Accepted Manuscript

Algorithmic design issues in adaptive differential evolution schemes: Review and

taxonomy

Rawaa Dawoud Al-Dabbagh, Ferrante Neri, Norisma Idris, Mohd Sapiyan Baba

PII: S2210-6502(17)30583-7
DOI: 10.1016/j.swevo.2018.03.008
Reference: SWEVO 383

To appear in: Swarm and Evolutionary Computation BASE DATA

Received Date: 14 July 2017

Revised Date: 9 February 2018
Accepted Date: 18 March 2018

Please cite this article as: R.D. Al-Dabbagh, F. Neri, N. Idris, M.S. Baba, Algorithmic design issues in
adaptive differential evolution schemes: Review and taxonomy, Swarm and Evolutionary Computation
BASE DATA (2018), doi: 10.1016/j.swevo.2018.03.008.

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and all
legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Algorithmic Design Issues in Adaptive Differential Evolution Schemes: Review and Taxonomy

Rawaa Dawoud AL-Dabbagh Ferrante Neri Norisma Idris Mohd Sapiyan Baba

Abstract The performance of most metaheuristic algorithms depends on parameters whose settings essentially
serve as a key function in determining the quality of the solution and the efficiency of the search. A trend that

PT
has emerged recently is to make the algorithm parameters automatically adapt to different problems during
optimization, thereby liberating the user from the tedious and time-consuming task of manual setting. These
fine-tuning techniques continue to be the object of ongoing research. Differential evolution (DE) is a simple yet

RI
powerful population-based metaheuristic. It has demonstrated good convergence, and its principles are easy to
understand. DE is very sensitive to its parameter settings and mutation strategy; thus, this study aims to
investigate these settings with the diverse versions of adaptive DE algorithms. This study has two main

SC
objectives: (1) to present an extension for the original taxonomy of evolutionary algorithms (EAs) parameter
settings that has been overlooked by prior research and therefore minimize any confusion that might arise from
the former taxonomy and (2) to investigate the various algorithmic design schemes that have been used in the

U
different variants of adaptive DE and convey them in a new classification style. In other words, this study
describes in depth the structural analysis and working principle that underlie the promising and recent work in
AN
this field, to analyze their advantages and disadvantages and to gain future insights that can further improve
these algorithms. Finally, the interpretation of the literature and the comparative analysis of the results offer
several guidelines for designing and implementing adaptive DE algorithms. The proposed design framework
M

provides readers with the main steps required to integrate any proposed meta-algorithm into parameter and/or
strategy adaptation schemes.
D

Keywords Metaheuristic algorithms; Optimization algorithms; Evolutionary Algorithms; Genetic Algorithm;

Parameter Control; Differential Evolution; Adaptive Differential Evolution; Review
TE

___________________________________

Rawaa Dawoud AL-Dabbagh, Corresponding Author

Department of Computer Science,
EP

University of Baghdad, Baghdad, Iraq

E-mail: [email protected]

Ferrante Neri
C

Centre for Computational Intelligence,

The Gateway, De Montfort University, UK
AC

E-mail: [email protected]

Norisma Idris
Department of Artificial Intelligence,
University of Malaya, Kuala Lumpur, Malaysia
Email: [email protected]

Mohd Sapiyan Baba

Computer Science Department
Gulf University for Science and Technology, Kuwait
Email: [email protected]

1
 ACCEPTED MANUSCRIPT

1 Introduction

Real-world optimization problems are found in a wide variety of applications, such as machine learning,
system design and nearly all areas (i.e. disciplines) of science and engineering. The extensive research interest
on providing the best solutions to these problems is ongoing. The main challenge such research confronts is
often related to uncertainties and/or noise that can lead to theoretical optima, which are by no means optimal or
practical in real life (Arnold and Beyer 2002; Jin and Branke 2005). Efficient algorithms are known and used to
solve typical optimization problems, whereas heuristic methods are used to solve difficult optimization problems

PT
(Junior et al. 2012a; Junior et al. 2012b). However, the increase in the complexity of the problems, in
conjunction with the need to decrease the time consumed for thorough problem analysis and tailored algorithm
design, implies a demand for robust algorithms with satisfactory performance. These algorithms should not

RI
merely adhere to specific problems but are applicable to a wide range of problems and yield good solutions
(although unnecessarily optimal) (Eiben and Smith 2003; Gendreau and Potvin 2010; Salvatore et al. 2010).
Under this class of optimization algorithms are metaheuristic algorithms, such as evolutionary algorithms (EAs)

SC
and swarm intelligence algorithms (SIA), which are used to find the parameters (or structures) that maximize or
minimize user-defined objective functions. The development of these algorithms continues to be heated and
remains an ongoing research task (Boussaïd et al. 2013; Zhang et al. 2017). Consequently, algorithmic designers

U
or developers must consider another challenge in addition to problem specifics; it is the well design of the
intrinsic part of the meta-algorithm itself (Yang et al. 2013). This includes certain algorithm specifics or
AN
capabilities, such as maintaining the diversity and mixing among the solutions. Well-designed algorithms must
have such capabilities retained for balancing global exploration and intensive local exploitation given that these
two antagonisms constitute the efficient search ability (Fister et al. 2013).
The dialects of all metaheuristics are generally based on the same generic framework whose details need to be
M

specified to obtain a particular algorithm. These details are customarily called algorithm parameters, such as
probability of mutation, probability of crossover, tournament size of selection and size of population. Many
D

studies on EAs examine the implementation of adaptive EAs (Cotta et al. 2008; Lobo et al. 2007). This type of
EAs, if well designed, can enhance the robustness and convergence performance of the algorithm by
dynamically updating the EA parameters for different objective function landscapes during evolution
TE

(Karafotias et al. 2015). This has led to the on-the-fly alteration for such parameters during evolution by
accounting for the actual search progress to achieve optimal convergence and liberate a user from the tedious
trials of tuning the parameters. As discussed in (Eiben et al. 1999) and (Eiben and Smith 2003), the main idea is
EP

to no longer choose the parameters in a semi-arbitrary fashion1 but to allow the parameters to adapt themselves
to the problem.
More than a decade ago, differential evolution (DE) algorithm emerged as a special and competitive form of
C

metaheuristics (Das et al. 2016). Owing to its distinct features, DE algorithm can be viewed from two different
perspectives. It can be classified as EA because of its recombination and selection operators, in addition to the
AC

names and descriptions of its control parameters. At the same time, DE algorithm can also be regarded as SIA
because of its structure that, over a number of generations, a group of initiated particles tend to converge their
values closer to the particle whose value is closest to the target at any given moment (Panigrahi et al. 2011;
Weber et al. 2010). Notwithstanding of its classification, DE algorithm and its numerous variants have
developed rapidly as simple and robust algorithms. Practitioners from different disciplines of science and
engineering have applied DE algorithms to address various optimization problems in their own fields, regardless
of whether these problems are continuous, combinatorial or mixed variable (Abderazek et al. 2017; Al-Dabbagh
et al. 2015a). DE has produced superior results across widely used benchmark functions (Elsayed et al. 2014;

1
The choices were often based on experience.
2
 ACCEPTED MANUSCRIPT

Wang et al. 2012) and real-world applications (Goudos et al. 2017; Zhang et al. 2010). Table 1 covers excerpts
of the most prominent milestones and epochs in the DE history starting from the time it was proposed by Storn
and Price in 1995 (Storn and Price 1995a; Storn and Price 1997) until present. The events in Table 1 are
presented chronologically. This arrangement can be observed from the years that correspond to each subject
(line). However, certain topics require citations from past to recent years. Consequently, two markers have been
used in the Juncture column. The first marker is a small circle that refers to a specific work or publication in the
DE literature. The second marker is a line that refers to an unspecified number of publications in the DE

PT
literature that are related to one topic starting from the year during which the first work regarding that topic was
published. For a subject with a line marker, related and recent references are cited as examples.
Claims and counterclaims have already been proposed, especially by engineers, regarding the rules in
choosing the appropriate control values of standard DE parameters with which to solve practical problems

RI
(Feoktistov 2006; Zou et al. 2011). However, the performance of DE algorithm depends heavily on the selected
mutation strategy and its associated control parameters. The sensitivity of the DE algorithm to its mutation
strategy and its corresponding control parameters can significantly deteriorate its performance if the strategy is

SC
improperly selected (Al-Dabbagh et al. 2015b; Wang et al. 2013). Choosing a suitable DE strategy and setting
its control parameters is difficult and requires much user experience. In the past few years, many researchers
have attempted to make the algorithm into a general and fast optimization method for any kind of optimization

U
problem by tuning its various constitutes, such as initialization, mutation, diversity enhancement, and crossover
of DE (Awad et al. 2017a; Das et al. 2014a). Multiple attempts have also been conducted to automatically adjust
AN
the algorithm’s parameters for single or multiple problems. The development of (self-)adaptive DE algorithms
has resulted in faster and more reliable convergence performance in many benchmark problems than classical
DE algorithms with manual parameter settings (Brest et al. 2007; Zhang and Sanderson 2009b; Zhu et al. 2013).
M

The present study aims to provide insights for fresh and experienced practitioners alike who are interested in
the field of parameter settings in metaheuristics. This study is designed from the ground up to support the issue
of controlling the values of various parameters of an evolutionary algorithm in general and of DE algorithm in
D

particular. The main objective of this study is to present a comprehensive procedural analysis (in SECTION 5)
that is conducted to investigate the various adaptive schemes utilized to automatically control the DE parameters
TE

and/or its mutation strategies. For this purpose, two taxonomies are proposed in this study. The first taxonomy
(in SECTION 3-Fig 2) is proposed to eliminate any ambiguity related to classify any adaptive EA. The new
classification comprises three levels of categories instead of two regarding the parameter control type
(deterministic, adaptive, self-adaptive) and the evidence (absolute, relative) used for determining the change of
EP

the parameter. The second taxonomy (in SECTION 5-Fig 3) is a new taxonomy proposed to classify the
adaptive DE algorithms into two categories (DE with adaptive parameters and DE with adaptive parameters and
mutation strategies). This study comprehensively describes the structural analysis and working principle that
C

underlie the promising work of these algorithms. The study also discusses the advantages and disadvantages of
these algorithms and suggests future insights that can further improve their algorithmic design. Eventually,
AC

protocols for future adaptive DE implementations are also offered (in SECTIONS 6-7).
The rest of this paper is organized as follows. SECTION 2 provides an overview of the published survey work
on adaptive DE algorithms. SECTION 3 presents an extended taxonomy of EA parameter settings. SECTION 4
discusses the standard procedural design of DE algorithm. SECTION 5 applies the two proposed taxonomies (in
Fig 2 and Fig 5) to multiple adaptive DE algorithms specifically as an example to convey the main purpose of
these taxonomies. A procedural analysis is then established on these algorithms to elucidate the conceptual
similarities and differences among them and the pros and cons of each adaptive scheme. SECTION 6 presents a
general framework that lists the steps that should be considered to create a meta-algorithm with parameter
control. SECTION 7 concludes the paper and summarizes the objectives addressed. Suggestions and future work
developments are also offered in this section.
3
 ACCEPTED MANUSCRIPT

Table 1: Historical elucidation of the differential evolution algorithm inventions and developments

Juncture
Subject Citation (publication)

1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
PT
The first technical report was written on DE. (Storn and Price 1995a)
The successful of DE was demonstrated at the First
International Contest on Evolutionary Optimization (1st

RI
(Storn and Price 1996)
ICEO).
Two journal articles published describing DE in detail and its
(Price and Storn 1997; Storn and Price 1997)

SC
very good results
DE was presented at the Second International Contest on
(Price 1997)
Evolutionary Optimization (2nd ICEO)

U
A compendium on DE ‘New Ideas on Optimization’ has
(Price 1999)
been summarized by Price.

AN
Preliminary recommendations on how to choose appropriate (Lampinen and Zelinka 2000; Price and Storn
parameter settings of DE. 1997; Price 1997; Price 1999; Storn 1996)
The first modifications of DE selection rule for constraints (Lampinen 2001; Lampinen 2002; Montes et al.
handling, were presented. 2004)

M
(Fan and Lampinen 2003; Feoktistov and Janaqi
New mutation schemes were presented and some other
2004a; Feoktistov and Janaqi 2004b; Mohamed
developed DE strategies.

D
2015; Mohamed and Mohamed 2017)
Two types of crossover schemes were considered: binomial

TE
and exponential. In 2011, two new crossover schemes were
(Lin et al. 2011; Price 1999; Zaharie 2007)
designed: consecutive binomial crossover and non-
consecutive exponential crossover.
Integration based randomization of individuals with explicit
EP
exploitative component: Memetic differential evolution and (Neri and Mininno 2010; Neri et al. 2008)
its variants.
Structure based randomization of individuals: Compact DE. (Mininno et al. 2011)
C

(Caraffini et al. 2013; Weber et al. 2011; Weber

Increase the search moves of DE with structured population:
AC

et al. 2010) more about parallel DE in (Das et al.

Parallel DE; Distributed DE; Micro-DE
2016)
(Abbass 2002; Brest and Maucec 2011; Das et al.
The era of developing adaptive DE started in 2002 and 2014a; Fan and Yan 2016; Suganthan et al. 2014;
continues to be a very hectic trend until current time (will be Teo 2006; Tvrdik 2009; Urfalioglu and Arikan
discussed in details in SECTION 5). 2011; Viktorin et al. 2017; Wang et al. 2013;
Zaharie 2002; Zhao et al. 2014; Zhu et al. 2013)
Specified number of DE publications Unspecified number of DE publications Unspecified number of DE publications in the field of parameter settings
4
 ACCEPTED MANUSCRIPT

Table 1- Continued

Juncture
Subject Citation (papers)

1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
PT
Many applications of DE to
engineering problems until 2015
(Das et al. 2016; Das and Suganthan 2011)
have been summarized in two

RI
surveys.
Optimization of multidimensional
(Uher et al. 2016)
discrete point clouds (PCs).

SC
DE in a wide range Maximization of trading profit in
(Monakhov et al. 2016)
of real-world financial market.
problems. Photonic structure designs. (Bor et al. 2016)

U
High-precision gear design. (Abderazek et al. 2017)
In communication industry, e.g.

AN
optimization of power consumption
(Goudos et al. 2017; Storn 2017)
of Long Term Evolution (4G LTE)
base stations.

M
Fast 3D path planning. (Yu et al. 2017)
(Dominguez-Isidro and Mezura-Montes 2017;
Elsayed et al. 2014; Kukkonen and Mezura-

D
DE in constrained optimization. Montes 2017; Mezura-Montes et al. 2010;
Mohamed 2017a; Mohamed and Sabry 2012;

TE
Significant Zamuda 2017)
improvements have
(Ali et al. 2012; Lobato and Jr. 2017; Rakshit et
been presented in
DE in multi-objective optimization. al. 2017; Robič and Filipič 2005; Sindhya et al.
many aspects in the
EP
2011; Wang et al. 2016)
overall framework
DE for global optimization in (Brest et al. 2009; Das et al. 2014b; Halder et al.
of DE which are still
dynamic environments. 2013; Novoa-Hernández et al. 2013)
ongoing.
C

(Li et al. 2017; Lynn et al. 2014; Novoa-

Multi-search or multi-population DE.
Hernández et al. 2013; Zaharie 2004)
AC

DE in Multifactorial Optimization
(Feng et al. 2017)
(MFO)
(Das et al. 2016; Das and Suganthan 2011;
Surveys that cover many published aspects of DE (till 2015). Mezura-Montes et al. 2008; Neri and Tirronen
2010; Suganthan et al. 2014)
Specified number of DE publications Unspecified number of DE publications Unspecified number of DE publications in the field of parameter settings

5
 ACCEPTED MANUSCRIPT

2 Related state-of-the-art work: in what way is this survey different to the others?

Since numerous DE variants have been proposed in the literature, many surveys on DE have been published.
Several survey papers have fully captured the performance aspects of adaptive and self-adaptive DE. Das and
Suganthan (2011) published a comprehensive survey that addressed nearly all scenarios where DE is applied,
such as DE and constrained optimization functions, important DE schemes for single-objective functions, DE in
complex environments, theoretical analysis development of DE and most contemporary engineering applications
of DE. The review of Das and Suganthan (2011) also briefly introduced the control of DE parameters and

PT
presented the most prominent and recent DE variants in the field. The survey of Neri and Tirronen (2010)
presented DE algorithms and their most recent advances in a classification format. The reviewed algorithms
were categorized into two classes. The first class is based on integrating DE with an additional component, such

RI
as local search methods. The second class is based on modifying the DE structure, such as algorithms with
adaptive schemes. The study also conducted detailed experiments based on a wide set of benchmark problems to
test the overall performance of these algorithmic classes.

SC
Chiang et al. (2013) published a new taxonomy for DE parameter control mechanisms based on the type of
parameter values (discrete and continuous), number of parameter values (multiple, individual and variable) and
the information used to adjust the parameter values (random, population, parent and individual).

U
Suganthan et al. (2014) published a review paper on adaptive DE variants. In their review, the authors
introduced a simple classification of adaptive DE algorithms according to the type of optimization problem that
AN
they are dealing with: adaptive DE for a single optimization problem, adaptive DE for other problem scenarios
and adaptive population topology in DE. Several adaptive DE algorithms, such as DE with a self-adaptive trial
vector generator (Brest et al. 2006) and DE with an adaptive trial vector generator and parameter control (Qin et
al. 2009) have been listed under the three proposed classes. Although several adaptive DE algorithms with
M

applications in multiple problem scenarios were included in this literature survey, no general adaptive
classification that covered the different aspects of published adaptive DE was proposed. Moreover, no thorough
D

procedural analysis with the pros and cons of these algorithms was provided. Lastly, this review is obsolete.
Das et al. (2016) published an updated version of Das and Suganthan’s 2011 survey paper. In their work, a
complete section was devoted to cover recent adaptive DE algorithms in a classification format, such as DE with
TE

parameters (
and ) and DE with adaptive population size (). However, we wish to raise some points
adaptive scalar factor (
) and crossover rate () parameters only, DE with both adaptive strategy and control

about this section of the survey.

EP

• Algorithms included under this classification are dated between 2011 and 2015. New versions of L-
SHADE, such as LSHADE-cnEpSin (Awad et al. 2017b), have been published recently. We have
C

referred to these versions in our study and other new versions of adaptive DEs have also been cited and
discussed.
AC

• We have found algorithms with special adaptive properties that have been included in classes with

included in the class where
and  are the only adaptive parameters; however, these algorithms also
different adaptive merits. For example, Sarker et al. (2014) and Tanabe and Fukunaga (2014) are

update the value of . Similarly, Zamuda et al. (2013) is included in the class where
and  are the
only adaptive parameters; however, this work also adapts the  value and the trial vector generation

included in the class of adaptive DE with ; however, these algorithms also exhibit the adaptation
schemes. Finally, Brest and Maucec (2011), Zamuda and Brest (2012) and Zamuda et al. (2013) are

property of multiple mutation strategies, which can be overlooked by readers.

• The discussion about the impact of the type of selected mutation strategy and parameter adaptation
scheme is nearly missing. The analysis about the pros and cons of each algorithm (such as in SECTION
6
 ACCEPTED MANUSCRIPT

5) and future work that can be conducted to extend the referenced studies (such as in SECTION 7) are
also missing.
• We have thoroughly discussed a critical issue regarding adaptive algorithms in our paper, which has
been overlooked by many survey studies on this topic. The issue being referred to is the classification of
parameter control (and its extended version) in any metaheuristic approach (SECTION 3). This
classification is proposed to eliminate any ambiguity related to the classification and understanding of
the adaptive characteristics of any adaptive metaheuristic and to complete any survey related to adaptive

PT
algorithms. For example, an adaptive probability-based scheme was used to adapt mutation strategies in

control scheme was used to select the values of
and  for each strategy during the run, which makes
(Bujok et al. 2014) according to the 2016 survey of Das et al. By contrast, a deterministic parameter

this algorithm different from others. These details concerning adaptive algorithms must be highlighted.

RI
Moreover, a survey paper typically attracts new researchers in the field. Thus, sufficient background about the

SC
topic it addresses should be provided.
Finally, the main objectives of these review papers are to report, classify and categorize DE versions presented
in the literature. Unlike them, the current article focuses on adaptation with the purpose to perform the
algorithmic design of DE algorithms. The interpretation of the literature and the comparative analysis of the

U
results offer several guidelines on how to design and implement adaptive DE algorithms (SECTION 6). The
proposed designing framework provides readers with the main steps required to integrate any proposed meta-
AN
algorithm into parameter and/or strategy adaptation schemes. Moreover, we design a state-of-the-art DE
schematic flow by customizing a distinct alphabetical index for each contribution aspect considering the strong
and multifaceted contribution trends of the DE algorithm and our tendency towards simplicity, as depicted in
M

Fig 1. As shown in the figure, some or even all areas of a DE algorithm can overlap in a common work. We
eventually found that surveys that deal with the control of DE parameters (Fig 1-b.2) are rare or obsolete
(Chiang et al. 2013; Das et al. 2016). This research gap has encouraged us to conduct the current comprehensive
D

survey to cover certain aspects related to DE parameter control.

TE
C EP
AC

Fig 1: Combination of differential evolution state-of-the-art flow diagram

7
 ACCEPTED MANUSCRIPT

3 Metaheuristics parameter settings: Extended taxonomy

The critical decision in implementing any meta-algorithm is on how to set the values for various parameters of
that algorithm. These values greatly affect the performance of the evolution. Parameter settings are commonly
composed of crossover rate, mutation step, population size, selection pressure, and penalty coefficient. It is an
important and promising aspect of metaheuristics. As an instance, the efficiency of an EA greatly depends on
the setting of these parameters, that is, by parameter tuning or parameter control (Al-Dabbagh et al. 2015b;
Karafotias et al. 2015). Parameter tuning is also called off-line setting and involves using several “standard”

PT
parameter setting values in advance and keeping these settings fixed during the run, whereas parameter control
is also called on-line setting and involves using another class of approaches where parameters are subject to
change or evolve as problem parameters are optimized. Scientists and practitioners typically tune EA parameters

RI
manually and are guided only by their experience and some rules of thumb. Parameter tuning often requires
tedious and time-consuming human involvement. Moreover, the process of any EA is essentially adaptive and
dynamic process; thus, using fixed parameters with constant values opposes this essence. Intuitively, the values

SC
of these parameters might be optimal at different stages of the evolution; any effort spent toward this direction is
indeed lost a priori (Angeline 1995; Brest et al. 2007; Cotta et al. 2008; Eiben et al. 1999; Eiben and Smith
2003). The downsides or limitations of parameter tuning are as follows (Lobo et al. 2007):

U
Parameter values tuned for a single problem may lead to a large difference in performance if these
AN
parameters were set to different values.
- A parameter tuned for one test problem and produced superior results may not be as effective in other
problems.
- EA parameters are intrinsically dependent; thus, tuning them independently is inconvenient.
M

An alternative form is parameter control, which refers to when an automated setting is applied on EA parameter
values. Globally, the automation of parameter settings encompasses three main categories (Cotta et al. 2008;
D

Eiben and Smith 2003; Lobo et al. 2007):

TE

- Deterministic parameter control – automation occurs when a deterministic rule is triggered to modify
the value of a strategy parameter in a fixed, predetermined manner without using any feedback from the
search.
- Adaptive parameter control – automation occurs during the evolution when the strategy parameter
EP

direction and/or magnitude are adjusted according to a pre-designed rule. Basically, automation
incorporates information gleaned from the feedback based on algorithm performance, such as the
quality of the individual fitness value, without being part of the evolution, where the new control
C

parameter value may or may not persists or propagates throughout the next iterations.
- Self-adaptive parameter control – automation occurs when the strategy parameters undergo genetic
AC

encoding and when the alteration is subject to evolution and pressure (i.e., mutation and crossover);
better parameter values tend to produce better individuals (i.e., solutions) with the highest chance to
survive and propagate for more off-springs.

Another important criterion that should be considered when discussing parameter control techniques is the
evidence of change in parameter value, which can be observed from the performance of operators, the diversity
of the population, and fitness values. Evidence can be absolute or relative. Absolute evidence is when a rule is
applied to alter a strategy parameter value on the basis of a predefined event feedback, such as updating the
probability of mutation rate in accordance with a fuzzy rule set, population diversity drops at some given value,
and even time elapses, rather than being relative to the performance of other values. By contrast, relative
8
 ACCEPTED MANUSCRIPT

evidence is when the strategy parameter value is altered according to the fitness of the offspring produced and
the better is rewarded; this change is specified relative, not deterministically, to one value present at any time.
Therefore, deterministic parameter control is impossible with relative evidence and thus for self-adaptive
parameter control with absolute evidence (Angeline 1995; Cotta et al. 2008; Eiben et al. 1999; Eiben and Smith
2003).
The aforementioned terminologies of the parameter setting of EAs have led to the taxonomy illustrated in Fig.
2. The new taxonomy is an extension of a former one suggested in (Eiben and Smith 2003) which caused some

PT
confusion among a number of researchers working in this field, particularly in distinguishing deterministic and
absolute adaptive rule, as well as relative adaptive rule and self-adaptive rule.

RI
U SC
AN
M
D

Fig 2: Extended taxonomy of parameters settings in EAs

TE

Accordingly, we investigated the subject of parameter control and found new subcategories that can be added to
the main one to address the ambiguity in classification. The definitions of these subcategories are as follows:
EP

- Fully-Deterministic Scheme and Partially-Deterministic Scheme, these two subcategories fall under
Deterministic Parameter Control category. Their main feature is not receiving any feedback from the
search during the evolution. Technically, a fully-predetermined rule is when a user makes a complete
C

counter-intuition on how to steer the control parameter to the desired direction and/or magnitude; for
instance, a rule is triggered on the basis of a certain number of generations that elapsed. By contrast, a
AC

partially-predetermined rule is when one uses random based scheme to alter, for example, mutation
probability after every 100 generations (Fogel et al. 1991).
- Progressive-Controlled Based, this subcategory is applied when some feedback from the search is
discriminated as measurements on the basis of user pre-determined rules. When these measurements
achieve a threshold, a corresponding adaptive rule is applied to update its relative parameter control.
Thus, the progress of updating parameter values, as well as the search, is controlled under
inconsiderable intuition. For instance, these measurements may be based on gathering information from
previous runs through data mining-based fuzzy-knowledge control (Liu and Lampinen 2005),
theoretical considerations (Smith and Smuda 1995), or practical experience encapsulation (Lis 1996). A

9
 ACCEPTED MANUSCRIPT

prominent example of this type of parameter control is the 1/5 success rule of Rechenberg (Rechenberg
1973; Schwefel 1977); this rule is applied at certain periodic intervals deterministically.
- Progressive-Uncontrolled Based (Learning Process-Based), this subcategory is the most prominent in
literature. This sub-classification falls between adaptive rule with relative evidence and self-adaptive
with evolution-process rule, because both rules are intersected on the basis of updating the control
parameter values associated with each individual, at each generation based on their corresponding
fitness value; better features of individuals will be propagated to the next populations over time. The

PT
only difference is that in progressive-uncontrolled rule feedback is gained from the search to allow the
parameter control to gradually adapt by applying a pre-specified strategy, which is most likely analogue
to that of crossover and mutation strategies (Hansen and Ostermeier 1996; Qin and Suganthan 2005;
Zhang and Sanderson 2009b); most of these strategies are learning schemes that collect experience from

RI
previous search. In such methods, the changes performed on the parameter values are fully uncontrolled,
because the adaptive rule is associated only with the “fittest” solutions and its corresponding control
parameters. This subcategory causes much confusion for some researchers working in this field (Zhang

SC
and Sanderson 2009b). For convenience, we use dashed-line connector to make it optional for
researchers who desire to stick with the former taxonomy, not to use the latter one, and include the
learning style under self-adaptive category, as shown in Fig 2.

U
Categories that are involved in the search and gradually evolved by either learning or evolution strategy as long
AN
as the search is not yet terminated, are considered as implicit parameter control, otherwise, are explicit
parameter control.
Ultimately, recent studies have examined the inclusion of algorithms’ operators under the new extended
M

classification (i.e. Fig. 2) as well to yield new algorithms with adaptive parameters and operators. These
algorithms have shown substantial performance improvements compared with algorithms with adaptive
parameters only (Hui and Suganthan 2016; Poláková 2017). Accordingly, self-adaptive algorithms are defined
D

as algorithms with operators and parameters encoded together with their associated solutions and evolved during
the run, whereas adaptive algorithms are defined as separate adaptation-based performance algorithms in which
TE

parameters and operators that yield good solutions are rewarded and propagated to the next generations
(Suganthan 2016), as discussed in detail in SECTION 5.

4 Differential Evolution: Standard structure

EP

Differential evolution (DE) is a robust and competitive metaheuristic method for real parameter optimization
(Storn and Price 1995b). Its effectiveness in solving a wide variety of optimization problems in science and

which makes it easy to implement. Classical DE features few control parameters (scalar factor
, crossover rate
C

engineering domains has been proven (Das and Suganthan 2011). This effectiveness is due to its simple concept,

 and population size ), although they greatly affect the performance of DE (Teo 2006; Zhang and
AC

of stages (mutation, crossover and selection) which is repeated in that order for
 generations.
Sanderson 2009b) which will discussed in the next section. After the initialization stage, DE enters a flow cycle

4.1 DE vector initialization

uniformly and randomly chosen points as initial samples to create population . Usually, each point  of 
DE, like many EAs, is a population-based optimizer that starts attacking the problem by generating multiple,

represents a vector of -dimension in a real parameter space ℝD like  = [ ,  , … ,  ] and is indexed with a
number between 1 and . In the first generation (
= 0), these vectors are initiated according to upper

10
 ACCEPTED MANUSCRIPT

 = [, , , , … , , ] and lower  = , , , , … , ,  bounds of each decision
variable (problem parameter)  of  as

,!" = , + $%&'[0,1], ∙ (, − , ) (1)

where $%&'[0,1] is a uniformly distributed random number generated between 0 and 1.

In DE community, the vectors generated at each stage of DE cycle have different names. For example, the

denoted as .

PT
vectors or population created at the beginning of each generation are called target vectors/target population and

4.2 DE mutation operation

RI
In the preceding stage of generating the target vectors  (, = 1,2, … , ), mutation operation is performed to
generate donor vectors/donor population . (, = 1,2, … , ). The main idea of this operation (see Eq. 2) is to
add a scaled difference of two target vectors (/ and / ) to a third vector /0 from the current population to

SC
yield the mutant vector . . The three indices ($1, $2 and $3) are mutually distinct, chosen from the range
[1, ], and do not coincide with the index , of the current vector. This operation is also called differential

. = / +
∙ (/ − /0 )

mutation.

U
(2)
In Eq. 2,
represents the current generation, and
is a control parameter that determines the exploration length
AN
(/ − /0 ) governing how far the offspring from point / should be generated.
∈ [0,1+[ always takes
positive value which cannot be greater than 1 (Price et al. 2005). DE has many nomenclatures in the literature
M

derived from its mutation formulas. The DE scheme shown in Eq. 2 is called DE/rand/1. Other DE schemes are
expressed below.
DE/best/1 . = 3456 +
∙ (/ − / ) (3)

. =  +
∙ 73456 −  8 +
∙ (/ − / )

D

DE/current-to-best/1 (4)

. = / +
∙ 7/ − /0 8 +
∙ (/9 − /: )

TE

DE/rand/2 (5)

DE/best/2 . = 3456 +
∙ 7/ − / 8 +
∙ (/0 − /9 ) (6)

where 3456 is the best solution found so far in the current generation
.
EP

4.3 DE crossover operation

C

introduce the trial vectors/trial population ; (, = 1,2, … , ). This operation is responsible for increasing the
The next stage that follows mutation in the DE flow cycle is crossover or vector perturbation operation to

diversity among the parameters of target  and . donor vectors. Under this operation, the target and donor
AC

vectors exchange their components with respect to predefined conditions to form the trial vector ; . DE family
has two types of crossover − binary (uniform) crossover and exponential crossover. In what follows, the binary
crossover is defined, which is also known as bin.

., ,= $%&', [0,1] ≤  ?$ @ = @/A

;, = <
, ?BℎD$E,FD
(7)

where $%&', [0,1] is a uniform random number between 0 and 1;  is a DE control parameter just like
. Its
value determines the variation probability among the perturbed vectors.  ∈ [0,1 + [always takes positive
11
 ACCEPTED MANUSCRIPT

value which cannot be greater than 1 (Price et al. 2005). @/A ∈ [1,2, … , ] is the index of a randomly selected
gene to ensure that the offspring ; inherits at least one component from the parent . .

4.4 DE evaluation and selection operations

In the final stage of DE, a greedy selection scheme takes place (see Eq. 8) to measure how far has been achieved

crossover stage (trial vector ; ) with the quality of its corresponding target vector  to determine who will
for the best performance. This scheme is done by comparing the quality of the solution obtained from the

survive to the next generation,  G (, = 1,2, … , ). The quality of solution is calculated using the objective

PT
function = (fitness function) designed for the problem in hand. This operation is described as

; ,= =7; 8 ≤ =7 8
 G
=< 

RI
 ?BℎD$E,FD
(for minimization problems) (8)

SC
From the above equation, if the trial vector has a lower or equal fitness function value, then it replaces the
corresponding target point in the next generation; otherwise, the old point is retained in the next generation. This
strategy guarantees that through generations the population will never deteriorate because the population either
gets better (with respect to minimizing the fitness value) or remains the same.

U
5 Adaptive Differential Evolution: Procedural analysis and comparison
AN
Generally, DE disposes three control parameters,

- Population size (): The total number of potential solutions in one generation
M

- Mutation factor (
): The amount of differentiation ratios that the perturbed solution can acquire
- Crossover rate (): The probability in which the offspring inherits the actual genes of a parent
individual
D

The literature reveals many recent and prominent adaptive DE variants that show efficiency and reliability in
TE

their performance. In this study, notable adaptive DE algorithms (as listed in Fig. 3, there are 28 adaptive DE
variants) have been reviewed and analyzed on a case-by-case basis according to a particular situation
implemented (i.e. parameter control scheme and/or adaptive DE mutation strategy) within the new proposed
EP

taxonomy. Due to space limitation, 19 algorithms out of 28 have been thoroughly presented and discussed; the
other 9 algorithms have been summarized in Table 8. These 19 algorithms are as follows:

• FADE (Fuzzy Adaptive DE) is a parameter adaptive DE in which the control parameters (
and ) of DE
C

jDE is a parameter adaptive DE in which the control parameters (
and ) of DE are adjusted using self-
are adjusted using fuzzy logic (Liu and Lampinen 2005).
•
AC

FiADE (Fitness-Adaptive DE) is a parameter adaptive DE that updates the values of
and  during
adaptive scheme (Brest et al. 2006).
•
evolution based on the fitness function values of the individuals in the population (Ghosh et al. 2011).
•
strategy that updates the values of
and  during the run in a successive learning period (Leon and Xiong
GADE (Greedy Adaptive Differential Evolution) is an adaptive DE algorithm with an effective greedy

DESAP is a parameter adaptive DE in which the control parameters (
,  and ) of DE are all adjusted
2016).
•
though evolution (Teo 2006).

12
 ACCEPTED MANUSCRIPT

• JADE is a parameter adaptive DE in which the control parameters (
and ) of DE are adjusted using
self-adaptive learning scheme (Zhang and Sanderson 2009b).
• DEGL (DE with global and local neighborhoods) is a parameter adaptive DE with a novel mutation
strategy. In this algorithm, new parameter control schemes that balance between the exploitation and

MDE_pBX is a parameter adaptive DE in which the control parameters (
and ) of DE are adjusted
exploration abilities in the DE mutation strategy during evolution is presented (Das et al. 2009b).
•

p-ADE is a parameter adaptive DE that adjusts the parameters of
and  and other control parameters
using self-adaptive learning scheme (Islam et al. 2012).
•

PT
related to its mutation scheme in an adaptive manner (Bi and Xiao 2011).
•
adaptation technique that controls the parameter settings of
and  based on a historical memory of
SHADE (Success-History based Adaptive DE) is an improved version of JADE algorithm. It proposes an

RI
successful solutions and it is updated dynamically during the run. The advantage of the historical memory is
to keep track of the control parameters values that generate the best solutions (Tanabe and Fukunaga 2013).

SC
• L-SHADE is an extended version of SHADE algorithm in which a linear population size reduction (LPSR)
is integrated to continually decrease the size of DE population using a linear function. This strategy has
exhibited efficiency in saving the computation cost of the origin algorithm (Tanabe and Fukunaga 2014).
• EsDEr-NR (Ensemble sinusoidal differential evolution with niching reduction) is an enhanced version

U
of the LSHADE-EpSin algorithm (L-SHADE with ensemble parameter sinusoidal adaptation) presented in

approaches and a Cauchy distribution are used as mixture to update the value of
during evolution. In
(Awad et al. 2016). EsDEr-NR features many significant adaptation merits in which two sinusoidal
AN
addition, the population size  has been reduced during generations using a novel niching-based reduction
scheme. Finally, a restart method is activated at the late stage of evolution to further improve the quality of
M

SaDE is a parameter and strategy adaptive DE in which the control parameters (
and ) of DE as well as
the solutions found so far (Awad et al. 2017a).
•
the DE strategies are adjusted using adaptive techniques (Qin et al. 2009).
D

• EPSDE is a new version of adaptive DE in which an ensemble of control parameters and strategies are
created then selected randomly for each individual (Mallipeddi et al. 2011).
TE

• NRDE (Noise Resilient DE) is an improved DE algorithm with strategy and parameter control proposed for
complex optimization problems, mainly, for functions corrupted with additive noise. This algorithm
proposes three new mechanisms to achieve this objective: adaptive switching between two alternative
EP

SaDE-MMTS is a parameter and strategy adaptive DE in which the control parameters (
and ) of DE as
mutation strategies, blending crossover, and threshold-based selection mechanism (Ghosh et al. 2017).
•
well as the DE strategies are adjusted using adaptive techniques. This algorithm is an integration of SaDE,

SaM (SaJADE) is a parameter and strategy adaptive DE in which the control parameters (
and ) of DE
JADE and local search algorithms (Zhao et al. 2011).
C

•
AC

as well as the JADE strategies are adjusted using adaptive techniques. This algorithm is an improvement of
the JADE (Gong et al. 2011).
•
new adaptive learning scheme to update the value of  during evolution. Both the mutation strategy and
EADE is an enhanced adaptive differential evolution algorithm that presents a novel mutation strategy and a

the adaptive scheme have successfully create the balance between the exploration and exploitation
capabilities of DE and prove effectiveness in solving large-scale optimization problems (Mohamed 2017b).
• EFADE is an enhanced fitness-adaptive differential evolution algorithm with novel mutation. In this

capabilities of DE. In addition, two novel adaptation schemes for controlling the values of
and  during
algorithm, a new triangular mutation strategy is presented to improve the exploration and exploitation

evolution are suggested (Mohamed and Suganthan 2017).

13
 ACCEPTED MANUSCRIPT

To convey the aforementioned information, this section is divided into THREE major subsections.

PT
RI
U SC
AN
M

Fig 3: New classification illustrates the position of each adaptive DE variant with respect to the type of adaptive
D

scheme(s) it applies
TE

5.1 DE with adaptive parameters and single strategy

In this subsection, the main characteristics and mechanisms of 12 remarkable adaptive DE versions are
stated in details on the basis of parameter adaptive schemes and DE strategies.
EP

5.1.1 Adaptive DE with single standard strategy

• FADE Algorithm
C

FADE uses the standard DE scheme DE/rand/1/bin. It updates the values of
and  in each generation using
AC

a mechanism, which is based on the fuzzy logic controller (FLC), whereby a fuzzy knowledge-based system is
used to update the control parameters online in a dynamic adaptive manner to the inconsistent situation.

&6H generations are calculated and then used as input to the FLCs. The values of the control parameters (i.e.
- The values of function values, population diversity (
), parameter vectors (), and their updates after

and ) are the outputs.

- The values of
and  are then assigned to the fuzzy sets membership functions.
- ‘9 × 2’ IF-THEN fuzzy rules statements are used to formulate the conditional statements that comprise fuzzy
logic.

14
 ACCEPTED MANUSCRIPT

- Mamdani’s fuzzy inference method is used as the fuzzy control strategy to map from the given inputs to an
output.
- Defuzzification is done to map from a space of fuzzy output to a space of real output.

• jDE Algorithm

jDE uses the standard DE strategy DE/rand/1/bin. It updates the values of
and  in a self-adaptive manner
based on adjusting the control parameters
and  by means of evolution and applied at the individual level.
First, each individual  !" ( , = 1,2, … ,  ) is associated with its corresponding control parameters
 and  .

PT
In the first generation, these parameters are initialized to
 !" = 0.5 and  !" = 0.9, then
 G and  G
are assigned to values generated according to uniform distributions in [0.1, 1] and [0, 1] respectively during

RI
0.1 + $%&' × 0.9, ,= $%&' < N
evolution as follows:

 G = L

 , ?BℎD$E,FD
(9)

SC
$%&'0 , ,= $%&'9 < N
 G = L
 , ?BℎD$E,FD
(10)

where $%&' ; @ ∈ P1,2,3,4R are uniform random values ∈ [0,1]. In this algorithm, N and N represent the

U
probability limits that permit the adjustment of
and  values, and they are both assigned to the same
AN
value 0.1.
A recent study has been devoted to present a thorough statistical analysis to show the influence of tuning the
randomness level parameter in the performance of jDE-like algorithms. The study has suggested values for the
M

control parameters and the means to select them to better steer the iteration-temporal performance of the
algorithms. Many open issues for the applications of the self-adaptive DE mechanisms into other mechanisms
have also been indicated such as, in the online learning of ensemble strategies (Zamuda and Brest 2015).
D

• FiADE
TE

In this algorithm, novel adaptation schemes for adjusting the values of
and  online are proposed and then
integrated with the standard DE/best/1/bin algorithm. The underlying idea behind these adaptive schemes is the
EP

best individual in the current generation will suffer from decreasing its step-size at the mutation stage to increase
its chances of searching nearby the optimum area (neighborhood). At the crossover stage, the same individual
will have its components propagated to its offspring more than the other individuals in the same population.
Using these strategies creates an appropriate balance between the exploitation and exploration abilities of DE.
C

the second attempt after (Ali and Törn 2004) who proposed a fitness-based adaptation scheme for
. Firstly, in
FiADE is implemented with regard to the fitness function values of the latest individuals obtained, and this is

FiADE, the following two different schemes (Scheme 1 and Scheme 2) are used to adjust the value of
during
AC

evolution.

 =
 ∙ S X
TUV
WGTUV
Scheme 1: (11)
where Δ= = |=( ) − =(3456 ) | and λ = + 10\9 . Number 10
TUV \9
"
by zero when Δ= = 0.
is added to avoid the problem of division

Scheme 2:
 =
 ∙ 71 − D \TUV 8 (12)

15
 ACCEPTED MANUSCRIPT

In both schemes,
 = 0.8 because this setting gives best results over various benchmark functions through
experiments. The two equations above clearly show that when Δ= → 0 then
 → 0. For scheme 1,
 → 0.8
when Δ= is large and λ/Δ= → 0. Likewise, in scheme 2,
 → 0.8 when Δ= is large and D TUV → 0.
These two schemes play a key role in satisfying the objective of generating the best values of
during evolution
in a consecutive manner. Whenever the value of Δ= is greater than a threshold value (in the original paper
(Ghosh et al. 2011) it is 2.4), scheme 2 will always have a greater value of
to keep the exploration process
active. On the contrary, when Δ= drops less than the value of the threshold, the search moves towards premature
convergence as the value of
decreases drastically. At this stage, scheme 1 is used to solve the degradation of
the evolution as it decreases the value of
at a lower rate, thus preventing the algorithm from falling into

PT
premature termination. Therefore, the adaptation scheme of
can be formulated as

 = _% (
 ,
 )

RI
(13)

where
 =
 ∙ 71 − D \TUV 8 and
 =
 ∙ S X.

TUV
WGTUV
The values of  have also been adapted in each generation based on the fitness values obtained for the donor

SC
vectors. As evidence of change, variable Δ=A``/_ = |=(. ) − =(3456 )| has been used as a measurement. If the
value of Δ=A``/_ is low, then the value of  should be higher because the donor vector is close to the best

U
the offspring. On the contrary, when the value of Δ=A``/_ is high, then the value of  should be lower.
individual located in the current population, and thus it possesses good genetic components to be propagated to

Another case is when the donor vector has better fitness value than even 3456 in the same population. In this
AN
case, the  should have been assigned to a very high value. Thus, the scheme for determining  can be

,= =(. ) ≤ =(3456 )
formulated as
M

 = b`56
DcFD
(14)
D

 =  +
(defgh \defVi )
GTUjkikl_V
TE

where b`56 has been assigned to a high constant value of 0.95. Eq. 14 clearly shows that when the value of
Δ=A``/_ → 0, then  moves towards  which is a high value and when Δ=A``/_ → ∞, then  moves
towards  which is a lower value.
EP

• GADE

GADE algorithm proposes an effective greedy mechanism to dynamically update the values of
and  during
C

the evolution. The main idea of this adaptive strategy is that at every learning period (LP), the current
and 
AC

values are compared with their neighboring setting values and then the settings with the best solution are

strategy such as DE/rand/1. It starts with initializing the values of
and  to 0.5 for both parameters. In each
selected. GADE algorithm has a very simple adaptation procedure that can be easily integrated with any DE

generation, the values of
 are selected randomly from the set no = P
− ' ,
,
+ ' R, and the values of 
are selected from the equation below.
 = $%&'p (qde , 0.2) (15)
where the value of the mean distribution, qde is selected randomly from the set nde = P − ' ,  ,  +
' R. The two sets, no and nde represent the neighboring setting values. After a specified learning period
(
%s) the values of
and  are then updated as follows:

16
 ACCEPTED MANUSCRIPT

= %$t _% ()

∈uv (16)

 = %$t _% (z)

w∈uxy (17)

where () and (z) are the progress rates that are used as criteria to evaluate and compare the algorithm
parameters assignments. PR is the average of the relative improvements (RI) from  tests, using the 
assignment to produce trial solutions, and it is calculated as follows:

() = { ∑{
! }(, @)


PT
(18)

=7 ~ 8 ∙ 10 − =7. ~ 8 ∙ 10 , ,= =7 ~ 8 ≥ =7. ~ 8

}(, @) = L
0, ?BℎD$E,FD
(19)

RI
where }(, @) is the relative improvement of the candidate assignment  of either the scaling factor or
crossover rate, € is the test index, and & is an integer number used to set the values of =7 ~ 8 ∙ 10 and =7. ~ 8 ∙

SC
10 within [1, 10] or [−10, −1] ranges.

5.1.2 Adaptive DE with single advanced strategy

• DESAP Algorithm
U
AN
o Advanced DESAP Mutation and Crossover Schemes
M

In DESAP the base strategy used is a bit different from the standard DE/rand/1/bin and of some sort similar
to the strategy introduced in (Abbass 2002).

Crossover Scheme: The crossover operator is performed first with some probability, $%&'(0,1) < / or
, = @, where @ is a randomly selected variable within individual ,. The updating strategy is as follows,
D

bH‚A = / +
∙ (/ − /0 )

TE

(20)

The ordinary amplification factor
is set to 1, thereby at least one variable in  must be changed. Otherwise the
value of bH‚A and its control parameters will be set to the same values associated with / .

Mutation Scheme: The mutation stage is implemented with some mutation probability, $%&'(0,1) < ƒ/ ;
EP

otherwise all the values will remain fixed.

bH‚A = bH‚A + $%&'&(0, ƒ/ )

C

(21)
AC

As can be seen from the equation above, DESAP mutation is not derived from any standard DE strategies.

o DESAP Parameter Control Schemes

DESAP not only updates the values of the mutation and crossover control parameters, ƒ and , but, rather, it
adjusts the population size  as well in a self-adaptive manner. All parameters undergo the evolution (i.e.
crossover and mutation) in a way analogue to their corresponding individuals. The term  has the same meaning
as  and ƒ refers to the probability of applying the mutation scheme whereas the ordinary
is kept fixed

size of both DESAP versions are initialized by generating, randomly, a population of size (10 × &) initial
during the evolution process. Mainly, two versions of DESAP have been applied (Rel and Abs). The population

17
 ACCEPTED MANUSCRIPT

vectors; where & denotes the number of design variables The mutation probability ƒ and crossover rate  are
both initialized to random values generated uniformly between [0,1]. The population size parameter  is
initialized in DESAP-Abs version to,
 = $?;&'(,&,B,%c ?;c%B,?& F,„D + $%&'&(0,1)) (22)
whereas in DESAP-Rel to,
 = $%&'(−0.5,0. 5) (23)

PT
The parameters , η and … are updated at the same level with their corresponding individuals using the same

Updating the crossover rate 

crossover and mutation schemes (see Eq. 20-21).

bH‚A = / +
∙ (/ − /0 )

RI
(24)

bH‚A = $%&'&(0,1) (25)

SC
ƒbH‚A = ƒ/ +
∙ (ƒ/ − ƒ/0 )
Updating the mutation probability η
(26)

ƒbH‚A = $%&'&(0,1) (27)

Updating the population size 

U
DESAP-Abs: bH‚A = / + ,&B(
∙ (/ − /0 ))
AN
(28)

DESAP-Rel: bH‚A = / + ,&B(
∙ (/ − /0 )) (29)

DESAP-Abs: bH‚A = bH‚A + ,&B($%&'&(0.5,1))

M

(30)

DESAP-Rel: bH‚A = bH‚A + $%&'&(0, ƒ/ ) (31)

The ordinary amplification factor
is set to 1. The evolution process of DESAP continues until it achieves a
D
TE

pre-specified population size †. Then, the new population size is calculated for the next generation as,
DESAP-Abs: †4‡ = $?;&'(∑ˆ
 /†) (32)

†4‡ = $?;&'(† + ( × †))

EP

DESAP-Rel: (33)
• JADE Algorithm
C

o Advanced JADE Mutation Schemes

AC

Different mutation versions of JADE have been proposed in (Zhang and Sanderson 2009a) and (Zhang and
Sanderson 2009b). The first new mutation scheme is called DE/current-to-pbest/1/bin (see Eq. 34), which has

the information of the best individual but also the information of the % good solutions in the current
less greedy property than its previous specification scheme DE/current-to-best/1/bin, because it utilizes not only

population.

. =  +
 . (3456 −  ) +
 . (/ − / ),

‰,
(34)

where  ∈ (0, 1] and 3456 is a random uniform vector chosen as one of the superior 100% vectors in the
‰,

introduced to store the recent explored inferior individuals that have been excluded from the search process. Š is
current population. The second mutation scheme with an external archive is denoted as Š, which is an archive

18
 ACCEPTED MANUSCRIPT

first initialized to be empty. Thereafter, failed solutions in the selection operation of each generation are added
to this archive. The new mutation operation is then reformulated as follows:

. =  +
 . (3456 −  ) +
 . (/ − ‹/ ),

‰,
(35)

where  and / are generated from  in the same way as in the original JADE, whereas ‹/ is randomly
generated from the union, Š ∪ . Eventually, randomly selected solutions are going to be removed from the
archive if its size exceeds a certain threshold to keep the archive within a specified dimension. Another JADE

PT
variant has been proposed to further increase the population diversity, and it is named archive-assisted DE/rand-

. = / +
 . ( − / ) +
 . (/ − ‹/0 )

to-pbest/1 as follows:
‰,
3456
(36)

RI
o JADE Parameter Control Schemes

JADE updates four control parameters (
, , qo and qde ) during evolution.

SC
Mutation factor (
) and location parameter of mutation probability distribution (qo ): The mutation probability

 is independently generated in each generation for each individual , according to the following formula:

 = $%&'p (qo , 0.1)

U
(37)
where $%&'p is a Cauchy distribution with location parameter qo and scale parameter 0.1. If
 ≥ 1, then the
value is truncated to be 1 or regenerated if
 ≤ 0. The location parameter qo is first initiated to be 0.5. In this
AN
step, JADE shows some similarities in updating the mean of the distribution, qde to the learning style used in

standard version of the PBIL uses learning rate s ∈ (0, 1] that must be fixed a priori. By utilizing Hebbian-
Population Based Incremental Learning (PBIL) algorithm (Baluja 1994; Baluja and Caruana 1995). The
M

inspired rule, the difference rate (1 − s) is then multiplied by the probability vector () that represents the
combined experience of the PBIL throughout the evolution, whereas s is multiplied by each bit (i.e. gene’s
D

location, qo . It is updated at the end of each generation after accumulating the set of all the successful mutation
value) of the current individual(s) used in updating. Likewise, JADE updates the mutation distribution mean

probabilities
 ’s at generation
, denoted by Žo . The new qde is updated as
TE

qo = (1 − p) ∙ qo + p ∙ _D%& (Žo ), (38)

where _D%& (. ) is Lehmer mean,
EP

∑v∈‘v o 
_D%& (Žo ) = ∑v∈‘v o
(39)

Crossover probability () and mean of crossover probability distribution (qde ): The crossover probability 
C

is updated independently for each individual according to a normal distribution,

 = $%&'& (q , 0.1),
AC

  de
(40)

with mean qde and standard deviation 0.1 and truncated to the interval (0, 1]. The mean qde is first initiated to
0.5. Then, qde is updated in each generation after accumulating the set of all the successful crossover
probabilities  ’s at generation
, denoted as Žde , thus calculate its _D%&’ (Žde ). The new qde is updated as

qde = (1 − p) ∙ qde + p ∙ _D%&’ (Žde ), (41)

where p is a positive constant ∈ (0,1] and _D%&’ (∙) is the usual arithmetic mean.

19
 ACCEPTED MANUSCRIPT

• DEGL (DE with Global and Local Neighborhoods) Algorithm

o Advanced DEGL Mutation Strategy

In this algorithm, a novel mutation strategy named Local and Global Neighborhood-Based Mutations is
proposed. This new scheme has been inspired by the fact that a proper balance must be achieved in the search
evolution between the two objectives, exploitation and exploration capabilities of the search technique. Two

mutation model. The local model is performed to generate a local donor vector c of each individual in the
models have been used to achieve the aforementioned objectives: local neighborhood model and global

PT
c =  + “ ∙ 7_3456V −  8 + ” ∙ 7‰ − • 8
current population as

RI
(42)
where _3456V is the best individual in the neighborhood of  , ‰ and • ; , – ∈ [, − €, , + €];  ≠ – ≠ ,;
€ is the neighborhood radius with value from 0 to ( − 1)/2. This strategy satisfies the first objective (i.e.
improve the local search). A global donor vector t is also generated for each individual as

SC
t =  + “ ∙ 7˜_3456 −  8 + ” ∙ 7/ − / 8 (43)
where ˜_3456 is the best individual in the entire current population; $1, $2 ∈ [1, ]; $1 ≠ $2 ≠ ,. This

U
strategy satisfies the second objective (i.e. improve the global search). In both models, “ and ” are the scaling
AN
factors. These two models are then combined to formulate a general model that produces the actual donor
individual as
. = E ∙ t + (1 − E) ∙ c (44)
where E is a scalar factor with a value in (0, 1).
M

o DEGL Parameter Control Schemes

D

DEGL disposes four new control parameters “, ”, E, and €. However, the most effective parameter control
among them is E because it creates the balance between the exploitation and exploration capabilities in DEGL.
TE

The number of control parameters has been subsequently reduced because “ = ” =
has been considered. In
DEGL, three adaptive schemes have been proposed to update the value of E.

1- Increasing Weight Factor: This scheme increases the value of E from 0 to 1 whenever the search process
EP

requires. When the search moves towards exploration the value of E is decreased to 0. When the exploitation is
required, the value of E is increased towards 1. For this reason, the initial vectors in the first generation are
assigned to E = 0, to activate the exploration moves; the contrary is applied at the final stages of the evolution
C

in which exploitation is required. The middle value of E, 0.5 results in balanced performance in DEGL. Two
AC

schemes have been proposed for this purpose.

Linear increment: E is increased linearly from 0 to 1.

E =


(45)

Exponential increment: E is increased exponentially from 0 to 1.

E = D ™ ∙ c&(2)š − 1


(46)

20
 ACCEPTED MANUSCRIPT

2- Random Weight Factor: For each individual the value of E is generated using uniformal random number,
E ~ (0, 1).

parameter E is embedded in the individual representation and undergoes evolution, this is known as self-
3- Self-Adaptive Weight Factor: As previously discussed in SECTION 3, when the value of the control

adaptive setting, which is expressed as follows:

É = E +
∙ 7E˜_3456 − E 8 +
∙ 7E/ − E/ 8 (47)

where E˜_3456 is the weigth factor associated with the best individual in the current population. E are randomly

PT
initialized in the first generation to values in the range (0.0, 1.0) and then updated during generations.

• MDE_pBX Algorithm

RI
o Advanced MDE_pBX Mutation and Crossover Schemes

SC
˜/žŸ  chosen from the –% group of individuals that were randomly selected from the current population for
Mutation Scheme: The new proposed mutation scheme DE/current-to-grbest/1/bin, utilizes the best individual

each target vector. The group size – of the MDE_pBX varies from 5% to 65% of the . The new scheme can

U
be described as
. =  +
 ∙ (˜/žŸ  −  + / − / ), (48)
AN
where / %&' / are two different individuals randomly selected from the current population, and they are
mutually different from the running individual  and ˜/žŸ  .

Crossover Scheme: The new proposed recombination scheme -Best has been defined as a greedy strategy,
M

which is based on the incorporation of a randomly selected mutant vector perturbed by one of the  top-ranked

the value of parameter  is reduced linearly in an adaptive manner (see Eq. 55).
D

individual selected from the current population to yield the trial vector at the same index. Throughout evolution,
TE

o MDE_pBX Parameters Control Schemes

Modifications applied to the adaptive schemes in MDE_pBX: The scalar factor
 and crossover rate  of each
EP

MDE_pBX, both qo and qde are subscribed to the same rule of adjusting. Firstly, their values are initialized to
individual are both altered independently in each generation using JADE schemes (see Eqs. 37 and 40). In

0.5 and 0.6 respectively. Subsequently, they are updated in each generation as follows:
qo = Eo ∙ qo + (1 − Eo ) ∙ _D%&‰`‡ (
5¡bb455 )
C

(49)

qde = Ede ∙ qde + (1 − Ede ) ∙ _D%&‰`‡ (5¡bb455 )

AC

(50)

where a set of successful scalar factors
5¡bb455 and a set of successful crossover probability 5¡bb455 are
generated from the current population. | | stands for the cardinality of each successful set. Variable & is set to
1.5 then the mean power _D%&‰`‡ of each set is calculated as follows:

_D%&¢`‡ (
5¡bb455 ) = ∑ ∈oŸ¤¥¥žŸŸ(  /|
5¡bb455 |)i

£
(51)

_D%&¢`‡ (5¡bb455 ) = ∑ ∈deŸ¤¥¥žŸŸ(  /|5¡bb455 |)i

£
(52)

including the calculation of the weight factors Eo and Ede ,

21
 ACCEPTED MANUSCRIPT

Eo = 0.8 + 0.2 × $%&'(0, 1) (53)

Ede = 0.9 + 0.1 × $%&' (0, 1) (54)

the qo and qde are formulized. As seen from Eqs. 53-54, the value of Eo uniformly and randomly varies within
the range [0.8, 1], whereas the value of Ede uniformly and randomly varies within the range [0.9, 1]. The small
random values used to perturb the parameters qo and _D%&¢`‡ reveal an improvement in the performance of
MDE_pBX as it emphasizes slight variations on these two parameters each time
is generated.

PT
Crossover amplification factor ( ): Throughout evolution, the value of parameter  is reduced linearly in the


−1
following manner:

RI
 = pD,c ¦ ∙ (1 − )§
2

(55)

where pD,c(z) is the “pD,c,&t” function that outputs the smallest integer ≥ z. The reduction monotony of the
parameter  creates the required balance between the exploration and exploitation capabilities of DE.

SC
• p-ADE Algorithm

o Advanced p-ADE Mutation scheme

U
A new mutation strategy called DE/rand-to-best/pbest/bin is used; which is, essentially, based on utilizing the
AN
best global solution and the best previous solution of each individual that are involved in the differential process,
thus bringing in more effective guidance information to generate new individuals for the next generation. The
M

detailed operation is as follows:

. = ¨ ∙ / + © ∙ 73456 −  8 +
 ∙ (‰3456 −  ) (56)

where 3456 denotes the best individual in the current generation

. / is a random generated individual where
D

$1 ∈ [1, ] and $1 ≠ ,. ‰3456 denotes the best ,6H previous individual picked up from the previous
generation. The mutation’s control parameters ¨ , © , and
 of the , 6H individual are updated using a
TE

dynamic adaptive manner. The most remarkable merit of this mutation technique is the inclusion of three
different working parts at the same time.
EP

Inertial Part (Inheriting part) represented by (¨ ∙ / ) where the current individual, . inherits traits from
another individual at generation
.
Social Part (Learning Part) represented by © ∙ 73456 −  8 where the current individual, . gains
information from the superior individual in the current generation
.
C

Cognitive Part (Private Thinking) represented by
 ∙ (‰3456 −  ) where the current individual, .
AC

reinforces its own perception through the evolution process.

The high values of both the inertial and the cognitive part play a key role in intensifying the exploration

part promotes connections among individuals, thus resulting to speed up the convergence rate. In -ADE there
searching space, thus improving its ability for finding the global solution. While the large values of the social

is an additional mechanism which is called classification mechanism. This classification mechanism is coupled
with the mutation scheme to be implemented on the whole population at each generation. Accordingly, the new
mechanism divides the population’s individuals into three classes:

22
 ACCEPTED MANUSCRIPT

range = − =4 < −ª(=  ), where =4 is the mean fitness values and ª(=  ) is the second moment of
Superior individuals: The first individuals’ category where the fitness values of these individuals fall in the

the fitness values of all individuals in the current generation. In this case, the exploration ability of the
search process is enhanced by further intensifying the inertial and cognitive parts in order to increase the
likelihood of the excellent individual to find the global solution in its neighborhood area. The corresponding
individual is generated as follows:
. = ¨ ∙ / +
 ∙ (‰3456 −  ) (57)

PT
the range = − =4 > ª(=  ). The individual in this case has poor traits since its place in the search space
Inferior individuals: The second individuals’ category where the fitness values of these individuals fall in

is far away from the global optimum. Therefore, the exploration search ability is also intensified for rapid

RI
convergence rate. The corresponding individual is generated as follows:
. = ¨ ∙ / + © ∙ 73456 −  8 (58)

SC
range −ª(=  ) < = − =4 < ª(=  ). The individuals in this category are not superior nor are they inferior;
Middle Individuals: The third individuals’ category where the fitness values of these individuals fall in the

therefore, the complete perturbation scheme (see Eq. 56) should be implemented entirely for further

U
improving both the exploitation and exploration abilities.
AN
o p-ADE Parameter Control Schemes

parameters ( ¨,
and ©) and crossover rate . A dynamic adaptive scheme has been proposed to jointly
p-ADE comprises four control parameters involved in the search process, including three mutation scheme
M

update the four parameters through the run as follows:

¨ = ¨ + (¨ − ¨ ) × ¬S2 − D  S × c &(2)XX ×  + ×  ®
 UV­ \UfVi
­

Ufgh
­ \U ­ (59)
fgh fVi
D

© = © + (© − © ) × ¬SD  S × c &(2)X − 1X × + × ®
 UV­ \UfVi
­

TE

 Ufgh
­ \U ­ 
(60)
fgh fVi

 =
 + (
 −
 ) × ¬S2 − D  S × c &(2)XX × + × ®
6  UV­ \UfVi
­

4  Ufgh
­ \U ­ 
(61)
fVi
EP

 =  + ( −  ) × ¬S2 − D  S × c &(2)XX × + × ®

6  UV­ \UfVi
­

4  Ufgh
­ \U ­ 
(62)
fVi
C

As seen from the above equations, the main adaptive scheme is equally captive to the influence of the number of
AC

within its specified range, ¨ ∈ [0.1, 0.9], © ∈ [0.3, 0.9],
∈ [0.3, 0.9] and  ∈ [0.1, 0.9] during the run.
generations achieved, as well as the fitness values. Technically, the value of each control parameter varies

• SHADE Algorithm

SHADE is an improved version of JADE algorithm specifically in updating the control parameter values of

and . SHADE kept the mutation scheme of JADE w archive (see Eq. 35) as it is but modified the adaptation
strategy of the
and  mean values (qo and qde ). The main idea of SHADE is to no longer generate the
values of
 and  from qo and qde as two distribution mean values are updated once in each generation but
are generated from a historical memory (†o , †de ) that stores the mean values of Žo and Žde each time these
23
 ACCEPTED MANUSCRIPT

two sets are updated with new values of qo and qde , respectively. This new modification improves the
performance of DE in general and JADE in specific due to the inclusion of more values in †o,~ and †de,~
which leads the search towards different search regions especially when dealing with problems of different
scenarios.

o SHADE Parameter Control Schemes

SHADE maintains a historical memory with ¯ entries for the mean values of Žo and Žde and denoted as †o,~
and †de,~ , respectively, and € ∈ [1, ¯]. In the first generation, the values of all †o,~ and †de,~ are initialized to

PT
0.5. When generating the values of
 and  in each generation, a new adaptation scheme is proposed as
follows:

 = $%&'p 7†o,/V , 0.18

RI
(63)

 = $%&'& 7†de,/V , 0.18

SC
(64)
where $ is randomly selected index within [1, ¯]. In case the values of
 and  exceed to 0 or 1, the same
adjusting procedure in JADE is followed. Before the end of each generation, the successful values of
 and 
are stored in Žo and Žde , respectively. The values of these two sets are then used to update the historical
memory of †o and †de as follows:

†o,~G = < U
_D%&E (Žo ) ,= Žo ≠ ∅
†o,~ ?BℎD$E,FD
AN
(65)

_D%&E’ (Žde ) ,= Žde ≠ ∅

†de,~
G
=<
†de,~ ?BℎD$E,FD
M

(66)

The value of € is first initialized to 1 and then incremented by 1 each time an element is added to the memory.
If € reaches a threshold (it is ¯ in SHADE), then € is set to 1. The weighted Lehmer mean _D%&E (Žo ) is
D

computed using the equation below.

∑~! E~ ∙ Žo,~
|± |
v 
TE

_D%&E (Žo ) =
∑~! E~ ∙ Žo,~
|± |
v (67)
EP

To avoid the bias towards small values that may be generated because of the arithmetic mean used in Eq. 41,
Peng et al. (2009) suggested using the following weighted mean (Peng et al. 2009):

_D%&E’ (Žde ) = ∑~! E~ ∙ Žde,~

xy |± |
(68)
C

E~ =
∆U³
´‘xy ´
∑³µ£ ∆U³
(69)
AC

where ∆=~ = ´=7;~ 8 − =7~ 8´. Finally, in JADE the value of  (see Eq. 34) is kept fixed during the run. In the
SHADE algorithm, the value of this parameter is associated with each  and altered during generations as

 = $%&'[ , 0.2]
follows:

where  = 2/.
(70)

• L-SHADE Algorithm

24
 ACCEPTED MANUSCRIPT

This algorithm is an improved version of SHADE algorithm in which a deterministic population resizing
scheme is incorporated to dynamically reduce the size of population according to a linear function. This scheme
is named as linear population size reduction (LPSR), and it is a specialized version of simple variable population

population size  linearly from the initial size which is 6 to the target population size which is 
sizing (SVPS) scheme proposed in (Laredo et al. 2009). The main concept of LPSR is to continually reduce the

 − 6
using the function below.
 G = $?;&' ¦™ š ∙ ª
+ 6 §
†Š¶_
ª
(71)

PT
Where †Š¶_
ª is the maximum number of fitness evaluation, and 
ª is the current number of fitness
evaluation. Based on the equation above, the worst ranking vectors ( −  G ) are deleted
whenever  G <  .

RI
L-SHADE has also suggested a slight improvement to the scheme of generating the values of  , in which a
terminal value “⊥” is assigned to the †de,/V when _%(Žde ) = 0 (i.e. all elements in Žde are 0). This

SC
modification has led to the following updated scheme:

0 ,= †de,/V !·
 = <
$%&'& 7†de,/V , 0.18 ?BℎD$E,FD
(72)

U
The 0 value assigned to  enforces only one parameter change in the original crossover scheme of DE, which
AN
the weighted Lehmer mean is used to calculate the mean values for both Žo and Žde (i.e. Ž) sets using the
implies to slow down the convergence of the algorithm, and it is effective in the multimodal problems. Finally,

equation below for both †o and †de .

∑|±|
~! E~ ∙ Ž~

_D%&E (Ž) =
M

∑|±|
~! E~ ∙ Ž~ (73)

∆=~
E~ =
D

∑! ∆=
|±| (74)
TE

∆=~ = ´=7;~ 8 − =7~ 8´ (75)

• EsDEr-NR Algorithm
EP

In EsDEr-NR algorithm, the advanced mutation strategy proposed in JADE is adopted (see Eq. 35). EsDEr-NR
has three improved parts: ensemble sinusoidal parameter adaptation scheme, niching-based population size
C

reduction, and restart method. They are presented briefly below.

AC

o EsDEr-NR Parameter Control Schemes

As previously suggested in (Awad et al. 2016), in every generation in the first half of the evolution process

∈ [1, fgh
] the values of
 are updated using either one of the two sinusoidal adaptation approaches as

presented below. The preference between these two approaches is implemented randomly.

 = ∙ SF,&(2… ∙ =$D– ∙
+ ¸) ∙ + 1X
sinusoidal decreasing  fgh \
 fgh
(76)
adjustment

 = ∙ ™F,&72… ∙ =$D– ∙
+ ¸8 ∙ + 1š
2

sinusoidal increasing
(77)
adjustment
25
 ACCEPTED MANUSCRIPT

where ¸ is the phase shift value, =$D– is set to a fixed value, and =$D– is set automatically using an adaptive
scheme as expressed below.

=$D– = $%&'p 7qU/4• , 0.18 (78)

Where qU/4• is the Cauchy distribution’s mean with 0.1 as a standard deviation. The value of qU/4• is calculated
using the usual arithmetic mean of the values in ŽU/4• set of the successful mean frequinces.
For the second half of the evolution, where
∈ [ fgh
,
 ], the values of
 are adapted in the same way

suggested in (Tanabe and Fukunaga 2014) using Eq. 63. For the values of  , they are adapted during the run

PT
using Eq. 64. Both
and  values adopted in Eqs. 73-75 are also suggested L-SHADE to calculate the
distribution mean values.

RI
o EsDEr-NR Restart Method

SC
A restart mechanism is activated at the late stages of the evolution when the population size is reduced and the
diversity of solutions is degraded. This mechanism is activated when the size of the population exactly reaches
20 individuals. Half of these individuals are then re-initialized using Eq. 1 and further modified using a

U
Gaussian walk formula as follows:

z =
%;FF,%&(q, ¹) + ($%&'(0,1) ∙ 3456 − $%&'(0,1) ∙  ) (79)
AN
where the mean, q is set to 3456 which is the best individual in the new generated group of individuals;  is the
, 6H individual in the same group of the new individuals; and ¹ is computed as follows:
c?t (
)
¹ = º ∙ ( − 3456 )º
M

(80)

± generations if the new individuals are better than the old individuals.
D

After these computations, a replacement occurs between the worst individuals and the new individuals in the
TE

o EsDEr-NR Niching-based population size reduction scheme

After half of the function evaluations have been consumed, a novel population size reduction mechanism is
EP

activated to ensure that all individuals will have equal chance to evolve before the reduction is performed. From
the start, each individual  will have its Euclidian distance, ª from the best individual, 3456 calculated as
follows:

ª = ª 7 , 3456 8 = »7, − 3456, 8 + 7, − 3456, 8 + ⋯ + 7, − 3456, 8
  
C

(81)

These individuals are then sorted according to their ª values and divided into two niche groups. The first niche
AC

contains the best individuals and the individuals in their neighborhood area (i.e. individuals with the smallest ED

, and since  is decreased after the first half of the evolution in each generation; the size of the two niches
values). The second individual contains the remaining individuals. These two niches have equal size which
{‰

is
is also linearly decreased using Eq.71 in L-SHADE. The individuals in the second niche should be eliminated
after satisfying different criteria.

5.2 DE with adaptive parameters and multiple strategies

26
 ACCEPTED MANUSCRIPT

In this subsection, the main characteristics and mechanisms of 7 adaptive DE versions are stated in detail on
the basis of parameter adaptive schemes and multiple adaptive DE strategies.

5.2.1 Adaptive DE with multiple standard strategies

• SaDE Algorithm

o SaDE Adaptive Strategies

PT
The main feature of SaDE is to automatically adapt multiple standard DE mutation strategies (DE/rand/1/bin,

corresponding control parameters during the evolution process using parameter, ~ (€ = 1,2,3,4).
DE/rand-to-best/2/bin, DE/rand/2/bin, and DE/current-to-rand/1 with no crossover) and update the

RI
Determine the probability of applying each candidate strategy to the current population (~ ): Initially, the
probabilities of applying each scheme ~ is set to 1/© to assign an equal probability for all strategies. The

SC
probability of applying each strategy is then updated every 50 generations in the following manner:

Ž~
~ =
∑~! Ž~
½
(82)

where

∑˜! \¢ &F~

U
AN
\ ˜
Ž~ = + ¾, =?$ € = 1,2, … , ©;
> s
∑˜!
\
\¢ &F~ + ∑˜! \¢ &=~
˜ \ ˜

where © is the number of strategies available for perturbation. s is the period assigned for learning in which
M

the learning process is activated only when

> s; s has been set to 50 generations. &F~ (Success Memory)
and &=~ (Failure Memory) are both memories generated by the €6H strategy and used to record the number of
D

¾ = 0.001, is added to avoid the possibility of the null rate of success. Once the sizes of these memories reach a
trial vectors that have succeeded or failed to enter the search process, respectively. The small value, that is,

certain threshold, (i.e. after s iterations), all previous records will be eliminated from these memories, (i.e.
TE

&F \¢ and &= \¢ ) to allow those vectors that are generated in the current iteration to be stored. Finally, Ž~ is
divided by ∑½ ~! Ž~ to guarantee that the resultant ~ is always summed to 1.
EP

o SaDE Parameters Control Schemes

Set the mutation factor
 values as independently generated in each generation according to Gaussian
C

distribution with mean 0.5 and standard deviation 0.3 as follows:

 = $%&'&(0.5,0.3)
AC

(83)

Accordingly, both the local (with small
 values) and global (with large
 values) search abilities will be kept
throughout the evolution, hence to generate, good mutant vectors.

Crossover Probability ( ) and the Mean Crossover Probability Distribution ( ): The strategy of
controlling the crossover probability  is an adaptive controlled scheme. It starts with independently
generating crossover probabilities  under Gaussian distribution with mean  and standard deviation 0.1
as follows:

27
 ACCEPTED MANUSCRIPT

 = $%&'&( , 0.1) (84)

The  values will remain fixed for five generations before the next generation has launched. Throughout these
generations,  values that are associated with successful trial vectors are recorded. By contrast, the value of
the median  is first initialized to 0.5 and then updated every 25 generations based on the successful
 values as follows:

~! 5¡b (€),

 = ½ ∑½

(85)

PT
where © denotes the number of successful  values accumulated over 25 generations, and 5¡b is the € 6H
 successful value.

RI
• EPSDE Algorithm

SC
o EPSDE Parameters Control Schemes and Strategies

EPSDE is unlike other adaptive DE variants, it is an ensemble of mutation strategies and parameter values of
DE. EPSDE does not involve a certain equation to modify the values of the control parameters. Rather, it

U
assigns for each member of the initial population a mutation strategy randomly selected from a pool of mutation
strategies with diverse characteristics and randomly takes values for the associated parameter from a pool of
AN
values. Throughout evolution, the population members that produce individuals better than the target vectors,
their mutation strategies and associated parameter values are retained for the next generation, whereas those that
fail to produce better individuals are reinitialized with a new mutation strategy and associated parameter values
from the respective pools or from the successful combinations stored with equal probability. EPSDE comprises
M

two pools.

Pool of mutation strategies: This pool includes the DE strategies that are involved in the evolution. These
D

strategies have been selected with the following diverse characteristics:

TE

1. Strategies relying on the best individual in the current population, DE/best/2/bin

2. Strategies bearing strong exploration capabilities, DE/rand/1/bin
3. Strategies being rotational invariant without crossover, DE/current-to-rand/1
EP

ranges. The pool of  values is taken in the range 0.1 − 0.9 in steps of 0.1. The pool of
values is taken in the
Pool of parameter control values: In this pool, the two crucial parameter values (
and ) are set to different

range 0.4 − 0.9 in steps of 0.1.

C

• NRDE Algorithm

In NRDE algorithm different mechanisms have been proposed to adapt the parameter control values of
and
AC

 as well as the two standard mutation strategies of DE (DE/rand/1 and DE/best/1). The work proposed in
NRDE is an extended version of a published work (Das et al. 2015) and as a new attempt after few attempts on
investigating the performance of DE on strongly noisy problems such as the one presented in (Caponio and Neri
2009).

o NRDE Parameters Control Schemes

The control parameters
and  of NRDE are altered using a simple mechanism that requires no prior
knowledge or feedback from the search space during the run such as learning strategy. For each offspring to be
28
 ACCEPTED MANUSCRIPT

generated, the value of
is selected randomly from the set of feasible values P0.5, 2R. At the same time, the
value of  is selected randomly from the interval [0,1] for each donor vector to be generated.

o NRDE Adaptive Mutation Strategies

NRDE proposes a simple mechanism to switch between two alternative DE mutation strategies (DE/rand/1 and

equally probabilistic switch which is 0.5. The reason behind the use of these two strategies is DE/rand/1 is
DE/best/1) without using a pool of strategies. The selection between these two strategies is implemented in an

PT
known as a very effective strategy in creating population with high diversity (i.e. highly explorative), whereas
DE/best/1 focuses on creating solutions by perturbing the best solutions found so far in various search directions
(i.e. highly exploitative).

RI
o NRDE Adaptive Crossover Strategies

between two alternative DE crossover strategies in an equally probabilistic manner which is 0.5. One of these
Using the same switching manner in mutation strategies, NRDE proposes a simple mechanism to switch

SC
strategies is newly proposed in this algorithm as follows:

¿ ∙ , + (1 − ¿) ∙ ., ,= $%&', ≤ 0.5 ?$ @ = @/

;, = <

U
, ?BℎDE,FD
(86)
AN
where the value of parameter ¿ is chosen randomly from three values 0.2, 0.5, 0.9. The main reason behind the
use of these three values is because when ¿ is small (i.e. 0.2), a large portion relative to the donor vector is used.
When ¿ is in the middle (i.e. 0.5), an equal portion is taken from both the target and donor vector. Finally, when
the value of ¿ is 0.9, a gene component is given very near to the target vector. The second crossover strategy is
M

the standard DE crossover defined in Eq.7.

D

o NRDE and Self-Adaptive Individual Selection Strategy

TE

NRDE adopted three cases in the survival selection step instead of two as in most of DE variants because the
main objective of NRDE is to tackle noise or uncertainties in the optimization problems, and the standard greedy
selection strategy may fail to distinguish whether the selected solution is absolutely superior or inferior in
EP

comparison to its parent. These cases are defined as follows:

=(; )
à ; , ,= ≤ 1,
Á =( )
 G
= =(; )
Â; , ,= ≤ 1 + ,
C

=( )
(87)
Á
À  , ?BℎD$E,FD
AC

where  is adapted in a self-adaptive manner using the following formula:

 G
=  ∙ (1 − “) (88)
where , is the generation index, and the value of “ is very small which is 0.04.  is initially set to 0.3.

5.2.2 Adaptive DE with multiple advanced strategies

• SaDE-MMTS Algorithm

29
 ACCEPTED MANUSCRIPT

SaDE-MMTS has been proposed to enhance the performance of the standard SaDE algorithm; by
incorporating SaDE with the JADE mutation strategy and integrating it with the modified multi-trajectory
algorithm (MMTS) to solve problems with complex characteristics and high dimensionality. This integration
can be encapsulated into three main parts: SaDE-MMTS = JADE mutation scheme + SaDE algorithm + MMTS
method (Local Search), as follows:

o SaDE-MMTS Advanced Adaptive DE Strategies

PT
JADE mutation strategy with external archive (see Eq. 35) is adopted and engaged with three crossover
operators (binomial and exponential), and no crossover option as well, to generate the trail vectors for the new
population. The expected number of perturbation strategies is three, and they are applied according to the

RI
probability ~ of applying each JADE with archive strategy in the current population (see Eq. 82).
strategy probability as in the SaDE algorithm. The selection of the mutation strategy is according to the

SC
o SaDE Parameters Control Schemes

The control parameters
and  are updated through evolution in the same manner used in SaDE (see Eqs.
83-84).

o MMTS method

U
AN
The original multi-trajectory (MTS) algorithm (Tseng and Chen 2007; Tseng and Chen 2008) is first
proposed to solve large scale global optimization problems. The main underlying idea of this algorithm is the
employment of randomly selected search combinations (i.e. agents) uniformly distributed over the whole search
M

space to seek out for better solutions. Each combination applies one of the three candidate local search methods

using the basic orthogonal array Ċ×~ , where & is the number of testing experiments and € is the number of
that better fits the search space characteristics of a solution’s neighborhoods. These combinations are generated
D

factors in each experiment.

TE

• SaM

SaM is a strategy adaptation mechanism that can be integrated with any DE variant to make it strategy

individual  from this pool according to Eq. 89.

adaptive. SaM creates a pool of strategies and selects the candidate strategy to be applied on the running
EP

Ž = Ń × ©Æ + 1 (89)
where ƒ ∈ [0,1) is a strategy parameter control variable. © is the total number of strategies in the pool and
C

Ž = 1,2, … , © the selected DE strategy. For example, suppose © = 4 and at a certain generation ƒ ∈ [0,0.25),
then based on the calculation of Eq. 89 the value of Ž is 1.
AC

SaM has suggested three approaches to update the value of ƒ during evolution. In this study, the first approach
has been considered which is inspired by the parameter adaptation equation of JADE. For each individual  at
generation
, a new value for ƒ is generated as follows:
$%&'&
= 1
ƒ = L
$%&'& (q5 , 0.1) ?BℎD$E,FD
(90)

where $%&'& (q5 , 0.1) indicates a normal random distribution of mean q5 and standard deviation 0.1. The mean
q5 is initialized to be 0.5 and then updated at the end of each generation in the same adaptation equation used in
JADE (see Eq. 41) as follows:
30
 ACCEPTED MANUSCRIPT

q5 = (1 − p) × q5 + p × _D%&’ (¯5 ) (91)

where ¯5 denotes as the set of all successful DE strategy parameters ƒ ’s at generation
.
SaM mechanism has been applied on JADE strategies to create a new approach called SaJADE. SaJADE
employed a pool of different JADE strategies: (1) DE/current-to-pBest/1/bin with no archive; (2) DE/current-to-

The parameter adaptive equations used to update the values of
 and  are also JADE schemes as in Eq. 37-
pBest/1/bin with archive; (3) DE/rand-to-pBest/1/bin with no archive; (4) DE/rand-to-pBest/1/bin with archive.

41. For updating the value of  , all strategies use Eq. 40. The difference is in updating the value of
 in

PT
which strategies 1 and 3 use Eq. 37; whereas strategies 2 and 4 use a modified version of Eq. 37 which is normal

 = $%&'& (qo , 0.1)

distribution generator as follows:
(92)

RI
• EADE Algorithm

Advanced Adaptive EADE Mutation scheme

SC
o

This algorithm proposes a novel mutation strategy that contributes to balance the tendency of both the
exploration and exploitation capabilities of DE, as well as enhancing its convergence rate. This new strategy has

U
two aims to guide the search process towards fine evolution. The first aim is to preserve the exploration
capability during evolution and avoid premature convergence at local optima by preventing the individuals from
AN
not always following the best individual in the current population. The second is to preserve at the same time the
exploitation capability by preventing the search process from following the bottom worst individual, thus,
guiding the search towards promising sub-regions in the search space. These two objectives can be met by
applying the new mutation strategy as follows:
M

. G
= / +
 ∙ 7‰_3456 − / 8 +
 ∙ (/ − ‰_‡`/56 ) (93)

where / is randomly chosen individual from the middle ( − 2(100%)) of the current population; ‰_3456
D

is a randomly chosen individual from the top 100 % individuals; ‰_‡`/56 is a randomly chosen individual
from the worst 100% individuals. In both cases,  ∈ (0, 1]; and
 and
 are uniformly and independently
TE

generated in the range (0,1). From the equation above, sharing information from the best to worst individuals in
the same population helps direct the search to different sub-regions in the search space by balancing the local

applied with probability of 0.5 in which the standard DE/rand/1/bin is applied otherwise.
EP

exploitation and global exploration tendencies in the DE population. In EADE, this new mutation strategy is

o EADE Parameter Control Schemes

C

EADE proposes a novel adaptive scheme for gradual change of the values of  that can take benefit from the
AC

experience of the individuals in the search space during evolution which, in turn, can considerably balance the
common trade-off between the population diversity and convergence speed. This scheme can be implemented by
applying a simple and straightforward procedure. This procedure has specific parameters to be defined and set to

Š= [ 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.95 ] of pre-specified values for . These values are selected
values before and during the run. The first parameter is the pool

randomly from Š for each  , using uniform distribution during evolution. Š is updated during and after a pre-
defined learning period (s = 10%
 ). Another parameter is _
c%t_s,FB , which is a flag that is set to
either 0 or 1 to indicate whether the target vector is better than the trial vector or the opposite, respectively.

%,c;$D_p?;&BD$_c,FB is a parameter used to monitor the progress of the individuals’ fitness values after s
completion. If no improvement is observed in the fitness, then the counter of that individual is increased by 1.
31
 ACCEPTED MANUSCRIPT

This process is repeated until †%_=%,c;$D_p?;&BD$ = 20 is reached. Another important parameter is

_%B,?_s,FB~ which is used to record the relative change improvement ratio (RCIR) for each ~ in A as

_,&7´=7; G 8´, ´=7 8´8

_%B,?_s,FB~ = _%B,?_s,FB~ + ¬1 − ®
_%7´=7; G 8´, ´=7 8´8
(94)

Equation 94 is implemented only when the _
c%t_s,FB = 1, which is the case where the trial vector is better
than the target vector. The complete pseudocode of EADE is available in (Mohamed 2017b) in Figs. 2 and 3.

PT
• EFADE Algorithm

o Advanced EFADE Mutation Scheme

RI
EFADE proposed a novel mutation strategy coined as triangular mutation. Its main concept is to enhance the
balance between the global exploration and local exploitation tendencies of DE by investing vectors with

SC
different properties in the same mutation rule. These vectors are as follows:

- 3456 , 34664/ and ‡`/56 are three random individuals selected from the current population ordered

̅b is called a convex combination vector of the triangle (three randomly selected vectors combined in
U
based on tournament selection.
-
AN
the form of linear equation) as:

̅b = E∗ ∙ 3456 + E∗ ∙ 34664/ + E0∗ ∙ ‡`/56 (95)

where E∗ ≥ 0 and ∑0! E∗ = 1; E∗ are called the real weights and are calculated using E∗ =
M

E / ∑0! E , , = 1,2,3. E is calculated using the following equation:

E = 1 − Ì Ì , , = 1,2,3
U( žŸ  )\U( V )
D

U( žŸ  )\U( fgh )

(96)

As a minimization problem, =(3456 ) = =( ) = _,& P=( )R , , = 1,2,3 is the individual with the
TE

minimum fitness value among the three randomly selected vectors. =( ) is the individual with
maximum fitness value in the current population.
EP

Using the vectors above, the new mutation scheme can be formulated as follows:

. G
= ̅b +
 ∙ 73456 − 34664/ 8 +
 ∙ 73456 − ‡`/56 8 +
0 ∙ 734664/ − ‡`/56 8 (97)
where
,
 and
0 are the mutation scalar factors that are generated in each generation using the new
C

adaptation scheme. The mutation scheme above is implemented interchangeably with DE/rand/1 scheme with
AC

respect to a nonlinear probability condition as

}= ™$%&'(0,1) ≥ (1 − ) š ÍℎD&



. G
= ̅b +
 ∙ 73456 − 34664/ 8 +
 ∙ 73456 − ‡`/56 8 +
0 ∙ 734664/ − ‡`/56 8
(98)

ªcFD . G
= / +
∙ 7/ − /0 8
where $%&'(0,1) is a uniform random number generator between 0 and 1, , ≠ $1 ≠ $2 ≠ $3 are indices
randomly selected from the current population, and
is set to 0.7.
32
 ACCEPTED MANUSCRIPT

o EFADE Parameter Control Schemes

EFADE proposes two adaptation schemes to update the values of
and  during evolution. These new
schemes have the advantage of not imposing extra parameters or determining a learning period while adapting
the control parameters.

Scalar factor adaptive scheme: The basic concept of this adaptation scheme is how to dynamically maintain a

purpose, three different scalar factors (
,
 and
0) have been involved in the new mutation scheme. The
proper balance between the local exploitation and global exploration aspects during generations. For this

PT
values of these parameters are updated in each generation as follows:

 = $%&'(0, € ), , = 1,2,3

RI
(99)

=(34664/ ) =(34664/ )
where
à ΠΠ+ ¾, ,= Î Î < 1
Á =(3456 ) =(3456 )

SC
€ =
 =(3456 )
Á Î=( Î + ¾, ?BℎD$E,FD,
À 34664/ )
=(‡`/56 ) =(‡`/56 )
à Î
Á =(3456 )
€ =
 =(3456 )
Î +
U
¾, ,= Î
=(3456 )
Î < 1
AN
Á Î=( Î + ¾, ?BℎD$E,FD,
(100)

À ‡`/56 )
=(‡`/56 ) =(‡`/56 )
à Î
Á =(34664/ )Î + ¾, ,= Î=(34664/ )Î < 1
M

€0 =
 =(34664/ )
Á Î =( Î + ¾, ?BℎD$E,FD,
À ‡`/56 )
D

where $%&'(0, € ) is a uniform random generator between 0 and € , and ¾ = 0.0001 is a small number added to
avoid zero value which implies no perturbation. If € > 1, then its value is truncated to 1. This adaptation
TE

scheme of
,
 and
0 have successfully achieved the desired balance of the two contradictory aspects by
controlling the amplification of the perturbation of the mutation scheme through generations.
EP

 of each target vector from recommended setting values in (Rönkkönen et al. 2005), distributed into two sets.
Crossover rate adaptive scheme: The main idea of this scheme is to select an appropriate control value for the

Then, the selection of these values is updated using learning strategy, depending on their experiences of
C

producing promising vectors through generations as follows:

}=
= 1 BℎD&
 , ,= $%&'(0,1) ≤ 0.5
AC

 = L 
 , ?BℎD$E,FD
ªcFD
(101)
 , ,= $%&'(0,1) ≤ 
 = L
 , ,=  < $%&'(0,1) ≤  + 
where  ∈ [0.05, 0.15] and  ∈ [0.9, 1] are the two suggested sets,  , @ = 1,2, … , _ is the probability of
selecting the  value from set @ which is first initialized to 1/@, and _ = 2 is the total number of sets.  is
dynamically updated during generations using the following equation:

33
 ACCEPTED MANUSCRIPT

 G = S(
− 1) ∙  \ + F X /
, (102)
5Ï­
F = ∑f
ϵ£ 5Ï
­ , (103)
where
5Ï­
F = ∑­ Ð Ð + ¾,
е£ 5Ï G∑е£ UÏ
­ (104)

where F is the probability of selecting the @ 6H set, F is the success ratio of the individuals who successfully

PT
survived to the next generation, &F and &= are two counters for the individuals who survived or failed in the
selection operations, respectively, and ¾ = 0.01 is a very small number to avoid the zero value of F . Arguably,

RI
the small values of  have exert effects at the early stage of evolution because of the high diversity the first
populations always have. By contrast, in the final generations, the large values of  can increase the diversity

SC
of population as the individuals will all be close to the best individual.

5.3 Discussion on adaptive DE variants

U
Nineteen adaptive DE algorithms have been presented in the previous subsections. These algorithms clearly
exhibit diversity in terms of characteristic, structure, complexity and algorithmic logic. Despite their advantages,
AN
these algorithms show shortcomings in some particular cases. In this subsection, comparisons have been
established among these methods according to the taxonomy provided in SECTION 3, their algorithmic design
similarities and differences and pros and cons of each algorithm are provided in Tables 2-7. In addition to these
M

19 DE variants, another 9 important and recent DE algorithms have been summarized in Table 8 to offer more
adaptive algorithmic design insights. Finally, an estimated rank of these algorithms has been presented in Fig. 4.
D

5.3.1 Adaptive DE conceptual similarities and differences

TE

In this subsection, we discuss how the aforementioned methods relate and differ from one another and from
the standard DE algorithm based on mutation strategy and parameter control schemes used in each algorithm.

• Comparison-based DE mutation scheme

EP

The 19 algorithms presented under this comparison differ in their mutation strategies, as seen in Table 2,
which summaries the mutation schemes employed in each algorithm and categorized according to the
taxonomy provided in SECTION 3. From Table 2, we can observe that the most common mutation strategy
C

is DE/current-to-pbest/1 with and without archive. This new strategy has been adopted in many DE
AC

algorithms such as JADE, SHADE, L-SHADE and EsDEr-NR algorithms because this strategy has proved
effective performance. All the algorithms adopt the classical crossover (¿,&) and the standard selection

DE/rand/1 mutation scheme, MDE_BX, which uses a new modified crossover scheme called -Best, and
operations in their main work except DESAP, which uses a modified crossover scheme similar to that of

NRDE that uses a simple mechanism to switch between two crossover schemes, one of these schemes is
newly proposed in NRDE; while SaDE, EPSDE, NRDE, SaDE-MMTS, SaJADE, EADE and EFADE
invent new adaptive schemes that can make a selection among a pool of candidate mutation schemes. In
addition, SaDE-MMTS uses three crossover aspects (binomial, exponential and no crossover) integrated
with their corresponding mutation strategies creating a pool of DE strategies to automatically adapt one of
them.

34
 ACCEPTED MANUSCRIPT

• Comparison-based DE parameter control schemes

The comparison in this subsection is based on the parameter setting taxonomy presented in SECTION 3.
Table 3 elucidates the main features of the adaptation scheme used to control the three main control
parameters (
, , ) in the 19 adaptive DE variants. This table shows how the adaptive concept is

none of the algorithms has considered adapting the population size  except DESAP, L-SHADE and
revealed by applying an adaptive rule to at least one of the algorithm components. From the same table,

EsDEr-NR algorithms. The focus is only on
and .

PT
5.3.2 Adaptive DE strengths and drawbacks

RI
In this subsection, conceptual strengths and drawbacks of the 19 algorithms have been discussed in terms of
modifying the main DE strategies, additional components that have been included in the original DE algorithm
and function as adaptive features. All the modifications and integrations proposed include extra moves in the

SC
original DE as well as create the proper balance between the exploitation and exploration characteristics.
However, drawbacks need to be considered.

• Comparison-based DE mutation strategy

U
In this subsection, the 19 DE algorithms have been compared based on the mutation scheme used in each
AN
algorithm. Tables 4 and 5 illustrate the comparison of algorithms in terms of pros and cons of each mutation
strategy.

• Comparison-based DE parameter control schemes

M

three main parameters (
,  and ) in the 19 DE variants. The two tables indicate common pros and cons
Tables 6 and 7 elucidate the points of strengths and drawbacks of the adaptation scheme used to control the
D

among certain algorithms because they use the same adaptation scheme(s).
TE

5.3.3 Supplemental adaptive DE variants

strategy or ensemble of strategies it utilizes and the adaptive scheme(s) it employs to adjust the values of ,

Table 8 gives a summary of 9 important and very recent adaptive DE algorithms each according to its mutation

and  control parameters. This table gives us an indication that the L-SHADE’s adaptive schemes have
EP

dominated over many other adaptive schemes employed in DE. Likewise, the JADE mutation strategy with
archive (see Eq. 35) has been adopted in many adaptive DE algorithms.
C

5.3.4 Adaptive DE schema ranking

AC

Fig. 4 depicts an estimated rank of 24 adaptive DE algorithms based on their average performance in
comparison with state-of-the-art DE algorithms (as presented in literature). In this figure, the sold arrow
indicates that the algorithm in the source node has inherited the mutation strategy from the destination node;
whereas, the intermittent line refers to the DE variants with less or almost equal performance in comparison with
the DE algorithm in the source node. We can also observe that Fig 4 is divided into distinct layers. Each layer
indicates that the DE algorithms placed in the same layer have almost equal performance.
Generally, Fig 4 shows that the adaptive DE algorithms with advanced mutation strategies such as JADE with
archive (Zhang and Sanderson 2009b), MDE_pBX (Islam et al. 2012) and DEGL (Das et al. 2009b) have gained
the overall best performance. This is because these advanced strategies could handle the deficiencies in the

35
 ACCEPTED MANUSCRIPT

standard DE strategies such as the greediness in the DE/best/1 strategy and successfully managed to create the
required balance between the exploration (i.e. the global search tendency) and exploitation (i.e. the local search
tendency) capabilities of the algorithm. Therefore, we can observe from the same figure that the DE algorithms
placed in the top layers such as SHADE (Tanabe and Fukunaga 2013) have already adopted these advanced
strategies (e.g. DE/current-to-pBest/1 strategy in Eq. 35) and incorporate them with the parameter adaptive
schemes to achieve the best algorithmic design performance. Likewise, adopting effective algorithms’ schemes
also applies to the parameter control schemes. LSHADE-cnEpSin (Awad et al. 2017b), L-covnSHADE (Awad

PT
et al. 2017c) and LSHADE-SPA (Mohamed et al. 2017) algorithms are already a modified versions of L-
SHADE algorithm’ parameter control schemes (Tanabe and Fukunaga 2014). Another important type of
adaptive DE variants is the dynamic selection of multiple DE strategies during the evolution process such as in
ESPSDE (Mallipeddi et al. 2011), HSPEADE (Mallipeddi 2013) and CoDE (Wang et al. 2011). This type of

RI
adaptive algorithms has always show good performance and still a very hectic research trend; because, it has
already been proved through experiments that at each stage of evolution there is a suitable DE strategy to
impose different search step size and guide the search towards better direction. As such, EDEV algorithm (Wu

SC
et al. 2018) outperforms JADE with archive, CoDE, SaDE and EPSDE algorithms with outstanding
performance as this algorithm proposed an adaptive scheme that manages to select the suitable adaptive DE
scheme during the run.

U
AN
M
D
TE
C EP
AC

36
 ACCEPTED MANUSCRIPT

Table 2: A summary of the type of mutation strategies used in 19 adaptive DE algorithms

Algorithm Mutation Scheme Main Feature of the Scheme Base Scheme

PT
FADE DE/rand/1 Fixed DE/rand/1
jDE DE/rand/1 Fixed DE/rand/1
FiADE DE/best/1 Fixed DE/best/1

RI
GADE DE/rand/1 Fixed DE/rand/1

SC
DESAP (non-DE standard scheme) With deterministic mutation probability updates Simple Perturbation Scheme

Utilization of the 100% best solutions and the Š

JADE wo DE/current-to-best/1 Utilization of the 100% best solutions DE/current-to-best/1
JADE w
DE/current-to-best/1 w archive DE/current-to-best/1

U
archive inferior solutions
Two stages mutation: Local and Global to

AN
Local and Global Neighborhood-
DEGL balance the exploitation and exploration DE/target-to-best/1
Based Mutations
capabilities

–% group of individuals
DE/current-to-t$3456 /1
Utilization of the best individual selected from

M
MDE_ÑBX DE/current-to-best/1

Utilization of the 100% best solutions and the Š

p-ADE DE/rand-to-best/best Adaptive Dynamic Structure PSO Standard Scheme (Kennedy and Eberhart 1995)

D
SHADE DE/current-to-best/1 w archive DE/current-to-best/1

Utilization of the 100% best solutions and the Š

inferior solutions

TE
L-SHADE DE/current-to-best/1 w archive DE/current-to-best/1
inferior solutions
EsDEr-NR DE/current-to-best/1 w archive Utilization of the 100% best solutions DE/current-to-best/1
EP
DE/rand/1; DE/rand/2; DE/current-to-rand/1;
SaDE Pool of standard DE strategies Adaptive PCB among the schemes
DE/rand-to-best/1
C

DE/best/1; DE/best/2; DE/rand-to-best/1; DE/rand-

EPSDE Pool of standard DE strategies Deterministic/Partial-predetermined to-best/2; DE/rand/1; DE/rand/2; DE/current-to-
AC

rand/1
LPB: Learning Process Based PCB: Progressive-Controlled Based EPB: Evolution Process Based

37
 ACCEPTED MANUSCRIPT

Table 2-Continued

Algorithm Mutation Scheme Main Feature of the Scheme Base Scheme

Equally probable switch between

NRDE Deterministic/ Partial-predetermined DE/rand/1 and DE/best/1

PT
two alternative mutation strategies
Pool of DE/current-to-best/1 w
SaDE-MMTS merged with two crossover schemes Adaptive PCB among the schemes DE/current-to-best/1 w

RI
and no crossover
DE/current-to-best/1 wo; DE/rand-to-best/1 wo;
SaJADE Pool of advanced DE strategies Adaptive LPB among the schemes

SC
DE/current-to-best/1 w; DE/rand-to-best/1 w;

same scheme. Top and bottom of 100% and

Utilizes three different types of individuals in the

middle of  − 2(100%).

U
Top, middle and bottom individual
EADE DE/rand/1
based mutation Deterministic/Partial-predetermined scheme is

AN
used to alternatively switch between the new
scheme and DE/rand/1 scheme.

M
Utilizes the best, better, and worst individual
selected randomly from the current population.
Triangulation scheme with linear
EFADE Deterministic/Partial-predetermined scheme is DE/rand/1

D
convex combination vector
used to alternatively switch between the new
scheme and DE/rand/1 scheme.

TE
LPB: Learning Process Based EP PCB: Progressive-Controlled Based EPB: Evolution Process Based
C
AC

38
 ACCEPTED MANUSCRIPT

Table 3: This table encompasses taxonomy of the adaptation scheme used to update the main control parameters in 19 adaptive
DE algorithms

Parameter control strategies based taxonomy

Algorithm ÒÑ Ó ÔÕ Ö×ØÙÚ ÛÖÜ×ÚÖÝ ÑÞÚÞßÙ×ÙÚà

PT
Name
Type of Type of Type of Setting Type of Type of Setting Type of Parameter Type of Type of
Setting Evidence Evidence Evidence Setting Evidence
;


RI
FADE Tuned × Adaptive/PCB Absolute Adaptive /PCB Absolute Adaptive /PCB Absolute
Self-adaptive/
jDE Tuned × Self-adaptive/EPB Relative Relative × × ×

SC
EPB
FiADE Tuned × Adaptive/LPB Relative Adaptive/LPB Relative × × ×
;}; ;
qde

U
GADE Tuned × Adaptive/LPB Relative Adaptive/LPB Relative Adaptive/LPB Relative

AN
Self-adaptive/ Self-adaptive/ Self-adaptive/
DESAP Relative Tuned × Relative η Relative

qo ; qde
EPB EPB EPB

qo ; qde
JADE wo Tuned × Adaptive/ LPB Relative Adaptive/ LPB Relative Adaptive/ LPB Relative

M
JADE w Tuned × Adaptive/ LPB Relative Adaptive/ LPB Relative Adaptive/ LPB Relative

“; ”; € Tuned ×

D
Deterministic/

TE E
DEGL Tuned × Tuned × Tuned × Partial-
Absolute and
predetermined;
Relative
self-
EP
qo ; qde
adaptive/EPB

Eo ; Ede
Adaptive/ LPB Relative
Deterministic/
C

MDE_ÑBX Tuned × Adaptive/ LPB Relative Adaptive/ LPB Relative

Partial- Absolute
AC

Ñ-ADE ¨; ©
predetermined
Tuned × Adaptive/ LPB Relative Adaptive/ LPB Relative Adaptive/ LPB Relative
LPB: Learning Process Based PCB: Progressive-Controlled Based EPB: Evolution Process Based

39
 ACCEPTED MANUSCRIPT

Table 3- Continued

ÒÑ Ó ÔÕ Ö×ØÙÚ ÛÖÜ×ÚÖÝ ÑÞÚÞßÙ×ÙÚà

Parameter control strategies based taxonomy
Algorithm
Name Type of Type of Type of Setting Type of Type of Setting Type of Parameter Type of Type of

†de ; †o ; E~

PT
Setting Evidence Evidence Evidence Setting Evidence
Adaptive/ LPB Relative


Deterministic/
SHADE Tuned × Adaptive/ LPB Relative Adaptive/ LPB Relative

RI
Partial- Absolute

†de ; †o ; E~ Adaptive/ LPB

predetermined

SC
Relative


Deterministic/
L-SHADE Adaptive/PCB Absolute Adaptive/ LPB Relative Adaptive/ LPB Relative
Partial- Absolute

=$D– ; qU/4• ;

U
predetermined

†o ; †de ; E~

AN
Adaptive/LPB Relative

z ; ¹
EsDEr-NR Adaptive/PCB Absolute Adaptive/ LPB Relative Adaptive/ LPB Relative Deterministic/
Partial- Absolute

M
predetermined

; Ž
Deterministic/

D
SaDE Tuned × Partial- Absolute Adaptive/ PCB Absolute Adaptive/ PCB Absolute
predetermined

TE
EPSDE Tuned × Adaptive/PCB Absolute Adaptive/PCB Absolute × × ×


Deterministic/ Deterministic/
NRDE Tuned × Partial- Absolute Partial- Absolute Adaptive /PCB Absolute
EP
predetermined predetermined

; Ž
Deterministic/
SaDE-
Tuned × Partial- Absolute Adaptive/ PCB Absolute Adaptive/ PCB Absolute
C

MMTS

qo ; qde ; q± ;S
predetermined
AC

SaJADE Tuned × Adaptive/ LPB Relative Adaptive/ LPB Relative Adaptive/ LPB Relative

_%B,?_s,FB Adaptive/ LPB

Deterministic/
EADE Tuned × Partial- Absolute Adaptive/ LPB Relative Relative

; F; F
predetermined
EFADE Tuned × Adaptive/ LPB Relative Adaptive/ LPB Relative Adaptive/ LPB Relative
LPB: Learning Process Based PCB: Progressive-Controlled Based EPB: Evolution Process Based

40
 ACCEPTED MANUSCRIPT

Table 4: DE algorithms points of strengths based on mutation strategy

Algorithm Points of strengths of the DE mutations

The standard DE/rand/1(Eq. 2) scheme is always considered as the fastest non greedy scheme with good convergence performance (Brest et al.
FADE; jDE
2006).

PT
The JADE mutation (Eq. 34) alleviates the problem of premature convergence because of its ability to intensify the population diversity (Zhang and
JADE
Sanderson 2009b) .
• Increase the population diversity as such the problem stem from the convergence rate was further reduced. This is so because, the superior and

RI
inferior solutions are both incorporated into the mutation strategy.
JADE with
SHADE and L-SHADE over JADE algorithm have the advantages that they dynamically update the value of  top solutions using a
• No significant computational overhead as the archive operation has been made very simple.
archive;
•

SC
SHADE;
L-SHADE; deterministic based technique.
EsDEr-NR The above points are discussed in (Awad et al. 2017a; Tanabe and Fukunaga 2013; Tanabe and Fukunaga 2014; Zhang and Sanderson 2009b)
using Eq. 35. The literature is already rich with other analysis and modified versions of SHADE algorithm as this algorithm proved effectiveness in

U
its performance (Viktorin et al. 2017).
The local and global neighborhood topology (ring) in the new DE mutation (Eq. 44) creates the proper balance between the exploitation and
DEGL

AN
exploration capabilities of the algorithm during evolution (Das et al. 2009b).
It weaknesses the tendency of premature convergence and alleviates the attraction of any fixed point in the fitness landscape. This modification in
MDE_ÑBX (Eq. 48) has led to reduce the greediness feature of the DE mutation scheme towards choosing the superior solutions for perturbation and making it

M
converges fast to a local point (Islam et al. 2012).
Ñ-ADE
The ability to be dynamically changed to three different schemes in (Eq. 56) according to the classification conditions upon the individuals’ quality
that are located in the same population (Bi and Xiao 2011).

D
Learning strategies (Eq. 82 and Eq. 89) have been applied to gradually evolve the selection of one mutation scheme from a pool of mutation
SaDE;
schemes in hand throughout generations. This characteristic allows further improvements in the DE moves, thus increasing the exploitation features
SaDE-MMTS;

TE
of the algorithm. This learning strategy affords flexibility to be extended to include more candidate mutation schemes in an attempt to solve complex
SaJADE
optimization problems (Gong et al. 2011; Qin et al. 2009; Zhao et al. 2011).
The selection of the mutation strategies is made randomly from a pool of mutations with different characteristics to avoid the influence of less
EPSDE
effective mutation strategies (Mallipeddi et al. 2011).
EP
The mutation strategy adopted (Ghosh et al. 2017) is simple but effective. It alternatively switches between two extreme mutation strategies
NRDE (DE/rand/1 Eq. 2 and DE/best/1 Eq.3). The use of both strategies will enhance the search process as each scheme has its capability to either provide
high explorative or high exploitative depending on the nature of strategy.
C

It (Eq. 93) successfully maintains the global exploration and local exploitation of the search process through investing components from the best,
EADE
middle and worst vectors in DE population (Mohamed 2017b).
AC

A novel mutation scheme named as triangulation mutation (Eq. 97-98) based on convex combination vector has been proposed to successfully
enhance the global and local search capabilities. This new scheme sorts three randomly chosen vectors into best, better and worst and employs them
EFADE
in the weighted convex equation and the main mutation scheme. The different components inherited from these vectors have helped to archive the
proper balance between the two contradictory aspects of the search process (Mohamed and Suganthan 2017).

41
 ACCEPTED MANUSCRIPT

Table 5: DE algorithms drawbacks based on mutation strategy

Algorithm Drawbacks of the DE mutation

Lack of Population diversity
FADE; jDE This problem of Eq. 2 will arise in optimizing high dimensional functions and also when the characteristic of the test problem is challenging

PT
(Brest et al. 2006; Liu and Lampinen 2005).
Lack of Exploitation and Exploration Ability
No significant improvements over the standard DE. The new crossover and mutation operations (Eq. 20-21) are simple and straightforward
DESAP
schemes, they did not bring about the desired performance and DESAP outperform the standard DE in only one out of five test problems (Teo

RI
2006).
JADE;

SC
JADE with Stagnation
archive; The population selective factor,  in Eq. 34-36 is tuned before the run and kept fixed during the evolution process. This strategy might lead to

% of the current population start to be close in values if not exactly same (Gong et al. 2011; Zhang and Sanderson 2009b; Zhao et al. 2011).
SaDE-MMTS; stagnation problem especially after several epochs of evolution when the population diversity rate is very low and the superior solutions in the

U
Although, SHADE and L-SHADE suggested using deterministic based technique to update the value of  as in Eq. 70, it still has the
SaJADE;
SHADE;

AN
L-SHADE; aforementioned problem may occur at the late stages of evolution (Tanabe and Fukunaga 2013; Tanabe and Fukunaga 2014).
EsDEr-NR

M
DEGL Limited test suit functions have been used to test the effectiveness of DEGL.
This implies that DEGL may not be able to model real-world complex problems (Das et al. 2009b).
Lack of strategy and parameter analysis &

D
Crossover greediness tendency
• The influence of the new mutation scheme in Eq. 48 on the diversity of population and convergence rate is not investigated.

TE
MDE_ÑBX
• The greediness of the new crossover scheme towards superior solutions even with the associated dynamic parameter,  (see Eq. 55) of the
top-ranking vectors may lead to premature convergence problem.
The above two points are discussed and investigated in (Islam et al. 2012).
EP
Ñ-ADE
Local optimum caused by greedy mutation
Selecting the best solutions from the previous and current population may lead the new strategy in Eq. 56 become greedier towards good
solutions, thus running the risk of falling into local optimum (Bi and Xiao 2011).
C

The use of equibrobale switch between the two mutation strategies is challenging
AC

NRDE The two mutation strategies adopted (Eq. 2 and Eq. 3) are simple and one of them is greedy strategy. This may drift the search towards falling
into local optima (Ghosh et al. 2017).

42
 ACCEPTED MANUSCRIPT

Table 6: DE algorithms points of strengths based on parameter control schemes

Algorithm Points of strengths of parameter control schemes

• The first attempt in using Fuzzy Control in DE parameter settings for the sake of reducing the user load from parameter tuning.
FADE • Possess robustness and fuzziness with problems in an imprecise environment.

PT
As discussed and presented in (Liu and Lampinen 2005).
jDE Adjust both
(see Eq. 9) and  (see Eq. 10) in a self-adaptive manner with few additional parameters (Brest et al. 2006).
The adaptation schemes of
(see Eq. 11-13) and  (see Eq. 14) are very simple and yet effective. The best feature of these schemes is

RI
FiADE
The greedy adaptation procedure for adjusting the
and  values used in GADE repeatedly finds the better parameter assignments by
their strategy to work alternatively as the evolution process requires (Ghosh et al. 2011).

SC
searching in the neighborhood area of the current assignments. It has two significant advantages (Leon and Xiong 2016).
GADE
• It offers a good chance for the better trial solutions to survive.
• It sustains the improvement rate of the fitness values in the next generations.
• The first attempt to demonstrate the possibility to produce an adaptive algorithm that not only updates the crossover (see Eq. 24-25)

U
and mutation rates (see Eq. 26-27) but also the population size parameter as well as in Eq. 22-23.
DESAP
•

AN
Downsize DE parameters’ setting by updating the population size and other control parameters in an adaptive manner.
As discussed and presented in (Teo 2006).
Adjusting the values of
and  in an adaptive characteristic based on a learning strategy that gains knowledge from previous
iterations and cumulates it into two learning parameters, qo and qde to be retained and used in the current population (see Eq. 37-41).
JADE; •

M
JADE with
archive; • Create the proper balance in maintaining the pressure on the population to move towards exploring more optimal solutions, as well as
SaJADE not to lose the exploitation features (Gong et al. 2011; Zhang and Sanderson 2009b).
The scalar factor E (see Eq. 45-47) has been used to control the behavior of the mutation during evolution. It helps to provide the required

D
DEGL

Modifies the original JADE scheme of adapting the values of
and . The new modifications to the control parameters schemes with the
balance between the local and global moves in the search space (Das et al. 2009a).

MDE_ÑBX
TE
combination of the new mutation and crossover schemes (see Eq. 49-55) in MDE_BX have greatly increased the ability of exploitation
and exploration, and the search process is directed to explore better search space regions, hence, escaping from the possibility of getting
EP
trapped in suboptimal solutions (Islam et al. 2012).
• Possess a unique merit over other adaptive DE algorithms by involving the number of iterations passed over and the fitness value in
Ñ-ADE
C

the updating process; for the sake of accelerating the convergence rate (Bi and Xiao 2011).

values for the control parameters ¨, ©,
, and  as expressed in Eq. 59-62.

• Creating the required balance between the exploitation and exploration features. This has been achieved through selecting the best
AC

• The former adaptation schemes of JADE for
and  have been improved by applying a new dynamic scheme based on accumulating
the successful parameter control settings in a historical memory updated during evolution (see Eq. 63-70 and Eq.72-75). This strategy

may have been included in Žde and Žo is resolved (Tanabe and Fukunaga 2013; Tanabe and Fukunaga 2014).
SHADE; L- has the advantage of guiding the search process towards better solutions as the negative impact caused by the poor setting values that
SHADE
• L-SHADE continually decreases the population size using a linear reduction equation that positively reduces the computation cost of
the algorithm (see Eq. 71) as presented in (Tanabe and Fukunaga 2014).

43
 ACCEPTED MANUSCRIPT

Table 6- Continued

Algorithm Points of strengths of parameter control schemes

• The two sinusoidal adjustments (increasing and decreasing) have the advantage of updating the value of
in such a way that drifts the

• The niche-based mechanism suggested to adapt the  (see Eq. 81) value has the advantage of overcome the problem of stagnation
search direction to different regions as expressed in Eq. 76-78.

PT
EsDEr-NR
that the algorithm is probably falls in when reducing the population size.
• The restart method (see Eq. 79-80) has the advantage of enriching the later generations of the search with new solutions.

RI
The aforementioned three points on EsDEr-NR performance is discussed in (Awad et al. 2017a).
SaDE ; Utilizes an adaptive learning strategy to adjust the crossover rate (see Eq. 85) and another learning scheme to adaptively select mutation
SaDE-MMTS strategy during the search process as provided in Eq. 82 (Qin et al. 2009; Zhao et al. 2011).

SC
The selection of the mutation and crossover parameter values is made randomly from a pool of values to avoid the influence of less
EPSDE

The adaptation mechanism used to modify the values of
and  is simple. It requires no knowledge from the search space or
effective parameter settings (Mallipeddi et al. 2011).

U
•
feedback. No means or variance, etc or any other parametric probability distribution parameters are involved. It only needs to know

AN
The additional control parameter  (see Eq. 87-88) has a significant effect on the performance of the new proposed DE selection
the feasible ranges of these parameters (Ghosh et al. 2017).
NRDE •
scheme. The use of this self adaptive parameter in conjunction with the new selection scheme will improve the search process from

M
getting trapped into a basin of optima attraction. This is practically achieved by including inferior solutions occasionally survived
using the new selection scheme (Ghosh et al. 2017).
The crossover rate has been modified during the run using an adaptive rule (see Eq. 94) that collects information from past experiences of

D
EADE
successful and unsuccessful individuals during and after a learning period (Mohamed 2017b).

values of
and  as expressed in Eq. 99-104. Both schemes show advantages to balance the local exploitation and global exploration

TE
No extra parameters or learning period burdens. The adaptive process is implemented using simple and clear schemes for updating the
EFADE
capabilities of the algorithm (Mohamed and Suganthan 2017).
C EP
AC

44
 ACCEPTED MANUSCRIPT

Table 7: DE algorithms drawbacks based on parameter control schemes

Algorithm Drawbacks of parameter control schemes

Lack of algorithm performance analysis
FADE It does not involve any relative consideration for the individual fitness value only absolute based on the knowledge gained from the

PT
fuzzy controller (Liu and Lampinen 2005).

There is a weakness in the relative consideration of the individual fitness value since the values of
and  are selected according to the
Lack of balance between the exploitation and exploration
jDE

RI
best fitness picked up from the current generation only, then updated according to a uniform distribution (see Eq. 9-10). This may lead
jDE to be biased towards exploration (Brest et al. 2006).
Limited test-suit
FiADE

SC
Its simplicity may lead FiADE to fail to model very complex optimization problems (Ghosh et al. 2011).
Greedy adaptive scheme
GADE The greediness tendency of the new adaptive scheme in GADE towards the parameter assignments that produce the best solutions may

U
lead to a stagnation problem in DE (Leon and Xiong 2016).
Lack of population diversity

AN
• It outperforms the standard DE over five test functions only.
DESAP • It was found that both DESAP’s versions yielded highly similar results in terms of the best solution obtained; although, in providing
more stability DESAP with absolute encoding was more favorable than DESAP with relative encoding.

M
This can be observed from the experiments provided in (Teo 2006).

The parameters p and  determine the adaptation rate of qde and qo and the greediness of the mutation strategy are kept fixed during the
JADE; Problem-dependent parameter selection
JADE with

D
archive; run (see Eq. 34-36). These two parameters have the ability to affect the overall performance of JADE mutation (Tanabe and Fukunaga
SaJADE 2013; Zhang and Sanderson 2009b).

TE
The adaptive schemes of
and  depend on the constraint =(; ) < =( ) because of the weighted scores associated with these
JADE with Inappropriate parameters update

schemes; so, when =(; ) = =( ) the weight becomes 0 resulting in inappropriate parameter settings (Tanabe and Fukunaga 2013).
archive; SHADE;
L-SHADE
EP
There are two additional control parameters – (the group size in the mutation operation in Eq. 48) and  (the number of the top-ranking
Lack of theoretical guidelines
MDE_ÑBX
vectors in the crossover operation in Eq. 55). These parameters bring about the effect to the performance of the mutation and crossover.
C

There is lack on the theoretical evidence of the importance of these two control parameters (Islam et al. 2012).

Ñ-ADE All of the four parameters (¨, ©,
, and ) should be adjusted through the run (see Eq. 59-62), simultaneously with adjusting the
Time consumption
AC

mutation scheme (see Eq. 56-58). This may require increasing the number of iterations needed to achieve the optimal solution (Bi and
Xiao 2011).

45
 ACCEPTED MANUSCRIPT

Table 7- Continued

Algorithm Drawbacks of parameter control schemes

Deterministic evidence rule for updating the mutation factor
• The mutation factor,
is updated through deterministic rule based, though, it has been set to different random values through the

PT
• It introduces additional learning parameters such as &F and &= to steer both learning strategies, thus making the algorithm
SaDE; evolution process (see Eq. 83).
SaDE-MMTS
cumbersome with many parameters (see Eq. 82).

RI
This can be clearly observed in (Qin et al. 2009).
Local optima
The procedure of adjusting the control parameter values is mostly implemented in a random way. There is no accumulating knowledge

SC
EPSDE through the evolution and the parameter value that produce inferior solution will be reinitialized at the same generation. This procedure in
conjunction with the random selection of the mutation scheme will in some cases divert the population to a wrong direction of the search
space and fall in local optima if the algorithm fails to select the appropriate parameter values and scheme (Mallipeddi et al. 2011).

Despite the fact that the adaptation strategy of
and  is simple and effective (Ghosh et al. 2017), it is arguably known that the search

U
Deterministic evidence rule for updating the mutation factor and crossover rate
NRDE

AN
through different problems landscapes requires different parameter settings during the run and getting feedback from the search will
always be necessary (Karafotias et al. 2015).
Lack of exploitation

M
In (Mohamed 2017b), it can be noted that,
EADE • The scalar factor has been set to a fixed value during the run; this may affect the exploitation ability of the mutation scheme and

• Although the adaptation scheme of  is quite simple, it has disposes many parameters to the algorithm.
reduces the chances of finding good solutions around the local regions of the target individual.

D
TE
C EP
AC

46
 ACCEPTED MANUSCRIPT

Table 8: Summary of 9 important and recent adaptive DE variants

Combinations of
and  values are created as [
= 1.0,  =

Algorithm Ref Mutation Strategy Parameter Control Schemes

0.1] , [
= 1.0,  = 0.9] and [
= 0.8,  = 0.2]. At each

Ensemble of DE mutation strategies “DE/rand/1bin”, “DE/rand/2/bin” and
composite DE (Wang et al.
“DE/current-to-best/bin” is created. One is chosen randomly for each
(CoDE) 2011) generation, each individual is assigned to randomly selected
individual and the best will survive for the next population.

PT
combination and the best will survive to the next generation.

the values of
and  on the basis of Harmony search

Harmony search It is a harmony based parameter selection approach used to select
This algorithm is an improvement of the EPSDE algorithm. In this algorithm
based parameter
(Mallipeddi the Harmony Search algorithm (HS) proposed in (Geem et al. 2001) is

RI
ensemble technique. In which a Harmony memory is used to store the
2013) employed to select the DE strategies (Eq. 35 and Eq. 48) from an ensemble of
adaptation for DE successful strategies and parameter control values to be used in
these strategies instead of random manner as in EPSDE.
• The value of  is adapted during evolution using the linear
(HSPEADE) the next generations.

SC
Ensemble
sinusoidal • The JADE current/to-pbest (Eq. 35) mutation strategy is adopted in this population size reduction technique proposed in LSHADE-

• The value of
is adapted during the run using an ensemble

differential algorithm. EpSin algorithm (Awad et al. 2016).

 = ââã with Euclidean neighborhood between the best individual and

•

U
covariance matrix The crossover operator is modified using a covariance matrix learning
(Awad et al.
every other individual in the population; where, â and âã are two
adaptation with of two adaptation schemes (sinusoidal approaches): adaptive
2017b)

AN
orthogonal matrices and is a diagonal matrix which contains the Eigen
Euclidean sinusoidal increasing adjustment and non-adaptive sinusoidal

• The value of  is adapted in all generations using normal

neighborhood decreasing adjustment (Eq. 76-78).
(LSHADE- values.
cnEpSin)

M
distribution as in L-SHADE (Eq. 72-75).
The value of  is adapted during evolution using the linear
Differential
• The JADE current/to-pbest mutation strategy (Eq. 35) is adopted in this •
crossover strategy
algorithm. population size reduction technique proposed in L-SHADE

D
based on
 = â  âã with Euclidean neighborhood between the best individual and • The value of
is adapted using Cauchy distribution and
• The crossover operator is modified using a covariance matrix learning algorithm (Eq. 71).
covariance matrix (Awad et al.
every other individual in the population; where, â and âã are two  value is adapted using normal distribution along with

TE
learning with 2017c)
orthogonal matrices and  is a diagonal matrix which contains the Eigen
Euclidean
their corresponding formulas as suggested in L-SHADE
neighborhood
values. algorithm (Eq. 63, Eq. 72 and Eq. 73-75).
The  value is adapted during the run from the maximum
(L-covnSHADE)
EP
•
• New mutation strategy has been proposed: value which is 1 to the minimum value which is close to 0
 +
` ∙ (/ − ` ) +
` ∙ (/ − /0 ) ,= ? ∈ Ž
; = L `
depending on the population diversity measure which is
C

Individual-
` +
` ∙ (34664/ − ` ) +
` ∙ (/ − /0 ) ,= ? ∈ } ',.
dependent named as relative diversity:
$',. =
',.6
AC

differential (Bujok and

Where ? the base is individual’s index; Ž is the set of superior and inferior
Where div is the current population diversity and ',.6 is
evolution with Tvrdík
population Size 2017)
individuals in the population.
adaptation the diversity of the initial population calculated using the
• Another perturbation scheme is proposed for the last individual with small
Both
and  values are adapted using normal distribution
(IDEbd) diversity equation proposed in (Polakova et al. 2016).
probability to enhance the local search before implementing the binomial
with distribution means ?/ and ,/, respectively.
scheme. •

47
 ACCEPTED MANUSCRIPT

Table 8- Continued

• The value of  is adapted during evolution using the linear

Algorithm Ref. Mutation Strategy Parameter Control Schemes

population size reduction technique proposed in L-SHADE

• The values of
and  are adapted using two semi

algorithm (Eq. 71).

PT
Part1: during this part the concentrate will be on  which
LSHADE with parameter adaptation (SPA) parts:
Semi-parameter
is updated using L-SHADE (Eq. 72-75); whereas
is
The JADE current/to-pbest mutation strategy (Eq. 35) is adopted in this
adaptation (Mohamed
algorithm.

RI
hybrid with CMA- et al. 2017)
Part2: during this part the value of
is updated using L-
ES (LSHADE- updated using uniform distribution within specified limits.

SHADE (Eq. 63 and Eq. 73-75); whereas, the value of 

SPA)

SC
is updated using the same scheme used in part1.
The first part is activated during the first half of evolution. The

A set of DE strategies ŽÄ546 = {Eq. 35 and (Sarker et al. 2014)} is created The values of
and  are generated form a set of fixed values
second part is activated during the second half of evolution.

U
and a suitable DE operator is selected with regard to the better performing in between 0 and 1 as combinations. The better combinations of F

AN
DE with automatic
adaptation of each stage of the evolution process. This is implemented using rank strategy and CR values are survived on the basis of ranking method

ŽÄ‰,5¡b p?_w,5¡b
(Elsayed et for the DE operators involved in the search process: similar to that of the operators selection:
%&€‰ = %&€w =
operators and
al. 2017)
}`‰ }w
control parameters

M
Where }`‰ is the number of individuals generated using DE operator ŽÄ`‰ . Where }w is the number of individuals influenced by p?_w .
(DE-AOPS)

• All indicators have equal population size and they are very

D
the  formula will be
Multi-population based framework (MPF) for ensemble of multiple DE much smaller than the reward indicator subpopulation. Thus

 = å ∙ 
variants is considered:
? = ä ?
Ensemble
multiple DE
of (Wu et al.
2018)
TE !…9
Where ? , ? , ?0 are three indicator subpopulations from three
different variants JADEw, CoDE and EPSDE respectively, and ?9 as reward
where  is the ? size and  be the size of ? . å is
the proportion between ? and ?. As such we have
∑!…9 å = 1. In EDEV, å = å = å0 .
EP
variants (EDEV)

subpopulation. ? is the overall population. •
and  are adapted based on JADEw, CoDE and EPSDE
each within its representative indicator subpopulation.
C

A novel mutation rule has been proposed to maintain the balance between the

are independently generated according to uniform

global exploration and the local exploitation as well as to improve the
AC

distribution in (0.1,1).
•
. G = / +
∙ (‰_žŸ  − ‰_‡`/56 )
(Mohamed convergence rate of DE, as follows:

 are generated with accordance to the new adaptation

Adaptive guided and
where ‰_žŸ  and ‰_‡`/56 are randomly chosen from the top and bottom
DE (AGDE) Mohamed •
100% vectors, respectively in  , / is randomly chosen from the middle
2017) scheme explained in Eq. 101-104.
[ − 2(100%)] vectors, and  ∈ (0%, 50%).

48
 ACCEPTED MANUSCRIPT

PT
RI
U SC
AN
M
D
TE
EP
C
AC

Fig 4: An estimated rank of the adaptive DE algorithms

49
 ACCEPTED MANUSCRIPT

6 Metaheuristic framework with parameter control

After an interpretation of the literature and a comparative analysis of the aforementioned algorithms’ design,
we offer some guidelines on how to design and implement adaptive DE algorithms. The proposed designing
framework provides the reader with the main steps that are required for integrating any proposed meta-algorithm
with parameter and/or strategy adaptation schemes. These steps have been inspired by many adaptive algorithms
already presented in literature. This constitutes also the various components of these algorithms, including
selection, recombination, mutation, and survival operators.

PT
Framework 1: General framework for constructing an adaptive metaheuristic with respect to parameters control
and perturbation strategies

RI
Step 1: INDIVIDUAL (SOLUTION) ENCODING, POPULATION REPRESENTATION AND SELF-ADAPTIVE
PARAMETERS

SC
A population is a set (i.e. array) of individuals (chromosomes, particles) where , refers to the number of individuals in
each generation. First, we have to encode the necessary information required for the problem analysis in the individual
structure. Each individual should represent a complete solution to the problem at hand. All together, these parameters are
called solution parameters (Eiben and Smith 2003). Additionally, if our intention is to work on self-adaptive algorithm, the

U
individual should implicitly include also the encoding of the control parameter(s), or the so-called strategy’s parameter(s)
that we would to undergo evolution via recombination and mutation. This requires an intelligent decision to be taken into
AN
account in such a way that better parameter values would tend to produce better individuals (i.e. solutions) with the highest
chance to survive and propagate for more offsprings (Brest et al. 2007; Brest et al. 2010) .

Step 2: INDIVIDUAL EVALUATION (SOLUTION VALIDATION) AND SELECTION

M

The definition of the fitness function is crucially important for a successful application. In any meta-algorithm, we have to
evaluate the fitness of each individual. There are many standard fitness functions that can be used to test any proposed
meta-method (Cheng et al. 2017; Tanabe and Fukunaga 2013). Throughout many problems, it is more natural to state the
D

fitness function as minimization with/without constraints rather than maximization of some utility objectives; however this
depends on the type of the problem/application at hand. This step is always connected to selection operation and survival of
TE

the fittest, as these two operations work dependent on the fitness evaluation (Elsayed et al. 2017; Zamuda 2017).

Step 3: INDIVIDUAL GENERATOR STRATEGY

One of the most important steps when using metaheuristic algorithms is how the candidate solutions will be changed in
EP

order to diversify the population with new solutions. Exploration and exploitation are the two cornerstones of problem
solving by search. For a successful evolutionary process, one has to get proper balance between exploration (to cover
sufficiently the solution space seeking out for better solutions), and exploitation (refining the solutions by combining
information gathered from good ones during the exploration phase). Also, diversity maintenance is important to prevent
C

premature convergence. This gives a motivation to study and provide sufficient exploration and exploitation techniques to
be incorporated into the meta-algorithm main paradigm (Mohamed 2017b; Mohamed and Suganthan 2017). The main
AC

individual generation strategies that should get a thorough understanding and then well applied are: recombination and
mutation. Several individual strategies can be involved in the same algorithm. The selection of the right individual
generator can be implemented implicitly during the run. This requires additional approaches that collect information during
evolution in order to make decision which is the best strategy to be applied for an epoch (Ghosh et al. 2017; Wu et al.
2018).

Step 4: STOPPING CRITERIA

The most common stopping condition used in the literature is to allow the algorithm to run to a maximum number of
iterations (Eiben and Smith 2003). A small number of iterations may not offer enough time for the algorithm to attain an
optimum especially when the size of the search space is large. On the other hand, once the optimum solution is reached
there will be no more gain from a very large number of iterations. In general, the maximum number of iterations allowed to
50
 ACCEPTED MANUSCRIPT

multiply by the swarm size gives an indication of the number of particles evaluated by the algorithm (Brest and Maucec
2011).

Step 5: INVESTIGATE THE INFLUENCE OF DIFFERENT PARAMETERS OF METAHEURISTICS

• Population size . To draw the effect of  on the population diversity and performance of the proposed meta-
This includes:

algorithm, also to see whether increasing the size of the population may prevent, or at least, reduce the chance of
the algorithm to be trapped in local minima (Awad et al. 2016; Tanabe and Fukunaga 2014).

PT
• Selection and perturbation operators (recombination and mutation) with their qualitative and quantitative
parameter setting. A determination should be made in advance to the type of parameter control settings that would
be changed in deterministic, adaptive and/or self-adaptive manner, the change evidence, and for which
parameter(s) e.g. mutation factor, crossover rate, tournament size, and so on (Bujok and Tvrdík 2017; Wang et al.

RI
2011) .
• Determine the scope of change which is either to be on the gene level, individual level, or the whole population
level (Brest et al. 2007; Das et al. 2009b).

SC
Step 6: EXPERIMENTAL RESULTS
This step is performed to reach twofold goals. First, to show the reliability and efficiency of the proposed methods,
experimental results will be evaluated using various test bed problems (Cheng et al. 2017). Second, the relative

U
performance of the proposed methods, when compared with other well-known state-of-the-art algorithms should be
evaluated (both in qualitative and/or quantitative terms), and reported when applied to the same test problem/application. In
AN
all these experiments, the influences of different versions, components, and parameters of the meta-algorithm on
performance are to be investigated, analyzed, and reported (Neri and Tirronen 2010; Zamuda and Brest 2015).
M

7 Summary, conclusion and future insights

The classification of adaptive metaheuristics and adaptive differential evolution (DE) algorithms is always an
on-going research area. The research on developing adaptive meta-algorithms has been such a hot topic. As
D

more and more new adaptive algorithms are proposed with new characteristics, the need for a general
classification that can cover all these types of algorithms becomes a large demand. These classifications provide
TE

knowledge to those researches who are interested in this type of algorithms on what have been implemented and
what improvements or developments can be added in this area as future work. In this study, two classifications
have been proposed for the purpose: First, extension taxonomy to the EA parameter settings that covers in
EP

general the type of parameters settings in evolutionary computations (SECTION 3). Second, general
classification to the adaptive DE algorithms that classifies these algorithms based on the parameters control of
the algorithm as well as the number of DE strategies employed in the implementation (SECTION 5). Also, this
study has implemented a number of tables (Tables 2-8) after a thorough algorithmic design investigation applied
C

on 28 recent and important adaptive DE variants. These tables summarize everything related to the algorithmic
AC

design of these algorithms, and the points of weaknesses and strengths of each algorithm.
In summary, the main focus of our literature analysis is the algorithmic design and we attempt to offer a
protocol for future DE implementations (SECTION 6). In addition, many studies can be conducted to extend or
enhance the adaptive DE algorithms we have presented, some of these future directions are stated in Table 9.

51
 ACCEPTED MANUSCRIPT

Table 9: Future insights suggested for more DE algorithmic design investigation

Suggestion Details
It would be interesting to use some multi-comparison statistical test such as Friedman test, ANOVA and Wilcoxon Rank to analyze
Multi-comparison statistical test
the differences among the state-of-the-art adaptive DE variants presented in this survey.
Improve the performance of the JADE
The selection of the best individuals, % of the population size in the mutation strategy can be implemented in an adaptive manner
mutation strategy and its variants (Zhang and

PT
based on the population diversity.
Sanderson 2009b)
The neighborhood size in DEGL can be selected based on empirical guidelines for different optimization problems. Another
Empirical and theoretical improvements in
suggestion is to study the performance of DEGL using different neighborhood topologies. These topologies can be, e.g. wheel

RI
DEGL (Das et al. 2009b)
shaped or star-shaped topologies.
The MDE_pBX algorithm is a platform for many modifications. 1) An analytical investigation on the effects of the two new

be a future MDE_BX development to include new operators such as ª/t$_¿DFB/1, ª/t$_¿DFB/2, etc., and then their
strategies (mutation and crossover) on the population diversity and convergence rate. 2) The connotation of a dynamic grouping can

SC
Improve the performance of the MDE_pBX
algorithm in different directions (Islam et al. effectiveness could be measured on different types of test functions. 3) The parameter, , may also be modified to be adaptive or at

two additional control parameters – (the group size in the mutation operation) and  (the number of the top-ranking vectors in the
2012) the very least dynamic during the evolution process, hence its performance effectiveness can further be investigated. 4) There are

U
1) Propose an adaptive mechanism for
and test the new algorithm on different optimization problems. 2) Integrate different
crossover operation), a theoretical guidelines of how to select the values of  and – can be investigated.

AN
Many future insights for EADE (Mohamed adaptive schemes for  and compare the results with different adaptive DE algorithms. 3) Investigate the performance of EADE in
2017b) solving complex problems such as constrained and multi-objective optimization problems. 4) Combine the novel mutation strategy

M
with other EAs such as genetic algorithm and particle swarm optimization.
It would be combining the new algorithm with other DE mutation schemes (standard or advanced). Another future direction would
Interesting improvement for EFADE
be conducted by increasing the number of  sets and study their effects on different optimization problems.
be integrating the new triangulation scheme with other adaptive DE algorithms such as SHADE. Finally, an experimental study can

D
(Mohamed and Suganthan 2017)
SHADE and L-SHADE have the same future

TE
work development (Tanabe and Fukunaga In which an adequate adaptive technique can be found to dynamically adjust the memory size parameter for new types of problems.
2014)
The neighborhood steps in GADE can be
This proposed idea will have a great effect on enhancing the exploration and exploitation capabilities of the GADE algorithm.
EP
adjusted dynamically (Leon and Xiong 2016)
The integration of EsDEr-NR with other DE The significant adaptation strategies used in EsDEr-NR can be integrated with other DE variants and the results obtained can be
variants (Awad et al. 2017a) examined.

In these two algorithms, the parameter
can be set to an adaptive rule that accumulate knowledge from the previous generations.
C

Enhance the adaptive scheme of the

parameters control in the SaDE and its
AC

variant such as SaDE-MMTS (Qin et al. 2009)

Improve the adaptive ensemble of EPSDE The random strategy of the EPSDE in selecting the parameters control and mutation strategies can be improved by accumulating

Integrate the adaptation schemes of Ó and ÔÕ

(Mallipeddi et al. 2011) knowledge regarding the performance of the control parameter values through certain number of generations.
This may lead to more competitive adaptive DE variants if these adaptation schemes of FiADE are integrated with other non

is investigated on how the values of
and  are adapted in the new proposed algorithms.
in FiADE with other State-of-the-art DE adaptive DEs, e.g. opposition-based DE (Rahnamayan et al. 2006). This step may lead to further future insight if a theoretical study
algorithms (Ghosh et al. 2011)
NRDE has many open issues as future work The switching mechanism of the strategy and parameter control can be further investigated in other complex problems in addition to
insights (Ghosh et al. 2017) noisy functions. The dynamic switching of the DE parameters can also be analytically investigated.

52
 ACCEPTED MANUSCRIPT

References

Abbass HA The self-adaptive Pareto differential evolution algorithm. In: Proceedings of the IEEE Congress
on Evolutionary Computation (CEC2002), Honolulu, HI, May 12-17 2002. IEEE Press, pp 831- 836
Abderazek H, Ferhat D, Ivana A (2017) Adaptive mixed differential evolution algorithm for bi-objective tooth
profile spur gear optimization The International Journal of Advanced Manufacturing Technology
90:2063–2073
Al-Dabbagh MD, Al-Dabbagh RD, Raja Abdullah RSA, Hashim F (2015a) A new modified differential
evolution algorithm scheme-based linear frequency modulation radar signal de-noising Engineering

PT
Optimization 74:771-787 doi:10.1080/0305215x.2014.927449
Al-Dabbagh RD, Mekhilef S, Baba MS (2015b) Parameters' fine tuning of differential evolution algorithm
Computer Systems Science and Engineering 30:125-139

RI
Ali M, Siarry P, Pant M (2012) An efficient Differential Evolution based algorithm for solving multi-objective
optimization problems European Journal of Operational Research 217:404-416
Ali MM, Törn A (2004) Population set based global optimization algorithms: Some modifications and
numerical studies Comput Oper Res 31:1703-1725

SC
Angeline PJ (1995) Adaptive and self-adaptive evolutionary computations Computational Intelligence: A
Dynamic Systems Perspective:152-163
Arnold DV, Beyer H-G (2002) Noisy Optimization With Evolution Strategies. Genetic Algorithms and
Evolutionary Computation Springer Science & Business Media, Dortmund, Germany

U
Awad NH, Ali MZ, Suganthan PN (2017a) Ensemble of Parameters in a Sinusoidal Differential Evolution with
Niching-based population Reduction Swarm and Evolutionary Computation
AN
doi:https://doi.org/10.1016/j.swevo.2017.09.009
Awad NH, Ali MZ, Suganthan PN (2017b) Ensemble sinusoidal differential covariance matrix adaptation with
Euclidean neighborhood for solving CEC2017 benchmark problems. Paper presented at the 2017 IEEE
Congress on Evolutionary Computation (CEC), San Sebastian, Spain, 5-8 June
M

Awad NH, Ali MZ, Suganthan PN, Reynolds RG (2016) An ensemble sinusoidal parameter adaptation
incorporated with L-SHADE for solving CEC2014 benchmark problems. Paper presented at the IEEE
Congress on Evolutionary Computation (CEC 2016), Vancouver, BC, Canada, 24-29 July
D

Awad NH, Ali MZ, Suganthan PN, Reynolds RG, Shatnawi AM (2017c) A novel differential crossover strategy
based on covariance matrix learning with Euclidean neighborhood for solving real-world problems.
TE

Paper presented at the 2017 IEEE Congress on Evolutionary Computation (CEC), San Sebastian, Spain,
5-8 June
Baluja S (1994) Population-based incremental learning: A method for integrating genetic search based function
optimization and competitive learning. Carnegie Mellon University, Pittsburgh, PA
EP

Baluja S, Caruana R (1995) Removing the genetics from the standard genetic algorithm Carnegie Mellon
University, Pittsburgh, PA
Bi X-J, Xiao J (2011) Classification-based self-adaptive differential evolution with fast and reliable convergence
performance Soft Computing 15:1581-1599 doi:10.1007/s00500-010-0689-5
C

Bor E, Turduev M, Kurt H (2016) Differential evolution algorithm based photonic structure design: numerical
and experimental verification of subwavelength λ/5 focusing of light Scientific Reports 6
AC

Boussaïd I, Lepagnot J, Siarry P (2013) A survey on optimization metaheuristics Information Sciences 237:82-
117
Brest J, Boskovic B, Greiner S, Zumer V, Maucec MS (2007) Performance comparison of self-adaptive and
adaptive differential evolution algorithms Soft Computing 11:617-629 doi:10.1007/s00500-006-0124-0
Brest J, Boskovic B, Zumer V (2010) An Improved Self-adaptive Differential Evolution Algorithm in Single
Objective Constrained Real-Parameter Optimization. Paper presented at the 2010 IEEE Congress on
Evolutionary Computation (CEC), Barcelona, Spain, 18-23 July
Brest J, Greiner S, Boskovic B, Mernik M, Zumer V (2006) Self-adapting control parameters in differential
evolution: A comparative study on numerical benchmark problems Ieee Transactions on Evolutionary
Computation 10:646-657 doi:10.1109/tevc.2006.872133
Brest J, Maucec MS (2011) Self-adaptive differential evolution algorithm using population size reduction and
three strategies Soft Computing 15:2157-2174 doi:10.1007/s00500-010-0644-5
53
 ACCEPTED MANUSCRIPT

Brest J, Zamuda A, Boskovic B, Maucec MS, Zumer V (2009) Dynamic optimization using Self-Adaptive
Differential Evolution. Paper presented at the IEEE Congress on Evolutionary Computation (CEC
2009), Trondheim, Norway, 18-21 May
Bujok P, Tvrdík J (2017) Enhanced individual-dependent differential evolution with population size adaptation.
Paper presented at the 2017 IEEE Congress on Evolutionary Computation (CEC), San Sebastian, Spain,
5-8 June
Bujok P, Tvrdík J, Poláková R (2014) Differential evolution with rotation-invariant mutation and competing-
strategies adaptation Paper presented at the 2014 IEEE Congress on Evolutionary Computation (CEC),
Beijing, China, 6-11 July

PT
Caponio A, Neri F (2009) Differential Evolution with Noise Analyzer. Paper presented at the Workshops on
Applications of Evolutionary Computation (EvoWorkshops 2009),
Caraffini F, Neri F, Poikolainen I, Ieee (2013) Micro-Differential Evolution with Extra Moves Along the Axes

RI
Proceedings of the 2013 Ieee Symposium on Differential Evolution (Sde)
Cheng R, Li M, Tian Y, Zhang X, Yang S, Jin Y, Yao X (2017) Benchmark Functions for CEC’2017
Competition on Evolutionary Many-Objective Optimization. School of Computer Science, University of
Birmingham, Birmingham, U.K.

SC
Chiang T-C, Chen C-N, Lin Y-C (2013) Parameter Control Mechanisms in Differential Evolution: A Tutorial
Review and Taxonomy Proceedings of the 2013 Ieee Symposium on Differential Evolution (Sde)
Cotta C, Sevaux M, Sörensen KE (2008) Adaptive and Multilevel Metaheuristics vol 136. Studies in
Computational Intelligence. Springer-Verlag, Berlin,Germany

U
Das S, Abraham A, Chakraborty UK, Konar A (2009a) Differential Evolution Using a Neighborhood-Based
Mutation Operator Ieee Transactions on Evolutionary Computation 13:526-553
AN
doi:10.1109/tevc.2008.2009457
Das S, Abraham A, Chakraborty UK, Konar A (2009b) Differential Evolution Using a Neighborhood-based
Mutation Operator IEEE Transactions on Evolutionary Computation 13:526 -553
Das S, Ghosh A, Mullick SS (2015) A Switched Parameter Differential Evolution for Large Scale Global
M

Optimization – Simpler May Be Better. Paper presented at the Mendel 2015: 21st International
Conference on Soft Computing, Advances in Intelligent Systems and Computing, Brno, Czech
Republic, 23 – 25 June
D

Das S, Mandal A, Mukherjee R (2014a) An Adaptive Differential Evolution Algorithm for Global Optimization
in Dynamic Environments IEEE Transactions on Cybernetics 44:966-978
TE

doi:10.1109/tcyb.2013.2278188
Das S, Mandal A, Mukherjee R (2014b) An Adaptive Differential Evolution Algorithm for Global Optimization
in Dynamic Environments IEEE Transactions on Cybernetics 44:966 - 978
Das S, Mullick SS, Suganthan PN (2016) Recent advances in differential evolution - An updated survey Swarm
EP

and Evolutionary Computation 27:1-30

Das S, Suganthan PN (2011) Differential Evolution: A Survey of the State-of-the-Art Ieee Transactions on
Evolutionary Computation 15:27-54 doi:10.1109/tevc.2010.2059031
Dominguez-Isidro S, Mezura-Montes E (2017) The Baldwin Effect on a Memetic Differential Evolution for
C

Constrained Numerical Optimization Problems. Paper presented at the Proceedings of the Genetic and
Evolutionary Computation (GECOO 2017), Berlin, Germany 15 -19 July
AC

Eiben AE, Hinterding R, Michalewicz Z (1999) Parameter control in evolutionary algorithms Ieee Transactions
on Evolutionary Computation 3:124-141 doi:10.1109/4235.771166
Eiben AE, Smith JE (2003) Introduction to Evolutionary Computing. Natural Computing Series, Second edn.
Springer-Verlag, Berlin,Germany
Elsayed S, Sarker R, Coello CC, Ray T (2017) Adaptation of operators and continuous control parameters in
differential evolution for constrained optimization Soft Computing doi:https://doi.org/10.1007/s00500-
017-2712-6
Elsayed SM, Sarker RA, Essam DL (2014) A self-adaptive combined strategies algorithm for constrained
optimization using differential evolution Applied Mathematics and Computation 241:267-282
doi:10.1016/j.amc.2014.05.018
Fan HY, Lampinen J (2003) A trigonometric mutation operation to differential evolution Journal of Global
Optimization 27:105-129 doi:10.1023/a:1024653025686
54
 ACCEPTED MANUSCRIPT

Fan Q, Yan X (2016) Self-Adaptive Differential Evolution Algorithm With Zoning Evolution of Control
Parameters and Adaptive Mutation Strategies IEEE transactions on Cybernetics 46:219 - 232
doi:10.1109/TCYB.2015.2399478
Feng L, Zhou W, Zhou L, Zhong JH, Da BS, Zhu ZX, Wang Y (2017) An empirical study of multifactorial PSO
and multifactorial DE. Paper presented at the 2017 IEEE Congress on Evolutionary Computation
(CEC), San Sebastian, Spain, 5-8 June
Feoktistov V (2006) Differential Evolution: In Search of Solutions vol 5. Springer Optimization and Its
Applications. Springer-Verlag, New York, United State
Feoktistov V, Janaqi S Generalization of the strategies in differential evolution. In: 18-th Annual IEEE

PT
International Parallel and Distributed Processing Symposium, Santa Fe, New Mexico, USA, April 26-30
2004a. IEEE Computer Society, pp 2341-2346
Feoktistov V, Janaqi S (2004b) New strategies in differential evolution - Design principle. In: I.C. Parmee

RI
editor, Adaptive computing in design and manufacture VI. Springer - Verlag London Limited, UK,
London, pp 335-346. doi:10.1007/978-0-85729-338-1_28
Fister I, Yang X-S, Fister I, Brest J, Fister D (2013) A Brief Review of Nature-Inspired Algorithms for
Optimization Elektrotehniski Vestnik/Electrotechnical Review 80:1-7 doi:arXiv:1307.4186

SC
Fogel DB, Fogel LJ, Atmar JW Meta-evolutionary programming. In: Conference on Signals, Systems and
Computers. 1991 Conference Record of the Twenty-Fifth Asilomar, San Diego, CA, USA, Nov 4-6
1991. pp 540-545 doi:10.1109/acssc.1991.186507
Geem ZW, Kim JH, Loganathan GV (2001) A New Heuristic Optimization Algorithm: Harmony Search

U
Simulation 76:60-68
Gendreau M, Potvin J-Y (eds) (2010) Handbook of Metaheuristics vol 146. International Series in Operations
AN
Research & Management Science, 2nd edn. Springer, London
Ghosh A, Das S, Chowdhury A, Giri R (2011) An improved differential evolution algorithm with fitness-based
adaptation of the control parameters Information Sciences 181:3749-3765
Ghosh A, Das S, Panigrahi BK (2017) A Noise Resilient Differential Evolution with Improved Parameter and
M

Strategy Control. Paper presented at the 2017 IEEE Congress on Evolutionary Computation (CEC), San
Sebastian, Spain, 5-8 June
Gong W, Cai Z, Ling CX, Li H (2011) Enhanced Differential Evolution With Adaptive Strategies for Numerical
D

Optimization IEEE Transactions on Systems Man and Cybernetics Part B-Cybernetics 41:397-413
doi:10.1109/tsmcb.2010.2056367
TE

Goudos SK, Deruyck M, Plets D, Martens L, Joseph W (2017) Optimization of Power Consumption in 4G LTE
Networks Using a Novel Barebones Self-adaptive Differential Evolution Algorithm Telecommunication
Systems 66:109–120 doi:doi:10.1007/s11235-017-0279-2
Halder U, Das S, Maity D (2013) A Cluster-Based Differential Evolution Algorithm With External Archive for
EP

Optimization in Dynamic Environments IEEE Transactions on Cybernetics 43:881- 897

Hansen N, Ostermeier A Adapting Arbitrary Normal Mutation Distributions in Evolution Strategies:
The Covariance Matrix Adaptation In: Proceedings of the 1996 IEEE International Conference on
Evolutionary Computation, Nagoya, Japan, 20-22 May 1996. pp 312-317
C

Hui S, Suganthan PN (2016) Ensemble and Arithmetic Recombination-Based Speciation Differential Evolution
for Multimodal Optimization IEEE TRANSACTIONS ON CYBERNETICS 46:64-74
AC

Islam SM, Das S, Ghosh S, Roy S, Suganthan PN (2012) An Adaptive Differential Evolution Algorithm With
Novel Mutation and Crossover Strategies for Global Numerical Optimization Ieee Transactions on
Systems Man and Cybernetics Part B-Cybernetics 42:482-500 doi:10.1109/tsmcb.2011.2167966
Jin Y, Branke J (2005) Evolutionary Optimization in Uncertain Environments-A Survey IEEE Transactions on
Evolutionary Computation 9:303-317
Junior HAeO, Ingber L, Petraglia A, Petraglia MR, Machado MAS (2012a) Global Optimization and Its
Applications. In: Stochastic Global Optimization and Its Applications with Fuzzy Adaptive Simulated
Annealing vol 35. Intelligent Systems Reference Library Springer-Verlag Berlin Heidelberg, pp 11-20
Junior HAeO, Ingber L, Petraglia A, Petraglia MR, Machado MAS (2012b) Metaheuristic Methods. In:
Stochastic Global Optimization and Its Applications with Fuzzy Adaptive Simulated Annealing, vol 35.
Intelligent Systems Reference Library. Springer-Verlag Berlin Heidelberg, pp 21-30

55
 ACCEPTED MANUSCRIPT

Karafotias G, Hoogendoorn M, Eiben AE (2015) Parameter Control in Evolutionary Algorithms: Trends and
Challenges IEEE Transactions on Evolutionary Computation 19:167 - 187
Kennedy J, Eberhart R Particle Swarm Optimization. In: 1995 Proceedings of IEEE International Conference on
Neural Networks, Perth, WA, Nov 27 - Dec 01 1995. IEEE Press, pp 1942-1948
Kukkonen S, Mezura-Montes E (2017) An experimental comparison of two constraint handling approaches used
with differential evolution. Paper presented at the 2017 IEEE Congress on Evolutionary Computation
(CEC), San Sebastian, Spain, 5-8 June
Lampinen J Solving problems subject to multiple nonlinear constraints by the differential evolution. In:
Proceedings of MENDEL'01 - 7th International Conference on Soft Computing, Brno, Czech Republic,

PT
June 2001. pp 50-57
Lampinen J A constraint handling approach for the differential evolution algorithm. In: Proceedings of the 2002
Congress on Evolutionary Computation - CEC2002, Honolulu, HI, May 12-17 2002. IEEE Press, pp

RI
1468-1473
Lampinen J, Zelinka I On stagnation of the differential evolution algorithm. In: Proceedings of MENDEL'00 -
6th International Mendel Conference on Soft Computing, Brno, Czech Republic, 2000. pp 76-83
Laredo JLJ, Fernandes C, Merelo JJ, Gagné C (2009) Improving genetic algorithms performance via

SC
deterministic population shrinkage. Paper presented at the 11th Annual conference on Genetic and
evolutionary computation (GECCO '09), Montreal, Québec, Canada, 08 - 12 July
Leon M, Xiong N (2016) Adapting Differential Evolution Algorithms For Continuous Optimization Via Greedy
Adjustment Of Control Parameters Journal of Artificial Intelligence and Soft Computing Research

U
6:103-118
Li X, Ma S, Hu J (2017) Multi-search differential evolution algorithm Applied Intelligence 47:231-256
AN
Lin C, Qing A, Feng Q (2011) A comparative study of crossover in differential evolution Journal of Heuristics
17:675-703 doi:10.1007/s10732-010-9151-1
Lis J Parallel genetic algorithm with dynamic control parameter. In: Proceedings of the 1996 IEEE Conference
on Evolutionary Computation, Nagoya, Japan, 1996. IEEE Press, pp 324-329
M

Liu J, Lampinen J (2005) A fuzzy adaptive differential evolution algorithm Soft Computing 9:448-462
doi:10.1007/s00500-004-0363-x
Lobato FS, Jr. VS (2017) Self-adaptive Multi-objective Optimization Differential Evolution. In: Multi-
D

Objective Optimization Problems. SpringerBriefs in Mathematics. Springer, Cham, pp 47-73.

doi:https://doi.org/10.1007/978-3-319-58565-9_4
TE

Lobo FG, Lima CF, Michalewicz Z (2007) Parameter Setting in Evolutionary Algorithms vol 54. Studies in
Computational Intelligence. Springer-Verlag, Berlin, Germany
Lynn N, Mallipeddi R, Suganthan PN (2014) Differential Evolution with Two Subpopulations. Paper presented
at the International Conference on Swarm, Evolutionary, and Memetic Computing (SEMCCO 2014),
EP

Bhubaneswar, India, 18-20 Dec

Mallipeddi R (2013) Harmony Search Based Parameter Ensemble Adaptation for Differential Evolution Journal
of Applied Mathematics doi:10.1155/2013/750819
Mallipeddi R, Suganthan PN, Pan QK, Tasgetiren MF (2011) Differential evolution algorithm with ensemble of
C

parameters and mutation strategies Applied Soft Computing 11:1679-1696

doi:10.1016/j.asoc.2010.04.024
AC

Mezura-Montes E, Miranda-Varela ME, Gómez-Ramón RdC (2010) Differential evolution in constrained

numerical optimization: An empirical study Information Sciences 180:4223-4262
Mezura-Montes E, Reyes-Sierra M, Coello CAC (2008) Multi-objective Optimization Using Differential
Evolution: A Survey of the State-of-the-Art. In: Advances in Differential Evolution, vol 143. Studies in
Computational Intelligence. Springer-Verlag Berlin Heidelberg, pp 173-196
Mininno E, Neri F, Cupertino F, Naso D (2011) Compact Differential Evolution Ieee Transactions on
Evolutionary Computation 15:32-54 doi:10.1109/tevc.2010.2058120
Mohamed AW (2015) An improved differential evolution algorithm with triangular mutation for global
numerical optimization Computers & Industrial Engineering 85:359-375
doi:https://doi.org/10.1016/j.cie.2015.04.012

56
 ACCEPTED MANUSCRIPT

Mohamed AW (2017a) A novel differential evolution algorithm for solving constrained engineering
optimization problems Journal of Intelligent Manufacturing:1-34 doi:https://doi.org/10.1007/s10845-
017-1294-6
Mohamed AW (2017b) Solving large-scale global optimization problems using enhanced adaptive differential
evolution algorithm Complex & Intelligent Systems 3:205–231 doi:https://doi.org/10.1007/s40747-017-
0041-0
Mohamed AW, Hadi AA, Fattouh AM, Jambi KM (2017) LSHADE with semi-parameter adaptation hybrid
with CMA-ES for solving CEC 2017 benchmark problems. Paper presented at the 2017 IEEE Congress
on Evolutionary Computation (CEC), San Sebastian, Spain, 5-8 June

PT
Mohamed AW, Mohamed AK (2017) Adaptive guided differential evolution algorithm with novel mutation for
numerical optimization International Journal of Machine Learning and Cybernetics:1-25
doi:https://doi.org/10.1007/s13042-017-0711-7

RI
Mohamed AW, Sabry HZ (2012) Constrained optimization based on modified differential evolution algorithm
Information Sciences 194:171-208
Mohamed AW, Suganthan PN (2017) Real-parameter unconstrained optimization based on enhanced fitness-
adaptive differential evolution algorithm with novel mutation Soft Computing:1-21

SC
doi:https://doi.org/10.1007/s00500-017-2777-2
Monakhov OG, Monakhova EA, Pant M (2016) Application of differential evolution algorithm for optimization
of strategies based on financial time series Numerical Analysis and Applications 9:150–158
Montes EM, Coello Coello CA, Tun-Morales EI Simple feasibility rules and differential evolution for

U
constrained optimization. In: Proceedings of the Third Mexican International Conference on Artificial
Intelligence (MICAI'2004), New York, April 2004. Springer-Verlag, pp 707-716
AN
Neri F, Mininno E (2010) Memetic Compact Differential Evolution for Cartesian Robot Control Ieee
Computational Intelligence Magazine 5:54-65 doi:10.1109/mci.2010.936305
Neri F, Tirronen V (2010) Recent advances in differential evolution: a survey and experimental analysis
Artificial Intelligence Review 33:61-106 doi:10.1007/s10462-009-9137-2
M

Neri F, Tirronen V, Ieee (2008) On Memetic Differential Evolution Frameworks: A Study of Advantages and
Limitations in Hybridization 2008 Ieee Congress on Evolutionary Computation, Vols 1-8:2135-2142
doi:10.1109/cec.2008.4631082
D

Novoa-Hernández P, Corona CC, Pelta DA (2013) Self-adaptive, multipopulation differential evolution in

dynamic environments Soft Computing- A Fusion of Foundations, Methodologies and Applications
TE

17:1861-1881
Panigrahi BK, Shi Y, Lim M-H (eds) (2011) Handbook of Swarm Intelligence: Concepts, Principles and
Applications vol 8. Adaptation, Learning, and Optimization. Springer, Chennai, India
Peng F, Tang K, Chen G, Yao X Multi-start JADE with knowledge transfer for numerical optimization. In: 2009
EP

Ieee Congress on Evolutionary Computation (CEC '09), Trondheim, NORWAY, MAY 18-21 2009. pp
1889-1895
Poláková R (2017) L-SHADE with competing strategies applied to constrained optimization. Paper presented at
the 2017 IEEE Congress on Evolutionary Computation (CEC), San Sebastian, Spain, 5-8 June
C

Polakova R, Tvrdık J, Bujok P (2016) Population-size adaptation through diversity-control mechanism for
differential evolution. Paper presented at the MENDEL, 22th International Conference on Soft
AC

Computing, Brno, Czech Republic,

Price K, Storn R (1997) Differential evolution: A simple evolution strategy for fast optimization Dr Dobb's
Journal of Software Tools 22:18-24
Price KV Differential evolution vs. the functions of the 2nd ICEO. In: Proceedings of 1997 IEEE International
Conference on Evolutionary Computation (ICEC '97), Indianapolis, In, April 1997. IEEE; IEEE Neural
Network Council (NNC), pp 153-157
Price KV (1999) An Introduction to Differential Evolution. In: New Ideas in Optimization. McGraw-Hill,
London,
Price KV, Storn RM, Lampinen JA (2005) Differential Evolution: A Practical Approach to Global
Optimization. Natural Computing Series, First edn. Springer-Verlag, Berlin, Germany

57
 ACCEPTED MANUSCRIPT

Qin AK, Huang VL, Suganthan PN (2009) Differential Evolution Algorithm With Strategy Adaptation for
Global Numerical Optimization Ieee Transactions on Evolutionary Computation 13:398-417
doi:10.1109/tevc.2008.927706
Qin AK, Suganthan PN Self-adaptive differential evolution algorithm for numerical optimization. In:
Proceedings of the 2005 Ieee Congress on Evolutionary Computation, Edinburgh, SCOTLAND, SEP
02-05 2005. IEEE; IEEE Computat Intelligence Soc; IEE; Evolut Programming Soc, pp 1785-1791
Rahnamayan S, Tizhoosh HR, Salama MMA Opposition-based differential evolution algorithms. In: 2006 IEEE
Congress on Evolutionary Computation, Vancouver, CANADA, 16-21 July 2006. IEEE Press, pp 1995-
2002

PT
Rakshit P, Chowdhury A, Konar A (2017) Differential evolution induced many objective optimization. Paper
presented at the 2017 IEEE Congress on Evolutionary Computation (CEC), San Sebastian, Spain, 5-8
June

RI
Rechenberg I (1973) Evolutionsstrategie: Optimierung technischer Systeme nach Prinzipien der biologischen
Evolution. Frommann-Holzboog,
Robič T, Filipič B (2005) DEMO: Differential Evolution for Multiobjective Optimization. Paper presented at the
International Conference on Evolutionary Multi-Criterion Optimization (EMO 2005),

SC
Rönkkönen J, Kukkonen S, Price VK Real-parameter optimization with differential evolution. In: The 2005
IEEE Congress on evolutionary computation CEC 2005, Edinburgh, Scotland, Sep 5 2005. IEEE press,
pp 506-513
Salvatore N, Caponio A, Neri F, Stasi S, Cascella GL (2010) Optimization of Delayed-State Kalman-Filter-

U
Based Algorithm via Differential Evolution for Sensorless Control of Induction Motors IEEE
Transactions on Industrial Electronics 57:385 - 394
AN
Sarker RA, Elsayed SM, Ray T (2014) Differential Evolution With Dynamic Parameters Selection for
Optimization Problems IEEE Transactions on Evolutionary Computation 18:689 - 707
Schwefel H-P (1977) Numerische Optimierung von Computor-Modellen mittels der Evolutionsstrategie: mit
einer vergleichenden Einführung in die Hill-Climbing-und Zufallsstrategie. Birkhäuser,
M

Sindhya K, Ruuska S, Haanpaa T, Miettinen K (2011) A new hybrid mutation operator for multiobjective
optimization with differential evolution Soft Computing 15:2041-2055 doi:10.1007/s00500-011-0704-5
Smith RE, Smuda E (1995) Adaptively resizing populations: Algorithm, analysis and first results Complex
D

Systems 9:47-72
Storn R On the usage of differential evolution for function optimization. In: Biennial Conference of the North
TE

America Fuzzy Information Processing Society (NAFIPS 1996), Berkeley, CA, New York, 1996. IEEE,
pp 519-523
Storn R (2017) Real-world applications in the communications industry - when do we resort to Differential
Evolution? Paper presented at the 2017 IEEE Congress on Evolutionary Computation (CEC), San
EP

Sebastian, Spain, 5-8 June

Storn R, Price K (1995a) Differential evolution: a simple and efficient adaptive scheme for global optimization
over continuous spaces. International Computer Science Institute (ICSI), USA
Storn R, Price K Minimizing the real functions of the ICEC'96 contest by differential evolution. In: IEEE
C

International Conference on Evolutionary Computation, May 1996. Nagoya, IEEE, New York, pp 842-
844
AC

Storn R, Price K (1997) Differential evolution–a simple and efficient heuristic for global optimization over
continuous spaces Journal of global optimization 11:341-359
Storn RM, Price KV (1995b) Differential Evolution - A Simple and Efficient Adaptive scheme for Global
Optimization over Continuous Spaces. International Computer Science Institute, Berkeley, CA, USA
Suganthan PN (2016) Future of Real Parameter Optimization Paper presented at the International Conference on
Soft Computing for Problem Solving (SocProS 2016) Thapar University, Topiala, India, 23-24 Dec
Suganthan PN, Das S, Mukherjee S, Chatterjee S (2014) Adaptation Methods in Differential Evolution: A
Review. Paper presented at the 20th International Conference on Soft Computing MENDEL 2014,
Brno, Czech Republic, 25 - 27 June
Tanabe R, Fukunaga A (2013) Evaluating the performance of SHADE on CEC 2013 benchmark problems.
Paper presented at the IEEE Congress on Evolutionary Computation (CEC 2013), Cancun, Mexico, 20-
23 June
58
 ACCEPTED MANUSCRIPT

Tanabe R, Fukunaga AS (2014) Improving the search performance of SHADE using linear population
reduction. Paper presented at the 2014 IEEE Congress on Evolutionary Computation (CEC), Beijing,
China, 6-11 July
Teo J (2006) Exploring dynamic self-adaptive populations in differential evolution Soft Computing 10:673-686
doi:10.1007/s00500-005-0537-1
Tseng L-Y, Chen C Multiple trajectory search for multiobjective optimization. In: 2007 Ieee Congress on
Evolutionary Computation, Singapore, SINGAPORE, SEP 25-28 2007. pp 3609-3616
Tseng L-Y, Chen C Multiple Trajectory Search for Large Scale Global Optimization. In: 2008 Ieee Congress on
Evolutionary Computation, Hong Kong, PEOPLES R CHINA, JUN 01-06 2008. pp 3052-3059

PT
Tvrdik J (2009) Adaptation in differential evolution: A numerical comparison Applied Soft Computing 9:1149-
1155 doi:10.1016/j.asoc.2009.02.010
Uher V, Gajdoš P, Radecký M, Snášel V (2016) Utilization of the Discrete Differential Evolution for

RI
Optimization in Multidimensional Point Clouds Computational Intelligence and Neuroscience 2016:1-
14
Urfalioglu O, Arikan O (2011) Self-adaptive randomized and rank-based differential evolution for multimodal
problems Journal of Global Optimization 51:607-640 doi:10.1007/s10898-011-9646-9

SC
Viktorin A, Senkerik R, Pluhacek M, Kadavy T (2017) Archive Analysis in SHADE. Paper presented at the
International Conference on Artificial Intelligence and Soft Computing, Zakopane, Poland,
Wang H, Rahnamayan S, Wu Z (2013) Parallel differential evolution with self-adapting control parameters and
generalized opposition-based learning for solving high-dimensional optimization problems Journal of

U
Parallel and Distributed Computing 73:62-73 doi:10.1016/j.jpdc.2012.02.019
Wang X, Dong Z, Tang L (2016) A multi-objective differential evolution algorithm with memory based
AN
population construction. Paper presented at the 2016 IEEE Congress on Evolutionary Computation
(CEC), Vancouver, BC, Canada, 24-29 July
Wang Y, Cai Z, Zhang Q (2011) Differential Evolution with Composite Trial Vector Generation Strategies and
Control Parameters IEEE Transactions on Evolutionary Computation 15:55-66
M

doi:10.1109/tevc.2010.2087271
Wang Y, Cai Z, Zhang Q (2012) Enhancing the search ability of differential evolution through orthogonal
crossover Information Sciences 185:153-177 doi:10.1016/j.ins.2011.09.001
D

Weber M, Neri F, Tirronen V (2011) Shuffle or update parallel differential evolution for large-scale
optimization Soft Computing 15:2089-2107 doi:10.1007/s00500-010-0640-9
TE

Weber M, Tirronen V, Neri F (2010) Scale factor inheritance mechanism in distributed differential evolution
Soft Computing 14:1187-1207 doi:10.1007/s00500-009-0510-5
Wu G, Shen X, Li H, Chen H, Lin A, Suganthan PN (2018) Ensemble of differential evolution variants
Information Sciences 423:172-186
EP

Yang X-S, Cui Z, Xiao R, Gandom AH, Karamanoglu M (eds) (2013) Swarm Intelligence and Bio-inspired
Computation: Theory and Applications. Elsevier,
Yu X, Chen W-N, Hu X-M, Zhang J (2017) Fast 3D Path Planning based on Heuristic-aided Differential
Evolution. Paper presented at the Proceedings of The Genetic and Evolutionary Computation
C

Conference (GECCO 2017), Berlin, July 15 - 19

Zaharie D Parameter adaptation in differential evolution by controlling the population diversity. In: Proc. of
AC

SYNASC'2002, Analele Univ. Timisoara,Timisoara, Roumania, 2002. seria Matematica-Informatica,

vol special issue. pp 281-295
Zaharie D A multi-population differential evolution algorithm for multi-modal optimization. In: Mendel'04 -
10th International Conference on Soft Computing, Brno, Czech Republic, June 2004. pp 17-22
Zaharie D A Comparative Analysis of Crossover Variants in Differential Evolution. In: Proceedings of the
IMCSIT, 2nd International Symposium Advances in Artificial Intelligence and Applications (AAIA'07),
2007. pp 171-181
Zamuda A (2017) Adaptive constraint handling and Success History Differential Evolution for CEC 2017
Constrained Real-Parameter Optimization. Paper presented at the 2017 IEEE Congress on Evolutionary
Computation (CEC), San Sebastian, Spain, 5-8 June
Zamuda A, Brest J (2015) Self-adaptive control parameters‫ ׳‬randomization frequency and propagations in
differential evolution Swarm and Evolutionary Computation 25:72–99
59
 ACCEPTED MANUSCRIPT

Zhang G, Pan L, Neri F, Gong M, Leporati A (2017) Metaheuristic Optimization: Algorithmic Design and
Applications Journal of Optimization 2017:1-2 doi:http:\\doi:10.1155/2017/1053145
Zhang J, Sanderson AC (2009a) Adaptive Differential Evolution: A Robust Approach to Multimodal Problem
Optimization vol 1. Adaptive, Learning, and Optimization. Springer-Verlag, Chennai, India
Zhang J, Sanderson AC (2009b) JADE: Adaptive Differential Evolution With Optional External Archive Ieee
Transactions on Evolutionary Computation 13:945-958 doi:10.1109/tevc.2009.2014613
Zhang X, Chen W, Dai C, Cai W (2010) Dynamic multi-group self-adaptive differential evolution algorithm for
reactive power optimization International Journal of Electrical Power & Energy Systems 32:351-357
doi:10.1016/j.ijepes.2009.11.009

PT
Zhao S-Z, Suganthan PN, Das S (2011) Self-adaptive differential evolution with multi-trajectory search for
large-scale optimization Soft Computing 15:2175-2185 doi:10.1007/s00500-010-0645-4
Zhao S, Wang X, Chen L, Zhu W (2014) A Novel Self-adaptive Differential Evolution Algorithm with

RI
Population Size Adjustment Scheme Arabian Journal for Science and Engineering 39:6149-6174
doi:10.1007/s13369-014-1248-7
Zhu W, Tang Y, Fang J-a, Zhang W (2013) Adaptive population tuning scheme for differential evolution
Information Sciences 223:164-191 doi:10.1016/j.ins.2012.09.019

SC
Zou D, Liu H, Gao L, Li S (2011) A novel modified differential evolution algorithm for constrained
optimization problems Computers & Mathematics with Applications 61:1608-1623
doi:10.1016/j.camwa.2011.01.029

U
AN
M
D
TE
C EP
AC

60