
algorithm begins with a high generation of function eval- adaptive local search. IEEE Trans. Evol. Comput. 12(1), 107–125
(2008)
uation and applies an environment adaption optimization 13. Radha Thangaraj, Millie Pant, Ajith Abraham, New mutation
method. The method enhanced the diversity in local and schemes for differential evolution algorithm and their application
global (search) areas. This proposed method checked all to the optimization of directional over-current relay settings.
the twenty-four benchmark functions for real-parameter Appl. Math. Comput. 216(2), 532–544 (2010)
14. I. Fajfar, J. Puhan, S. Tomazic, A. Burmen, On Selection in
and global optimizer. The proposed IBDE version is Differential Evolution, English Edition. (Elektrotehniki Vestnik,
compared to five modern DE algorithms, and the results are 2011), pp. 275–280
verified concerning dimensions of 20 & 40, and as shown 15. Y. Wang, Z. Cai, Q. Zhang, Differential evolution with composite
in Table 3. In Figs. 1, 2, 3, and 4, the rank-sum test has trial vector generation strategies and control parameters. IEEE
Trans. Evol. Comput. 15, 55–66 (2011)
been checked and statistically on 20 & 40 dimensions for 16. S.M. Islam, S. Das, S. Ghosh, S. Roy, P.N. Suganthan, An
better performance hypothesis testing. This algorithm of adaptive differential evolution algorithm with novel mutation and
the future scope controls the tuning of control parameters crossover strategies for global numerical optimization, IEEE
by the internal and external dynamic selection of the entire Trans. Syst. MAN Cyber. Part B Cyber. 42(2), 482–500 (2012)
17. Ruhul A. Sarker, Saber M. Elsayed, Tapabrata Ray, Differential
population. evolution with dynamic parameters selection for optimization
problems. IEEE Trans. Evol. Comput. 18(5), 689–707 (2014)
18. W. Gong, Z. Cai, Y. Wang, Repairing the Crossover Rate in
Adaptive Differential Evolution Applied soft commuting, (Else-
References vier, Feb 2014), pp. 149–168
19. Xinyu Zhou, Wu. Zhijian, Hui Wang, Shahryar Rahnamayan,
Enhancing differential evolution with role assignment scheme.
1. R. Storn, K. Price, Differential Evolution–A Simple and Efficient Soft Comput. 18, 2209–2225 (2014)
Adaptive Scheme for Global Optimization Over Continuous 20. Wenchao Yi, Liang Gao, Xinyu Li, Yinzhi Zhou, A new differ-
Spaces (International Computer Science Institute, Berkeley, ential evolution algorithm with a hybrid mutation operator and
1995) self-adapting control parameters for global optimization prob-
2. M.G.H. Omran, A. Salman, A.P. Engelbrecht, Self-adaptive dif- lems. ApplIntell 42, 642–660 (2015)
ferential evolution, in Proceedings of the International Confer- 21. Qinqin Fan, Xuefeng Yan, Self-adaptive differential evolution
ence on Computational Intelligence and Security, Xi’an, China, algorithm with discrete mutation control parameters. Exp. Syst.
pp. 192-199, 15–19 Dec 2005 Appl. 42, 1551–1572 (2015)
3. E. Mezura-Montes, J. Vel azquez-Reyes, C.A. Coello, A com- 22. http://coco.gforge.inria.fr/
parative study of differential evolution variants for global

123
J. Inst. Eng. India Ser. B (February 2022) 103(1):107–117 117

23. Rainer Storn, Kenneth Price, Differential evolution—a simple 35. Z. Z. Liu, Y. W. S. Yang, Z. Cai, Differential evolution with a
and efficient heuristic for global optimization over continuous two-stage optimization mechanism for numerical optimization, in
spaces. J. Global Opt. 11(4), 341–359 (1997) 2016 IEEE Congress on Evolutionary Computation (CEC)
24. Jingqiao Zhang, Arthur C. Sanderson, JADE: adaptive differen- (Vancouver, BC, 2016), pp. 3170–3177
tial evolution with optional external archive. IEEE Trans. Evol. 36. Shailendra Pratap Singh, Anoj Kumar, Homeostasis mutation
Comput. 13(5), 945–958 (2009) based differential evolution algorithm. J. Intell. Fuzzy Syst.
25. N. Hansen et al., Real-parameter black-box optimization bench- 32(5), 3525–3537 (2017)
marking 2009: noiseless functions definitions (2009) 37. Hu. Zhongbo, Su. Qinghua, Xianshan Yang, Zenggang Xiong,
26. S. Finck, N. Hansen, R. Ros, A. Auger, Real-parameter black-box Not guaranteeing convergence of differential evolution on a class
optimization benchmarking 2009: presentation of the noiseless of multimodal functions. Appl. Soft Comput. 41, 479–487 (2016)
functions, in Technical Report 2009/20, Research Center PPE, 38. Y. Wang, Z. Z. Liu, J. Li, H. X. Li, J. Wang, On the selection of
2009. Updated February, (2010) solutions for mutation in differential evolution. Front. Comput.
27. S. Finck, N. Hansen, R. Ros, A. Auger, Real-parameter black-box Sci. 12, 297–315. (2016)
optimization benchmarking2010: presentation of the noiseless 39. Yong Wang, Zhi-Zhong. Liu, Jianbin Li, Han-Xiong. Li, Gery G.
functions. http://coco.lri.fr/downloads/download15.02/bbobdoc Yen, Utilizing cumulative population distribution information in
functions.pdf differential evolution. Appl. Soft Comput. 48, 329–346 (2016)
28. N. Hansen, A. Auger, S. Finck, R. Ros, Real-parameter black-box 40. Yong Wang, Han-Xiong. LI, Tingwen Huang, Long Li, Differ-
optimization benchmarking 2010: Experimental Setup, in Tech- ential evolution based on covariance matrix learning and bimodal
nical Report, RR-7215, INRIA (2010) distribution parameter setting. Appl. Soft Comput. 18, 232–247
29. S. Das, P.N. Suganthan, Differential evolution: a survey of the (2014)
state-of-the-art. IEEE Trans. Evol. Comput. 15(1), 4–31 (2011) 41. S.P. Singh, A. Kumar, Multiobjective differential evolution using
30. P. Posik, V. Klems, JADE, an adaptive differential evolution homeostasis based mutation for application in software cost
algorithm, benchmarked on the BBOB noiseless testbed. estimation. Appl. Intell. 48(3), 628–650 (2018)
GECCO’12, Philadelphia, PA, USA, (2012) 42. S.P. Singh, A. Kumar, Software cost estimation using home-
31. L. Pal, Benchmarking a Hybrid Multi Level Single Linkage ostasis mutation based differential evolution, in 2017 11th
Algorithm on the BBOB Noiseless Testbed, GECCO’13, Ams- International Conference on Intelligent Systems and Control
terdam, Netherlands, (2013) (ISCO), (2017), pp. 173–181
32. G.H. Wu, R. Mallipeddi, P.N. Suganthan, R. Wang, H.K. Chen, 43. S.P. Singh, V.P. Singh, A.K. Mehta, Differential evolution using
Differential evolution with multi-population based ensemble of homeostasis adaption based mutation operator and its application
mutation strategies. Inf. Sci. 329, 329–345 (2016). for software cost estimation, J. King Saud Univ. Comput. Inf.
https://doi.org/10.1016/j.ins.2015.09.009 Sci., (2018)
33. Y. Wang, Z.Z. Liu, J. Li, H.X. Li, G.G. Yen, Utilizing cumulative 44. S. P. Singh, Cost estimation model using enhance based differ-
population distribution information in differential evolution. ential evolution algorithm, Iran J. Comput. Sci. (Springer),
Appl. Soft Comput. 48, 329–346 (2016) (2019), pp. 1–12
34. Y. Wang, B. Xu, G. Sun, S. Yang, A two-phase differential
evolution for uniform designs in constrained experimental Publisher’s Note Springer Nature remains neutral with regard to
domains. IEEE Trans. Evol. Comput., (in press). jurisdictional claims in published maps and institutional affiliations.
https://doi.org/10.1109/TEVC.2017.2669098

123

You might also like