Neural Computing and Applications

https://doi.org/10.1007/s00521-019-04580-4 (0123456789().,-volV)(0123456789().
,- volV)

ORIGINAL ARTICLE

Hybridizing grey wolf optimization with neural network algorithm

for global numerical optimization problems
Yiying Zhang1 • Zhigang Jin1 • Ye Chen1,2

Received: 19 January 2019 / Accepted: 19 October 2019

Ó Springer-Verlag London Ltd., part of Springer Nature 2019

Abstract
This paper proposes a novel hybrid algorithm, called grey wolf optimization with neural network algorithm (GNNA), for
solving global numerical optimization problems. The core idea of GNNA is to make full use of good global search ability
of neural network algorithm (NNA) and fast convergence of grey wolf optimizer (GWO). Moreover, both NNA and GWO
are improved to boost their own advantages. For NNA, an improved NNA is given to strengthen the exploration ability of
NNA by discarding transfer operator and introducing random modification factor. For GWO, an enhanced GWO is
presented, which adjusts the exploration rate based on reinforcement learning principles. Then the improved NNA and the
enhanced GWO are hybridized by dynamic population mechanism. A comprehensive set of 23 well-known unconstrained
benchmark functions are employed to examine the performance of GNNA compared with 13 metaheuristic algorithms.
Such comparisons suggest that the combination of the improved NNA and the enhanced GWO is very effective and GNNA
is clearly seen to be more successful in both solution quality and computational efficiency.

Keywords Artificial neural networks  Reinforcement learning  Grey wolf optimizer  Numerical optimization

1 Introduction mathematical principles at each iteration and may need

other information like gradient, initial points and hessian
In the real world, optimization problems can be found in matrix [1, 2]. Although deterministic methods can be
almost all engineering fields. Solving optimization problem viewed as available options for some simple and ideal
is to find an optimal solution from all possible solutions of optimization problems, they are not effective in solving
the given constrained space to maximize or minimize its complex optimization problems such as large-scale, mul-
objective function. Optimization approaches can be timodal and highly constrained engineering optimization
broadly divided into two types: deterministic methods and problems [3, 4]. Metaheuristic algorithms are modern
metaheuristic methods. As conventional optimization optimization methods, which commonly operate by some
approaches, deterministic methods apply specific defined principles and randomness to imitate natural phe-
nomena and are proving to be better than conventional
optimization methods in solving complex practical opti-
& Zhigang Jin mization problems.
[email protected] A lot of metaheuristic algorithms have been developed
Yiying Zhang over the last two decades. These algorithms can be roughly
[email protected] separated into the following four categories according to
Ye Chen different types of inspiration:
[email protected]
1. Swarm intelligence algorithms. These algorithms are
1
School of Electrical and Information Engineering, Tianjin inspired from some behaviour of animal and plant,
University, 92 Weijin Road, Tianjin 300072,
People’s Republic of China
such as foraging process of bird flocking in particle
2
swarm optimization (PSO) [5], obligate brood parasitic
College of Applied Science and Technology, Hainan
behaviour of some cuckoo species in cuckoo search
University, 58 People Avenue, Haikou 570228, People’s
Republic of China (CS) [6], echolocation behaviour of bats in bat

123
 Neural Computing and Applications

algorithm (BA) [7], leadership hierarchy and hunting new hybrid method based on NNA and GWO to solve
mechanism of grey wolves in grey wolf optimizer global optimization problems.
(GWO) [8], social behaviour of fireflies based on their NNA is one of the newest metaheuristic algorithms,
flashing characteristics in firefly algorithm (FA) [9], which is inspired by biological nervous systems and arti-
swarming behaviour of salps in salp swarm algorithm ficial neural networks (ANNs). Benefiting from the unique
(SSA) [10] and social behaviour of humpback whales structure of ANNs, NNA shows the following outstanding
in whale optimization algorithm (WOA) [11]. characteristics: (1) strong portability. NNA can be easily
2. Evolutionary algorithms. These algorithms are inspired applied to different optimization problems without adjust-
from some behaviour of the biological evolution, such ing other parameters except common control parameters
as reproduction, mutation and selection. Genetic algo- like population size and stopping criterion. Here, it is worth
rithm (GA) [12] and differential evolution (DE) [13] mentioning that common control parameters are essential
are two classical evolutionary algorithms. What’s for all population-based optimization algorithms; (2)
more, biogeography-based optimization (BBO) is a powerful ability of global search. NNA can find the global
recent proposed evolutionary algorithm, which imitates optimum with the minimum possibility of getting trapped
the biological migration behaviour in nature [14]. in local minima compared with many existing meta-
3. Human activity-related algorithms. These algorithms heuristic algorithms [17]. Therefore, NNA is easy to
are inspired from human activities, such as traditional implement and has good global search ability. However,
teaching mode in teaching–learning-based optimiza- NNA has the slow convergence speed that restricts its
tion (TLBO) [15], vehicle passing mechanisms on two- applications in some specific types of optimization prob-
lane highways in passing vehicle search (PVS) [1], sine lems with limitations like computation resource constraints
and cosine waves in sine cosine algorithm (SCA) [16] and high accuracy.
and artificial neural networks in nervous network GWO is a recent proposed metaheuristic algorithm,
algorithm (NNA) [17]. which mimics the hunting mechanism of grey wolves.
4. Physics-based algorithms. These algorithms are GWO has fast convergence speed and good global search
inspired from some physical phenomenon, such as ability, which has been used to solve many optimization
the principle of solid annealing in simulated annealing problems [24–27]. However, GWO may tend to conver-
algorithm (SA) [18] and the water cycle process in gence to local minima in solving some complex opti-
water cycle algorithm (WCA) [4]. mization problems. Once GWO got trapped in local optima
at the end of optimization process, it is very difficult for
The no-free-lunch (NFL) theorem [19] was proposed in
GWO to escape from local optima due to the limited
1997, which is an important theoretical basis of develop-
exploration ability [8, 28–30].
ment of optimization algorithms. According to NFL, there
Motivated by the characteristics of NNA and GWO, this
is no metaheuristic best suited for solving all optimization
paper presents an effective hybrid algorithm called grey
problems. In other words, one metaheuristic algorithm may
wolf optimization with neural network algorithm (GNNA).
achieve very promising results on a set of optimization
The core idea of this hybrid technique is to make the best
problems while it may present poor performance on
of good global search ability of NNA and fast convergence
another set of optimization problems. Thus NFL encour-
of GWO. Note that GNNA is not a simple combination of
ages researchers to develop new optimization algorithms or
NNA and GWO. Firstly, an improved NNA is proposed to
improve some existed ones for solving a wider range of
strengthen exploration of NNA by discarding the transfer
optimization problems. A technique which combines two
operator and introducing the random modification factor,
or more metaheuristic algorithms to produce hybrid
and an enhanced GWO is presented to balance exploration
methods that contain advantages of all constitutive algo-
and exploitation of GWO by adjusting the exploration rate
rithms can be a possible solution [20]. In [21], a hybrid
according to reinforcement learning principles. Then the
GSA-GA was presented to solve constrained optimization
improved NNA and the enhanced GWO are hybridized
problems. Xiong et al. [22] proposed a hybrid differential
based on dynamic population mechanism. More specifi-
evolution with whale optimization algorithm for parameter
cally, in each iteration, population is divided into two
extraction of solar photovoltaic models. A novel hybrid
halves according to fitness values of individuals. Then the
between the electromagnetism-like algorithm and the fire-
best half of individuals are optimized by the enhanced
fly algorithm for discrete structural optimization was
GWO to improve the convergence speed of GNNA, and the
introduced in [23]. Note that all above mentioned hybrid
worst half of individuals are processed by the improved
algorithms have been proven to be more efficient compared
NNA to strengthen the ability of GNNA in escaping from
with corresponding original methods. Considering the
local optima. At last, all individuals construct a new pop-
effectiveness of the hybrid technique, this paper proposes a
ulation after one iteration and the best individual is selected

123
Neural Computing and Applications

from the new population as the current optimum solution. 2. Update weight matrix.
The proposed GNNA is examined using 23 well-known The renewal equation of the weight matrix can be
unconstrained benchmark test functions, and its efficiency expressed as:
and superiority are highlighted against 13 metaheuristic   
 t t 
algorithms along with statistical tests. wtþ1
i ¼  wi þ 2  c  w t
opt  wi ; i ¼ 1; 2; . . .; N
The rest of this paper is organized as follows: Sect. 2 ð4Þ
describes NNA and GWO. Section 3 elaborates the
improved NNA, the enhanced GWO and the proposed where c is a random number between 0 and 1 and wtopt
GNNA in detail. The experimental results are illustrated in is the target weight vector. The target weight vector
Sect. 4. Section 5 presents conclusions and future work. wtopt and the target solution xtopt are updated simulta-
neously. If xtopt is equal to xts ðs 2 ½1; N Þ at time t,wtopt is
equal to wts .
2 Background 3. Bias operator.
In the bias operator, modification factor bt is used to
2.1 Neural network algorithm determine the bias proportion, and it is updated by [17]

NNA is inspired by the structure of ANNs and biological btþ1 ¼ 0:99  bt ð5Þ
nervous systems, which is a population-based optimization In NNA, bias operator consists of two parts: bias of
algorithm. In general, ANNs are used for prediction pur- population and bias of weight matrix. For bias of
poses in most cases, which try to close the gap between population, a random number Nbt is generated firstly
predicted solutions and given target solution by frequently according to the following equation:
adjusting the values of weight functions. However, an
optimization algorithm is aimed at finding an optimum Nbt ¼ dbt  De ð6Þ
solution from the given search space of the optimization Then a set T is produced by randomly selecting Nbt
problem. In order to adapt ANNs to be suitable for using as integers between 0 and D. Thus the bias of population
an optimization technique, the current best solution is can be described as (Let l ¼ ðl1 ; l2 ; . . .; lN Þ and u ¼
viewed as the target solution in NNA. The core idea of ðu1 ; u2 ; . . .; uN Þ are the lower and the upper bounds of
NNA is to close the gap between the target solution and the design variables, respectively)
other solutions included in the population by adjusting the  
xti;TðkÞ ¼ lTðkÞ þ uTðkÞ  lTðkÞ  u; k ¼ 1; 2; . . .; Nb
values of weight functions. NNA includes the following
four major stages: ð7Þ
1. Generate the new population. where u is a random number between 0 and 1, lTðkÞ and
In NNA, every individual i has its corresponding uTðkÞ are the upper bound and the lower bound of the
h i
weight vector wti ¼ wti;1 ; wti;2 ; . . .; wti;N , which can be TðkÞth variable. As for bias of weight matrix, a random
number Nwt is generated firstly by the following
defined as
equation:
X
N
wti;j ¼ 1; 0\wti;j \1; i ¼ 1; 2; . . .; N ð1Þ Nwt ¼ dbt  N e ð8Þ
j¼1
Then Nwt integers between 0 and N are randomly
where N is population size and t is the current number selected, which form a set Q. Thus the bias of weight
of iterations. Thus the process of generating the new matrix can be defined as
population can be described as
wti;QðbÞ ¼ c; b ¼ 1; 2; . . .; Nw ð9Þ
X
N
xtþ1
new;j ¼ wti;j  xti ; j ¼ 1; 2; . . .; N ð2Þ where c is a random number between 0 and 1.
i¼1 4. Transfer operator.
xtþ1 ¼ xti þ xtþ1 i ¼ 1; 2; . . .; N ð3Þ In NNA, the transfer operator is to transfer the current
i new;i ;
solution to a better solution towards the current best
where xti is solution of the ith individual at time t, wti;j is solution, which can be expressed as
the jth weight value of the ith individual at time t and
xtþ1
new;j is the weighted solution of the jth individual at
time t þ 1.

123
 Neural Computing and Applications

 
xtþ1
i ¼ xtþ1
i þ 2  /  xtopt  xtþ1
i ; i ¼ 1; 2; . . .; N
ð10Þ
where / is a random number between 0 and 1.
These steps show NNA is very simple and easy to carry
out. In addition, Fig. 1 presents the pseudo code of NNA.

2.2 Grey wolf optimizer

GWO is inspired by the leadership hierarchy and hunting

mechanism of grey wolves in nature [8]. The grey wolves
are strictly divided into four different levels according to
social dominant hierarchy from top down, namely the alpha
wolf (a), the beta wolf (b), the delta wolf (d) and the omega Fig. 2 Social dominant hierarchy of the grey wolf
wolf (x), which is shown in Fig. 2.
In terms of the mathematically model of GWO, it is At ¼ 2at :r1  at ð12Þ
based on the following assumptions [8]: (1) a, b and d t
C ¼ 2r2 ð13Þ
denote the fittest solution, the second best solution and the
third best solution, respectively; (2) the rest solutions are where r1 and r2 are the random vectors between 0 and
supposed to x; (3) the three best wolves (a, b and d) have 1, and at is the exploration rate. According to the
 
better knowledge about the potential location of prey than authors of GWO, at ¼ at1 ; at2 ; . . .; atN is linearly
the x wolves; (4) x wolves follow the three best wolves. decreased from 2 to 0 over the course of iterations,
Figure 3 shows the pseudo code of GWO, and it mainly which can be described as [8]
includes the following two steps:
2t
ati ¼ 2  ; i ¼ 1; 2; . . .; N ð14Þ
1. Encircling prey. Tmax
This encircling behaviour of grey wolves can be
where Tmax is the maximum number of iterations.
described by the following equations [8]:
  2. Hunting.
 
Xtþ1 ¼ X tp  At :Ct :X tp  X t  ð11Þ The hunt of grey wolves is usually guided by the
alpha grey. Sometimes the beta wolf and the delta wolf
where Xtp is the position vector of prey, At and Ct are may also participate in hunting. Note that it is very
the coefficient vectors, and Xt is the position vector of hard to get the location of prey in an abstract search
a grey wolf. The coefficient vectors At and Ct are space. Thus the authors of GWO suppose that the alpha
calculated as follows [8]: wolf, the beta wolf and the delta wolf have better
knowledge of the potential location of prey to simulate

Fig. 1 Pseudo code of standard NNA Algorithm

NNA 01. begin
02. Initialize the weight matrix and the population including N solutions
03. Calculate the fitness value of each solution and then set the optimal solution and the optimal weight
04. repeat
05. Generate the new population X t +1 via Eqs. (2) and (3) and update the new weight matrix W t +1 via Eq.(4)
06. for each individual i ∈ N do
07. if rand ≤ β t
08. Perform bias operator for updating the solution xit +1 by Eqs. (6-7) and the weight wit +1 by Eqs.(8-9)
09. else
10. Perform the transfer function operator for updating the solution xit +1 via Eq.(10)
11. end if
12. end for
13. Generate the new modification factor β t +1 via Eq. (5)
t +1 t +1
14. Calculate the fitness value of each solution and find the optimal solution xopt and the optimal weight xopt
15. Until (stop condition=false)
16. Post process results and visualization
17. end

123
Neural Computing and Applications

Fig. 3 Pseudo code of standard

GWO

the hunting behaviour of grey wolves. Mathematically, ht ¼ m ð19Þ

this hunting firstly obtains the three best solutions and
then updates other solutions, which can be expressed as Nbt ¼ dht  De ð20Þ
[8] Nwt ¼ dht  N e ð21Þ
 
X1 ¼ X ta  At1 :Ct1 :X ta  X t  ð15Þ where m is a random number between 0 and 1. Figure 4
  shows the pseudo code of the INNA.
 
X2 ¼ X tb  At2 :Ct2 :X tb  Xt  ð16Þ
  3.2 The enhanced GWO
X3 ¼ X td  At3 :Ct3 :X td  X t  ð17Þ
X1 þ X2 þ X3 In original GWO, exploration rate plays a key role to
Xtþ1 ¼ ð18Þ
3 balance exploration and exploitation, which is linearly
decreased from 2 to 0 as shown in Eq. (14). This method of
controlling the exploration rate proves efficient for solving
some optimization problems while it has the following
3 The proposed algorithms drawbacks: (1) once GWO got trapped in local optima at
the end of optimization process, it is very difficult for
This section is divided into three subsections. Firstly, an GWO to escape from local optima due to the limited
improved NNA is presented. Then an enhanced GWO are exploration ability [30]; (2) according to this method, every
introduced based on the reinforcement learning principles. grey wolf has the same exploration rate at every iteration.
At last, GNNA is proposed combining with the improved But search operators with similar behaviour may lead to the
NNA and the enhanced GWO. loss of diversity in the search space [31]; (3) the search
process of GWO is nonlinearly and highly complicated.
3.1 The improved NNA However, Eq. (14) is a linear equation, which cannot truly
reflect the search process [25].
The bias current plays a key role in the dynamics of the Reinforcement learning (RL) focus on goal-directed
neural networks model, which is always tied to a sur- learning, which is one of the powerful artificial intelligence
rounding condition to make the output of each neuron techniques motivated by behaviourist psychology and
respect the surrounding condition [17]. The bias operator in includes five elements: a discrete set of environment states,
NNA simulates the function of the bias current, which is an environment, a discrete set of agent actions, a reward
employed to adjust the exploration ability of NNA. How- assignment mechanism and a learning agent [31, 32]. In
ever, the running probability of the bias operator will order to overcome the above drawbacks of the original
gradually decline and the exploration ability will gradually GWO in controlling exploration rate, this paper proposes
weaken with the modification factor decreasing in NNA. In an enhanced GWO (EGWO). Considering the drawbacks
order to make full use of the good exploration ability of of calculating the exploration rate in original GWO, a new
NNA in our proposed GNNA, an improved NNA (INNA) method of adjusting exploration rate is introduced based on
is given in this paper. More specifically, INNA discards the reinforcement learning principles in the proposed enhanced
transfer operator and introduces the random modification GWO (EGWO), which can be defined as
factor ht into the bias operator. In INNA, the parameters Nbt
and Nwt can be rewritten as

123
 Neural Computing and Applications

Fig. 4 Pseudo code of INNA

    
atþ1
i ¼ ati :min ati  w þ s  g fitþ1  fit ; w ; 3.3 The proposed GNNA
ð22Þ
i ¼ 1; 2; . . .; N
As mentioned previously, the proposed GNNA is aim at
where w is a constant, s is penalty factor, fit is the fitness combining the good global search of INNA and fast con-
value of individual i at time t and gðÞ is a control function. vergence of EGWO. In GNNA, the information exchange
gðÞ can be defined as (take the minimum problem of EGWO and INNA is based on dynamic population
example): mechanism. More specifically, this mechanism can be
8
< 1 if x\0 (Decrease action) summarized as follows: (1) the individuals processed by
gð x Þ ¼ 1 if x [ 0 (Increase action) ð23Þ EGWO and INNA are from the same population. More
: specifically, the optimized population is divided into two
0 if x ¼ 0 ðKeep action)
halves based on the fitness values of individuals. The best
Moreover, considering the value of at decreasing from 2 to half of population is processed by EGWO to accelerate
0 in the basic GWO, w is set to 2. As for penalty factor s, it convergence speed of GNNA, and INNA is used to opti-
should not be too big to avoid disturbing the performance mize the worst half of population to make it avoid getting
advantages of the basic algorithm and is set to 0.01 in our trapped into local optimum to some extent; (2) after the
experiment. Figure 5 presents the reinforcement learning- individuals are optimized by EGWO and INNA, they form
based EGWO schematic. In Fig. 5, the agent (EGWO) a new population; (3) INNA and EGWO share the current
received the state (the current fitness value of one indi- global optimal solution to accelerate the convergence speed
vidual at time t þ 1 is superior, inferior or equal to its of GNNA. Figure 7 presents the flowchart of GNNA and a
fitness value at time t) and a reward associated with the last step-wise procedure for execution of GNNA is given in the
 
sate transition (s  g fitþ1  fit ) at time t. Then EGWO will following.
take an action (keeping, increasing or decreasing ati ) to
Step 1 Initialization information.
make a transition to new state, and the cycle will be
Initialize parameters. These parameters include the
repeated at time t þ 1. Figure 6 shows the pseudo code of
maximum number of function evaluations Tmax , popula-
the EGWO.
tion size N, the lower bounds of design variables l, the
upper bounds of design variables u, dimension of
problem D and fitness function f ðÞ. In addition, the
current of number iteration time t is 0, the current
number of function evaluations Tcurrent is set to 0, the
penalty factor s is set to 0.01, and the constant w is equal
to 2.
(1:1) Initialize population. A random population X t is
generated based on the initialization parameters,
which can be described as

Fig. 5 The reinforcement learning-based EGWO schematic

123
Neural Computing and Applications

Fig. 6 Pseudo code of EGWO

Start

Initialize parameters

Initialize population and weight matrix by Eqs. (24-28)

Calculate fitness value of each individual, select the

optimum solution and update the current number of
function evaluations by Eq. (29)

Yes Meet the stopping

The optimum solution
criteria ?

No

The population is divided into twohalves based on fitness values

End of individuals. The best half of population is optimized by
EGWO and the worst half of population is processed by INNA

The worst solution is replaced with the global optimal

select three best wolves ( α, β and σ)
solution. The target solution and the target weight are
selected

Generate the new population by Eqs. (2) and (3) Calculate position vector of each grey wolf by Eq. (15-18)

Update the vectors A and C by Eq. (12) and Eq.

t t
Update the weight matrix by Eq. (4)
(13), respectively
INNA for the worst EGWO for the best
half of population half of population
Perform the bias operator for updating population by Eqs.
(7), (9) and (20). Perform the bias operator for updating Calculate the fitness value of population
weight matrix by Eqs. (9), (19) and (21).

Calculate the fitness value of population Update the exploration rate by Eqs. (25), (26) and (27)

The new population obtained by INNA The new population obtained by EGWO

Construct a new population

Select the optimum solution and update the current number of function
evaluations by Eq. (29)

Fig. 7 The flowchart of GNNA

123
 Neural Computing and Applications

 T each individual i, the bias operator is imple-

X t ¼ xt1 ; xt2 ; . . .; xtN
2 t 3 mented by Eqs. (7), (9), (19), (20) and (21). At
x1;1 xt1;2 . . . xt1;D
6 xt2;1 xt2;2 . . . xt2;D7 last, the new population Xtþ1worst is obtained by
6 7
¼ 6. .. .. 7 ð24Þ INNA and the fitness values of individuals are
4 .. . . 5
calculated.
xtN;1 xtN;2 . . . t
xN;D (5:3) Perform EGWO for the population Xtbest . Firstly,
xti;j ¼ li þ ðui  li Þ  l ð25Þ the three best wolves (a, b and d) are selected.
Secondly, the position of each grey wolf is
where l is a random number between 0 and 1. calculated by Eqs. (15–18). Thirdly, the vectors
(1:2) Initialize weight matrix. In GNNA, INNA is At and Ct via Eqs. (12) and (13), respectively.
employed to optimize the worst half of popula- Fourthly, the fitness value of each wolf is
tion. Thus the weight matrix W t can be initialized calculated. At last, the exploration rate of each
by individual is updated according to Eqs. (25–27)
 T and the new population X tþ1
W t ¼ wt1 ; wt2 ; . . .; wtE best is generated by
2 t 3 EGWO.
w1;1 wt1;2 . . . wt1;E
6 wt2;1 wt2;2 . . . wt2;E 7 Step 6 Construct population.
6 7
¼ 6. .. .. 7 ð26Þ
4 .. . . 5 Population X tþ1 tþ1
best and population X worst compose the new
tþ1
wtE;1 wtE;2 . . . wtE;E population X .
Step 7 Population evaluation.
wti;j ¼ t; i ¼ 1; 2; . . .; E; j ¼ 1; 2; . . .; E The optimal solution Gt is selected and the current
ð27Þ number of function evaluations is updated by Eq. (29).
wt Then Step 4 is executed.
wti ¼ PE i t ð28Þ
j¼1 wi;j

where t is a random value between 0 and 1 and E 4 Numerical experiment and analysis
is equal to 0:5  N.
Step 2 Population evaluation. In this section, 23 well-known complex numerical opti-
The fitness values of individuals are calculated and the mization problems are used to examine the performance of
optimal solution Gt is selected. the proposed GNNA. This section is divided into the fol-
Step 3 Update the number of function evaluations. lowing three parts. Section 4.1 introduces the experiment
The current number of function evaluations Tcurrent is settings including benchmark test functions, parameters
updated by settings and evaluation metrics. Section 4.2 is to compare
the performance between GNNA and variants of NNA and
Tcurrent ¼ Tcurrent þ N ð29Þ GWO. The performance comparisons between GNNA and
Step 4 Termination criteria. several state-of-the-art metaheuristic algorithms are pre-
If the current number of function evaluations Tcurrent is sented in Sect. 4.3.
greater than the maximum number of function evalua-
tions Tmax , the algorithm stops; otherwise, go to Step 5. 4.1 Experiment settings
Step 5 Population optimization.
4.1.1 Benchmark test functions
(5:1) The population is divided into two halves based
on the fitness values of individuals. The best half The selected 23 well-known test functions have been listed
of population and the worst half of population are in Table 1, which have been employed to evaluate the
marked as X tbest and Xtworst , respectively. optimization performance of different algorithms
(5:2) Perform INNA for the population Xtworst . Firstly, [2, 28, 31, 33–35]. More specifically, these functions have
the worst solution of the population X tworst is different features including unimodal (F1–F4), multimodal
replaced with the optimal solution Gt . Secondly, (F5–F19) and rotated multimodal (F20–F23). In general,
target solution xtopt and the target weight wtopt are multimodal functions are more complicated than unimodal
selected. Thirdly, the population Xtworst and the functions, which have more than one local minimum.
weight matrix W t are updated according to Nineteen of 23 functions are multimodal functions as
Eqs. (2–3) and (4), respectively. Fourthly, for shown in Table 1, which provides a better examination of

123
Neural Computing and Applications

Table 1 Twenty-three benchmark test functions. ‘‘U’’, ‘‘M’’ and ‘‘R’’ stand for unimodal, multi-model and rotated, respectively
No. Name Definition Range Type
PD
F1 Sphere f ð xÞ ¼ 2
i¼1 xi
½100; 100 U
PD 2
F2 Step f ð xÞ ¼ i¼1 ðbxi þ 0:5cÞ ½100; 100 U
PD
F3 Noise f ð xÞ ¼ i¼1: i x4i
þ random½0; 1Þ ½1:28; 0:64 U
PD QD
F4 Schwefel 2.22 f ð xÞ ¼ i¼1: jxi jþ i¼1 jxi j
½10; 10 U
F5 Schwefel 2.21 f ð xÞ ¼ maxi fjxi j; 1  i  Dg ½100; 100 M
PD 2 QD
F6 Griewank 1
f ð xÞ ¼ 4000 i¼1 xi 
xiffi
i¼1 cos i þ 1
p ½600; 600 M
P
F7 Alpine f ð xÞ ¼ D i¼1 jxi sinðxi Þ þ 0:1xi j
½10; 10 M
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P  PD 
F8 Ackley ½32; 32 M
f ð xÞ ¼ 20 exp 0:2 D1 D i¼1 xi
2  exp 1
D i¼1 cos 2pxi þ 20 þ e
PD
F9 Rastrigin f ð xÞ ¼ 2
i¼1 xi  10 cosð2pxi Þ þ 10 ½5:12; 5:12 M
qffiffiffiffiffiffiffiffiffiffiffiffiffi
P 
F10 Schaffer sin2
D
x2 0:5
½100; 100 M
i¼1 i
f ð xÞ ¼ 0:5þ  PD 2
1þ0:001 x2
i¼1 i

F11 Noncontinuous Rastrigin PD 2 xi ; jxi j\0:5 ½5:12; 5:12 M

f ð xÞ ¼ i¼1 yi  10 cosð2pyi Þ þ 10; yi ¼
0:5ðroundð2xi ÞÞ; jxi j  0:5
PD
F12 Cigar f ð xÞ ¼ x21 þ 106 6
i¼2 xi
½100; 100 M
PD
F13 Tablet 6
f ð xÞ ¼ 10  þ x21 6
i¼2 xi
½1; 1 M
P h   i
F14 Rosenbrock f ð xÞ ¼ D1 2 2
þðxi  1Þ2 ½10; 10 M
i¼1 100 xiþ!  xi
P  2 
F15 Bohachevsky f ð xÞ ¼ D1 2
i¼1 xi þ xiþ1 0:3 cosð3pxi Þ  0:4 cosð4pxiþ1 Þ þ 0:7
½15; 15 M
P  2 2
F16 Pixon&price f ð xÞ ¼ ðx1  1Þ2 þ D ½10; 10 M
i¼2 i 2xi  xi1
P
F17 Sum of different powers f ð xÞ ¼ D i¼1 jxi j
iþ1 ½1; 1 M
 PD P 
F18 Cosine mixture f ð xÞ ¼  0:1 i¼1 cosð5pxi Þ  D 2
i¼1 xi þ 0:1D
½1; 1 M
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PD 2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PD 2ffi
F19 Salomon ½100; 100 M
f ð xÞ ¼ 1  cos 2p i¼1 xi þ 0:1 i¼1 xi

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
PD 2ffi  P 
F20 Rotated Ackley ½32; 32 R, M
f ð xÞ ¼ 20 exp 0:2 1
D i¼1 zi  exp D1 D
i¼1 cos 2pzi þ 20 þ e; z ¼ x M
PD
F21 Rotated Rastrigin f ð xÞ ¼ 2
i¼1 zi  10 cosð2pzi Þ þ 10; z ¼ x M ½5:12; 5:12 R, M
PD 2 QD
F22 Rotated Griewank f ð xÞ ¼ 1
i¼1 xi 
xiffi
i¼1 cos i þ 1; z ¼ x
p M ½600; 600 R, M
4000
(
F23 Rotated noncontinuous PD yi ; jyi j\0:5 ½5:12; 5:12 R, M
2
Rastrigin f ð xÞ ¼ i¼1 zi  10 cosð2pzi Þ þ 10; zi ¼ ;y ¼ x M
0:5ðroundð2yi ÞÞ; jyi j  0:5

optimization performance of the proposed method in grey wolf optimizer (IGWO) [25]. The parameters of the
solving complex optimization problems. In addition,M is compared algorithms were taken from the original litera-
an orthogonal matrix in Table 1. All test functions should ture and have been shown in Table 2. Moreover, some
be minimized, which have a global minimum at f ðxÞ ¼ 0. parameters of GWO, SSA, SCA, WOA, IGWO and
Besides, the test dimension of all functions is set to 30. AGWO are related to the maximum number of function
evaluations, which are not listed in Table 2 and can be
4.1.2 Parameters settings found in the corresponding references.
In addition, considering the results of a single run might
In order to better show the efficiency and robustness of the be unreliable, all algorithms applied were executed in 30
proposed algorithm in solving the complex optimization independent runs for the same test function. Then the
problems, 13 optimizers are selected to compare with average values of the obtained results were recorded. The
GNNA. These algorithms include DE, CS, GWO, SCA, maximum number of function evaluations was considered
NNA, TLBO, WOA, SSA, PSO, INNA, EGWO, aug- as stopping condition, which was 5000 multiples by
mented grey wolf optimizer (AGWO) [24] and improved dimension size for all functions [17].

123
 Neural Computing and Applications

Table 2 Parameter settings of the applied algorithms

Algorithm Parameters References

GWO Population size = 30 [8]

NNA Population size = 50 [17]
WOA Population size = 30 [11]
SSA Population size = 30 [10]
SCA Population size = 30 [16]
TLBO Population size = 50 [15]
CS Population size = 15, discover probability = 0.25 [6]
INNA Population size = 50 –
EGWO Population size = 30, penalty factor = 0.01, control constant = 2 –
GNNA Population size = 30, penalty factor = 0.01, control constant = 2 –
AGWO Population size = 30 [24]
IGWO Population size = 30 [25]
DE Population size = 100, differential amplification factor = 0.5, crossover probability constant = 0.9 [13]
PSO Population size = 40, inertia weight = 0.7298, personal learning coefficient = 1.4986, social learning [36]
coefficient = 1.4986

4.1.3 Evaluation metrics rank test with a significance level a = 0.05 are produced by
using the mean optimal solutions of 30 runs of GNNA and
To compare the optimization performance among different the compared algorithms to solve benchmark test functions.
algorithms, three quality indicators are used in our exper- Tables 4 and 6 present the test results. In these tables, ‘‘H’’
iments, which can be described as follows. is marked as ‘‘1’’, which means there is significant differ-
The first one is the value-based method. Mean value and ence between GNNA and the compared algorithm; ‘‘H’’ is
standard deviation are good indicators to evaluate accuracy marked as ‘‘0’’, which indicates there is no significant
and stability of one algorithm. The smaller the mean value difference between GNNA and the compared algorithm.
is, the stronger the global optimization ability of the ‘‘S’’ is marked as ‘‘?’’, which imply the proposed method
algorithm is; the smaller the standard deviation is, the more is superior to the compared algorithm; ‘‘S’’ is marked as
stability the algorithm is. Tables 3 and 5 show the statis- ‘‘=’’, which indicates the proposed method has the same
tical results of mean value and standard deviation. In these performance with the compared algorithm. Moreover, the
tables, ‘‘ME’’, ‘‘Std’’ and ‘‘TR’’ stand for mean value, stand last column of the each of these tables under the heading
deviation and tied rank, respectively. Moreover, the best w/t/l represents the win, tie and lose counts of the proposed
results are highlighted in bold. GNNA over the compared algorithms.
The second one is the rank-based method. Tied rank
(TR) [31, 34] is employed to compare the performance 4.2 GNNA comparison with variants of GWO
among the applied algorithms. More specifically, TR and NNA
assigns rank 1 to the algorithm with the best mean value;
rank 2 to the second best and rank M (the number of the GNNA is inspired from basic GWO and NNA. Thus it is
applied algorithms) to the Mth best. If several algorithms necessary to compare the optimization performance among
have the same mean values, they will share the average of GNNA, NNA, GWO, and some recent variants of NNA
ranks. Tables 3 and 5 show the results of tied rank. and GWO. In this section, the applied algorithms include
Moreover, the last column of the each of these tables under GNNA, NNA, GWO, INNA, EGWO, IGWO and AGWO.
the heading average rank represents the mean tied rank of The experimental results obtained by the applied algo-
every algorithm for 23 benchmark test functions. rithms on 23 benchmark test functions with 30-dimensional
The third one is the statistical test-based method. Wil- are presented in Table 3, and the best results are high-
coxon signed-rank test [37] is used to compare the per- lighted in boldface. Figure 8 gives the statistical results
formance between the proposed GNNA and the other based on Table 3. In Fig. 8, ‘‘Win’’ means the number of
algorithms, which has been widely used to compare the functions that GNNA can offer better mean values than the
performance between two optimization algorithms compared algorithm; ‘‘Tie’’ indicates the number of func-
[38–41]. In this research, the results of Wilcoxon signed- tions that GNNA can achieve the same mean values with

123
Neural Computing and Applications

Table 3 The experimental results obtained by seven different algorithms for 23 benchmark functions
No. Metric GWO EGWO AGWO IGWO NNA INNA GNNA

F1 ME 0.00E100 0.00E100 0.00E100 1.19E-306 9.56E-22 1.64E?04 0.00E100

SD 0.00E100 0.00E100 0.00E100 0.00E100 1.18E-21 3.33E?03 0.00E100
TR 2.5 2.5 2.5 5 6 7 2.5
F2 ME 0.00E100 0.00E100 0.00E100 0.00E100 0.00E100 1.58E?04 0.00E100
SD 0.00E100 0.00E100 0.00E100 0.00E100 0.00E100 2.60E?03 0.00E100
TR 3.5 3.5 3.5 3.5 3.5 7 3.5
F3 ME 1.16E-04 1.12E-04 1.08E-04 5.87E-04 2.20E-03 6.78E?00 2.11E205
SD 6.80E-05 7.77E-05 7.90E-05 2.33E-04 1.63E-03 2.22E?00 2.32E205
TR 4 3 2 5 6 7 1
F4 ME 7.40E-178 9.96E-261 5.66E-282 9.24E-178 2.53E-11 5.70E?01 0.00E100
SD 0.00E100 0.00E100 0.00E100 0.00E100 8.02E-11 1.40E?01 0.00E100
TR 4 3 2 5 6 7 1
F5 ME 1.72E-88 1.57E-146 2.53E-145 3.25E-85 2.09E-04 3.75E?04 0.00E100
SD 9.40E-88 7.64E-146 1.38E-144 1.64E-84 2.06E-04 3.83E?03 0.00E100
TR 4 2 3 5 6 7 1
F6 ME 0.00E100 0.00E100 0.00E100 2.00E-03 8.93E-03 1.67E?02 0.00E100
SD 0.00E100 0.00E100 0.00E100 4.07E-03 1.73E-02 3.27E?01 0.00E100
TR 2.5 2.5 2.5 5 6 7 2.5
F7 ME 0.00E100 0.00E100 0.00E100 2.74E?00 3.73E-06 2.98E?01 0.00E100
SD 0.00E100 0.00E100 0.00E100 2.12E?00 1.09E-05 3.94E?00 0.00E100
TR 4 2 2 6 5 7 2
F8 ME 8.11E-15 6.81E-15 6.45E-15 1.38E-14 3.70E-11 1.77E?01 8.88E216
SD 1.47E-15 1.70E-15 1.79E-15 4.89E-15 3.48E-11 6.97E-01 0.00E100
TR 4 3 2 5 6 7 1
F9 ME 0.00E100 0.00E100 0.00E100 2.72E?01 3.75E-10 2.40E?02 0.00E100
SD 0.00E100 0.00E100 0.00E100 4.17E?00 1.94E-09 2.98E?01 0.00E100
TR 2.5 2.5 2.5 6 5 7 2.5
F10 ME 3.13E-03 3.13E-03 3.13E-03 3.13E-03 4.06E-03 4.99E-01 3.70E217
SD 9.86E-12 1.76E-10 3.71E-10 2.49E-08 1.44E-03 2.46E-04 8.42E217
TR 2 3 4 5 6 7 1
F11 ME 0.00E100 0.00E100 0.00E100 2.64E?01 3.43E-15 2.10E?02 0.00E100
SD 0.00E100 0.00E100 0.00E100 6.68E?00 4.78E-15 2.64E?01 0.00E100
TR 2.5 2.5 2.5 6 5 7 2.5
F12 ME 0.00E100 0.00E100 0.00E100 2.54E-307 7.73E-22 1.75E?04 0.00E100
SD 0.00E100 0.00E100 0.00E100 0.00E100 1.86E-21 3.03E?03 0.00E100
TR 2.5 2.5 2.5 5 6 7 2.5
F13 ME 0.00E100 0.00E100 0.00E100 0.00E100 2.07E-23 3.28E?00 0.00E100
SD 0.00E100 0.00E100 0.00E100 0.00E100 5.20E-23 5.74E-01 0.00E100
TR 3 3 3 3 6 7 3
F14 ME 2.65E?01 2.67E?01 2.65E?01 2.72E?01 2.08E?01 2.90E?05 7.56E201
SD 8.79E-01 8.02E-01 6.62E-01 6.48E-01 2.29E-01 1.13E?05 2.20E201
TR 4 5 3 6 2 7 1
F15 ME 0.00E100 0.00E100 0.00E100 0.00E100 1.48E-17 1.02E?03 0.00E100
SD 0.00E100 0.00E100 0.00E100 0.00E100 8.11E-17 1.79E?02 0.00E100
TR 3 3 3 3 6 7 3
F16 ME 6.67E-01 6.67E-01 6.67E-01 7.02E-01 6.67E-01 1.47E?05 3.78E201
SD 2.60E-08 4.55E-08 4.77E-09 1.08E-01 2.11E213 5.34E?04 4.32E-01
TR 5 4 3 6 2 7 1

123
 Neural Computing and Applications

Table 3 (continued)
No. Metric GWO EGWO AGWO IGWO NNA INNA GNNA

F17 ME 0.00E100 0.00E100 0.00E100 0.00E100 1.69E-38 8.92E-03 0.00E100

SD 0.00E100 0.00E100 0.00E100 0.00E100 4.51E-38 8.24E-03 0.00E100
TR 3 3 3 3 6 7 3
F18 ME 0.00E100 0.00E100 0.00E100 1.48E-17 1.63E-16 4.11E?00 0.00E100
SD 0.00E100 0.00E100 0.00E100 8.11E-17 2.18E-16 4.47E-01 0.00E100
TR 2.5 2.5 2.5 5 6 7 2.5
F19 ME 1.07E-01 9.99E-02 9.99E-02 1.73E-01 2.10E-01 1.54E?01 0.00E100
SD 2.54E-02 1.58E-10 1.22E-10 5.21E-02 4.81E-02 1.54E?00 0.00E100
TR 4 2 3 5 6 7 1
F20 ME 8.88E216 8.88E216 8.88E216 8.88E216 8.88E216 2.34E-06 8.88E216
SD 0.00E100 0.00E100 0.00E100 0.00E100 0.00E100 3.11E-06 0.00E100
TR 3.5 3.5 3.5 3.5 3.5 7 3.5
F21 ME 1.87E?01 2.05E?01 1.37E?01 5.05E?01 1.15E?02 2.91E?02 0.00E100
SD 1.58E?01 2.11E?01 3.06E?01 1.10E?01 3.85E?01 2.32E?01 0.00E100
TR 3 4 2 5 6 7 1
F22 ME 4.60E-03 6.62E-03 4.70E-03 4.43E-02 7.71E-03 1.44E?02 0.00E100
SD 1.26E-02 1.72E-02 1.29E-02 3.76E-02 9.24E-03 2.52E?01 0.00E100
TR 2 4 3 6 5 7 1
F23 ME 2.25E?01 3.36E?01 2.97E?01 5.28E?01 1.28E?02 2.54E?02 0.00E100
SD 1.25E?01 1.41E?01 2.75E?01 1.54E?01 3.10E?01 2.04E?01 0.00E100
TR 2 4 3 5 6 7 1
Average rank 3.22 3.00 2.70 4.96 5.27 7.00 1.87

the compared algorithm. According to Table 3 and Fig. 8, algorithms in terms of global search ability and stability.
GNNA is superior or equal to the compared algorithms in This indicates the combination of EGWO and INNA is
terms of mean value. More specifically, INNA is inferior to very successful in solving optimization problems. How-
GNNA on all test functions. GNNA is superior to NNA on ever, we also can find an interesting phenomenon that the
all test functions except F2 and F20. GNNA outperforms proposed EGWO outperforms the basic GWO while the
IGWO on all test functions except F2, F13, F15, F17 and basic NNA surpasses the proposed INNA according to the
F20. In addition, GWO, EGWO and AGWO show out- sorted results. We can explain this phenomenon from two
standing global search ability, which can offer the same aspects as follows. Firstly, compared with the basic GWO,
results with GNNA on F1, F2, F6, F7, F9, F11, F12, F13, F15, the proposed EGWO adjusts the exploration rate according
F17, F18 and F20 while they still cannot surpass GNNA on to the reinforcement learning principles, which can better
the another 11 test functions. Besides, based on results of balance exploration and exploitation of the basic GWO.
the average rank, the applied algorithms can be sorted into The experimental results demonstrate the improved strat-
the following order from good to bad: GNNA, AGWO, egy in EGWO is helpful to enhance the global search
EGWO, GWO, IGWO, NNA and INNA. capability of the basic GWO. Secondly, the basic NNA
Table 4 gives the statistical results produced by Wil- balance its exploration and exploitation by bias operator
coxon signed-rank test. As can be seen from Table 4, and transfer operator, respectively. However, the proposed
GNNA outperforms INNA on all test functions. GNNA is INNA discards the transfer operator to make the most of
superior or equal to GWO, IGWO and NNA on all test exploration of NNA. In other words, exploration and
functions. It also may be noted that EGWO and AGWO exploitation are serious imbalance in the INNA. Thus NNA
show strong global search ability. EGWO and AGWO are can show better optimization performance than INNA.
equal to GNNA on 12 and 13 functions, respectively. In order to compare the convergence performance
However, EGWO and AGWO are still inferior to GNNA among different algorithms, Fig. 9 shows several typical
on 11 and 10 functions, respectively. averaged curves over 30 independent runs. The selected
Based on the above analysis, the proposed GNNA shows functions include 2 unimodal functions (i.e. F1 and F4), 10
better optimization performance than the compared multimodal functions (i.e. F5, F6, F7, F9, F10, F11, F12, F13,

123
Neural Computing and Applications

Table 4 Wilcoxon signed-rank test results between GNNA and six compared algorithms
No. GNNA versus
GWO EGWO AGWO IGWO NNA INNA
P value H S P value H S P value H S P value H S P value H S P value H S

F1 1.73E-06 1 ? 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?

F2 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ?
F3 1.92E-06 1 ? 9.32E-06 1 ? 3.52E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F4 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F5 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F6 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 6.10E-05 1 ? 2.69E-05 1 ? 1.73E-06 1 ?
F7 1.73E-06 1 ? 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F8 1.01E-07 1 ? 6.25E-07 1 ? 7.86E-07 1 ? 1.05E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F9 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ? 1.64E-06 1 ? 1.73E-06 1 ?
F10 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.00E-06 1 ? 1.73E-06 1 ?
F11 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ? 1.14E-04 1 ? 1.73E-06 1 ?
F12 1.73E-06 1 ? 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F13 1.73E-06 1 ? 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F14 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F15 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ?
F16 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ? 1.73E-06 1 ?
F17 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ? 1.73E-06 1 ?
F18 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 9.77E-04 1 ? 1.73E-06 1 ?
F19 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 5.00E-07 1 ? 1.73E-06 1 ?
F20 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ?
F21 1.23E-05 1 ? 5.61E-06 1 ? 6.10E-05 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F22 6.25E-02 0 = 3.13E-02 1 ? 1.25E-01 0 = 8.86E-05 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F23 1.73E-06 1 ? 1.73E-06 1 ? 1.23E-05 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
w/t/l 14/9/0 11/12/0 10/13/0 18/5/0 20/3/0 23/0/0

compared algorithms converge easily to the local optima in

solving complex rotated functions. Here, it is worth men-
tioning that GNNA, GWO, AGWO and EGWO can con-
verge to the optimum solutions on F1, F6, F7, F9, F11, F12,
F13, F17 and F19 while GNNA can converge faster than
GWO, AGWO and EGWO on these functions.

4.3 GNNA comparison with state-of-the-art

algorithms

In this section, GNNA is compared with several state-of-

the-art algorithms. The selected algorithms include two
classical algorithms (i.e. PSO and DE) and five recent
Fig. 8 The compared results of mean values between GNNA and six
proposed algorithms (i.e. CS, TLBO, SCA, SSA and
other algorithms
WOA).
F17 and F19) and 3 rotated functions (i.e. F21, F22 and F23). Table 5 presents the experimental results obtained by
As can be seen form Fig. 9, the proposed GNNA not only GNNA and seven state-of-the-art algorithms for 23
converges quickly towards the global optimal solution but benchmark test functions with 30-dimensional and the best
also offers higher accuracy. Moreover, we also can find the results are highlighted in boldface. Figure 10 gives the
statistical results based on Table 5. ‘‘Win’’ and ‘‘Tie’’ have

123
 Neural Computing and Applications

(a) F1 (b) F4 (c) F5

(d) F6 (e) F7 (f) F9

(g) F10 (h) F11 (i) F12

(j) F13 (k) F17 (l) F19

(m) F21 (n) F22 (o) F23

Fig. 9 Several typical convergence curves obtained by GNNA and six compared algorithms

123
Neural Computing and Applications

Table 5 The experimental results obtained by GNNA and seven state-of-the-art algorithms for 23 benchmark functions
No. Metric DE PSO CS SCA TLBO WOA SSA GNNA

F1 ME 2.46E-16 3.12E-74 3.99E-25 9.77E-25 2.19E-267 0.00E100 4.95E-09 0.00E100

SD 2.55E-16 7.79E-74 8.79E-25 4.68E-24 0.00E100 0.00E100 1.00E-09 0.00E100
TR 7 4 5 6 3 1.5 8 1.5
F2 ME 0.00E100 4.33E-01 0.00E100 0.00E100 0.00E100 0.00E100 6.97E?00 0.00E100
SD 0.00E100 8.17E-01 0.00E100 0.00E100 0.00E100 0.00E100 4.59E?00 0.00E100
TR 3.5 7 3.5 3.5 3.5 3.5 8 3.5
F3 ME 9.37E-03 4.43E-03 7.03E-02 6.46E-03 3.07E-04 2.95E-04 1.52E-02 2.11E205
SD 2.04E-03 2.42E-03 3.34E-02 5.28E-03 9.71E-05 4.19E-04 6.45E-03 2.32E205
TR 6 4 8 5 3 2 7 1
F4 ME 7.47E-08 9.50E-32 1.15E-15 6.12E-29 8.99E-134 0.00E100 4.79E-01 0.00E100
SD 4.35E-08 4.68E-31 3.01E-15 2.55E-28 1.03E-133 0.00E100 1.01E?00 0.00E100
TR 7 4 6 5 3 1.5 8 1.5
F5 ME 7.81E-01 6.26E-08 2.49E-02 1.28E?02 1.95E-57 7.94E?01 3.31E-07 0.00E100
SD 5.15E-01 7.05E-08 2.92E-02 2.92E?02 6.96E-57 2.25E?02 1.52E-07 0.00E100
TR 6 3 5 8 2 7 4 1
F6 ME 3.11E-15 1.15E-02 4.02E-03 3.90E-04 1.02E-13 6.30E-04 1.60E-02 0.00E100
SD 1.32E-14 1.62E-02 6.80E-03 1.81E-03 5.60E-13 3.45E-03 1.42E-02 0.00E100
TR 2 7 6 4 3 5 8 1
F7 ME 2.07E-09 1.31E-15 2.86E-01 3.34E-08 4.10E-269 0.00E100 2.63E?00 0.00E100
SD 5.49E-09 5.37E-16 2.16E-01 1.83E-07 0.00E100 0.00E100 1.63E?00 0.00E100
TR 5 4 7 6 3 1.5 8 1.5
F8 ME 6.79E-09 6.72E-01 2.49E?00 1.64E?01 6.22E-15 3.49E-15 1.92E?00 8.88E216
SD 2.70E-09 7.66E-01 1.13E?00 7.55E?00 1.81E-15 2.46E-15 7.94E-01 0.00E100
TR 4 5 7 8 3 2 6 1
F9 ME 1.72E?02 3.97E?01 1.44E?01 1.75E?00 1.05E?01 0.00E100 7.35E?01 0.00E100
SD 1.14E?01 1.08E?01 4.50E?00 6.77E?00 6.39E?00 0.00E100 1.96E?01 0.00E100
TR 8 6 5 3 4 1.5 7 1.5
F10 ME 1.41E-02 8.67E-03 4.05E-02 3.33E-03 3.13E-03 2.19E-03 1.20E-02 3.70E217
SD 2.99E-03 2.93E-03 2.40E-02 7.86E-04 9.93E-10 1.46E-03 2.68E-03 8.42E217
TR 7 5 8 4 3 2 6 1
F11 ME 1.46E?02 3.30E?01 1.93E?01 1.81E?01 1.89E?01 0.00E100 8.11E?01 0.00E100
SD 1.57E?01 1.12E?01 4.49E?00 2.50E?01 6.05E?00 0.00E100 1.90E?01 0.00E100
TR 8 6 5 3 4 1.5 7 1.5
F12 ME 3.49E-16 2.41E-73 7.39E-25 1.70E-24 4.07E-268 0.00E100 5.21E-09 0.00E100
SD 2.43E-16 8.76E-73 2.89E-24 9.23E-24 0.00E100 0.00E100 8.56E-10 0.00E100
TR 7 4 5 6 3 1.5 8 1.5
F13 ME 6.73E-20 1.79E-76 3.81E-28 1.50E-30 8.87E-271 0.00E100 2.74E-01 0.00E100
SD 1.01E-19 6.00E-76 1.50E-27 5.72E-30 0.00E100 0.00E100 1.33E-01 0.00E100
TR 7 4 6 5 3 1.5 8 1.5
F14 ME 1.29E?01 1.51E?01 2.74E?01 2.75E?01 1.82E?01 2.51E?01 3.94E?01 7.56E201
SD 1.09E?00 1.17E?01 2.74E?01 6.49E-01 1.09E?00 2.34E-01 2.55E?01 2.20E201
TR 2 3 6 7 4 5 8 1
F15 ME 3.33E-15 3.34E?00 5.87E?00 0.00E100 0.00E100 0.00E100 9.94E?00 0.00E100
SD 4.78E-15 1.79E?00 2.20E?00 0.00E100 0.00E100 0.00E100 2.38E?00 0.00E100
TR 5 6 7 2.5 2.5 2.5 8 2.5
F16 ME 6.67E-01 6.67E-01 6.93E-01 6.67E-01 6.67E-01 6.67E-01 7.17E-01 3.78E201
SD 1.37E-12 1.60E216 8.47E-02 2.36E-04 3.57E-15 1.84E-06 9.01E-02 4.32E-01
TR 4 4 7 4 4 4 8 1

123
 Neural Computing and Applications

Table 5 (continued)
No. Metric DE PSO CS SCA TLBO WOA SSA GNNA

F17 ME 6.54E-35 1.66E-166 3.38E-25 2.97E-39 0.00E100 0.00E100 3.69E-08 0.00E100

SD 3.58E-34 0.00E100 1.72E-24 8.56E-39 0.00E100 0.00E100 1.84E-08 0.00E100
TR 6 4 7 5 2 2 8 2
F18 ME 0.00E100 5.42E-01 1.03E-01 0.00E100 0.00E100 0.00E100 1.29E?00 0.00E100
SD 0.00E100 2.58E-01 1.30E-01 0.00E100 0.00E100 0.00E100 3.45E-01 0.00E100
TR 3 7 6 3 3 3 8 3
F19 ME 2.02E-01 4.47E-01 8.20E-01 1.17E-01 1.03E-01 1.37E-01 5.10E-01 0.00E100
SD 1.08E-02 1.31E-01 2.71E-01 3.79E-02 1.83E-02 6.68E-02 9.60E-02 0.00E100
TR 5 6 8 3 2 4 7 1
F20 ME 8.88E216 8.88E216 8.88E216 8.88E216 8.88E216 8.88E216 4.50E-10 8.88E216
SD 0.00E100 0.00E100 0.00E100 0.00E100 0.00E100 0.00E100 4.83E-10 0.00E100
TR 4 4 4 4 4 4 8 4
F21 ME 1.91E?02 3.93E?01 8.98E?01 7.06E?01 1.08E?01 7.01E?01 6.64E?01 0.00E100
SD 9.78E?00 1.44E?01 1.91E?01 7.15E?01 5.51E?00 6.31E?01 1.82E?01 0.00E100
TR 8 3 7 6 2 5 4 1
F22 ME 4.93E-04 1.18E-02 1.48E-03 6.48E-03 0.00E100 6.89E-03 1.02E-02 0.00E100
SD 1.88E-03 1.32E-02 6.83E-03 3.55E-02 0.00E100 1.53E-02 9.63E-03 0.00E100
TR 3 8 4 5 1.5 6 7 1.5
F23 ME 1.65E?02 4.90E?01 8.09E?01 1.10E?02 2.91E?01 8.12E?01 8.77E?01 0.00E100
SD 1.29E?01 1.35E?01 1.95E?01 5.23E?01 1.72E?01 6.59E?01 2.37E?01 0.00E100
TR 8 3 4 7 2 5 6 1
Average rank 5.46 4.73 5.93 5.00 2.89 3.20 7.17 1.61

offer the same results with GNNA on 12 functions (i.e. F1,

F2, F4, F7, F9, F11, F12, F13, F15, F17, F18 and F20) while
WOA is still inferior to GNNA on 11 functions (i.e. F3, F5,
F6, F8, F10, F14, F16, F19, F21, F22 and F23). Besides, based
on the results of the average rank, the applied algorithms
can be sorted into the following order from good to bad:
GNNA, TLBO, PSO, CS, WOA, SSA, SCA and DE.
Table 6 shows the statistical results produced by Wil-
coxon signed-rank test. As can be seen from Table 6,
GNNA outperforms SSA on all test functions. GNNA is
superior or equal to DE, PSO, CS, SCA, TLBO and WOA
on all test functions. More specifically, DE, PSO, CS, SCA
and TLBO are inferior to GNNA on 20, 22, 21, 19 and 17
functions, respectively. Note that WOA show strong global
Fig. 10 The compared results of mean values between GNNA and
seven state-of-the-art algorithms search ability, and it is equal to GNNA on 12 functions.
However, GNNA still can surpass WOA on nearly half of
the same meanings in Figs. 8 and 10. According to Table 5 functions.
and Fig. 10, GNNA is superior or equal to the compared To sum up, the proposed GNNA outperforms the com-
algorithms in terms of mean value. More specifically, pared algorithms in terms of global search ability and
GNNA outperforms SSA on all test functions. GNNA can stability. Furthermore, Fig. 11 shows several typical aver-
surpass DE, PSO, CS, SCA and TLBO on all test functions aged curves over 30 independent runs to compare the
except three (i.e. F2, F18 and F20), one (i.e. F20), two (i.e. convergence performance among different algorithms. The
F2 and F20), four (i.e. F2, F15, F18 and F20) and six (i.e. F1, selected functions include F1, F3, F5, F6, F7, F8, F9, F10,
F15, F17, F18, F20 and F22) functions, respectively. In F11, F12, F13, F14, F17, F18 and F23. As can be seen form
addition, WOA shows strong competitiveness, which can Fig. 11, the proposed GNNA not only converges quickly

123
 Table 6 Wilcoxon signed-rank test results between GNNA and seven state-of-the-art algorithms
No. GNNA versus
DE PSO CS SCA TLBO WOA SSA
P value H S P value H S P value H S P value H S P value H S P value H S P value H S
Neural Computing and Applications

F1 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.00E?00 0 = 1.73E-06 1 ?

F2 1.00E?00 0 = 1.95E-03 1 ? 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.66E-06 1 ?
F3 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.64E-05 1 ? 1.73E-06 1 ?
F4 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.00E?00 0 = 1.73E-06 1 ?
F5 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F6 1.75E-05 1 ? 1.92E-04 1 ? 9.77E-04 1 ? 1.25E-01 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ?
F7 1.73E-06 1 ? 1.64E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.00E?00 0 = 1.73E-06 1 ?
F8 1.73E-06 1 ? 1.63E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 8.21E-07 1 ? 7.90E-05 1 ? 1.73E-06 1 ?
F9 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 7.81E-03 1 ? 2.56E-06 1 ? 1.00E?00 0 = 1.73E-06 1 ?
F10 1.73E-06 1 ? 1.68E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 2.38E-05 1 ? 1.70E-06 1 ?
F11 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 3.79E-06 1 ? 1.72E-06 1 ? 1.00E?00 0 = 1.73E-06 1 ?
F12 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.00E?00 0 = 1.73E-06 1 ?
F13 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.00E?00 0 = 1.73E-06 1 ?
F14 1.73E-06 1 ? 2.35E-06 1 ? 3.18E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ?
F15 1.80E-05 1 ? 1.64E-06 1 ? 1.67E-06 1 ? 1.00E?00 0 = 1.00E?00 0 ? 1.00E?00 0 = 1.73E-06 1 ?
F16 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ? 3.11E-05 1 ?
F17 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ?
F18 1.00E?00 0 = 1.68E-06 1 ? 1.22E-04 1 ? 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ?
F19 1.73E-06 1 ? 1.51E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 1.90E-06 1 ? 1.36E-06 1 ?
F20 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.00E?00 0 = 1.73E-06 1 ?
F21 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 5.61E-06 1 ? 1.73E-06 1 ? 2.70E-05 1 ? 1.73E-06 1 ?
F22 1.73E-06 1 ? 5.94E-05 1 ? 8.86E-05 1 ? 1.00E?00 0 = 1.00E?00 0 = 7.81E-03 1 ? 1.73E-06 1 ?
F23 1.73E-06 1 ? 1.73E-06 1 ? 1.73E-06 1 ? 2.56E-06 1 ? 1.73E-06 1 ? 2.70E-05 1 ? 1.73E-06 1 ?
w/t/l 20/3/0 22/1/0 21/2/0 17/6/0 17/6/0 10/13/0 23/0/0

123
 Neural Computing and Applications

(a) F1 (b) F3 (c) F5

(d) F6 (e) F7 (f) F8

(g) F9 (h)F10 (i) F11

(j) F12 (k) F13 (l) F14

(m) F17 (n) F18 (o) F23

Fig. 11 Several typical convergence curves obtained by GNNA and seven state-of-the-art algorithms

123
Neural Computing and Applications

towards the global optimal solution but also offers higher References
accuracy. On the contrary, the compared algorithms con-
verge easily to the local optima in solving complex opti- 1. Savsani P, Savsani V (2016) Passing vehicle search (PVS): A
mization problems such as F5, F6, F10 and F23. In addition, novel metaheuristic algorithm. Appl Math Model 40:3951–3978.
https://doi.org/10.1016/j.apm.2015.10.040
GNNA and WOA can converge to the optimum solutions 2. Zhang J, Xiao M, Gao L, Pan Q (2018) Queuing search algo-
on F1, F7, F9, F11, F12, F13, F17 and F18, while GNNA can rithm: a novel metaheuristic algorithm for solving engineering
converge faster than WOA on these functions. optimization problems. Appl Math Model 63:464–490. https://
doi.org/10.1016/j.apm.2018.06.036
3. Rao RV, Savsani VJ, Vakharia DP (2012) Teaching–learning-
based optimization: an optimization method for continuous non-
5 Conclusions linear large scale problems. Inf Sci 183:1–15. https://doi.org/10.
1016/j.ins.2011.08.006
In order to solve the global numerical optimization problems, 4. Eskandar H, Sadollah A, Bahreininejad A, Hamdi M (2012)
Water cycle algorithm—a novel metaheuristic optimization
this paper presents an efficient hybrid method called grey wolf method for solving constrained engineering optimization prob-
optimization with neural network algorithm (GNNA). The lems. Comput Struct 110–111:151–166. https://doi.org/10.1016/j.
core idea of GNNA is to make full use of the good global compstruc.2012.07.010
search ability of NNA and fast convergence of GWO. Firstly, 5. Kennedy J, Eberhart R (1995) Particle swarm optimization. In:
Proceedings of ICNN’95—international conference on neural
NNA and GWO are improved to enhance the optimization networks, vol 4, pp 1942–1948
performance and INNA and EGWO are obtained. Then INNA 6. Yang X, Deb S (2009) Cuckoo search via Lévy flights. In: 2009
and EGWO are hybridized based on dynamic population world congress on nature biologically inspired computing
mechanism. In the optimization process, the population is (NaBIC), pp 210–214
7. Yang X-S (2010) A new metaheuristic bat-inspired algorithm. In:
divided into two halves according to the fitness values. The González JR, Pelta DA, Cruz C et al (eds) Nature inspired
best half of population is optimized by EGWO and the worst cooperative strategies for optimization (NICSO 2010). Springer,
half of population is processed by INNA. Berlin, pp 65–74
A comprehensive set of 23 well-known unconstrained 8. Mirjalili S, Mirjalili SM, Lewis A (2014) Grey wolf optimizer.
Adv Eng Softw 69:46–61. https://doi.org/10.1016/j.advengsoft.
benchmark functions are employed to examine the perfor- 2013.12.007
mance of GNNA compared with 13 metaheuristic algorithms 9. Gandomi AH, Yang X-S, Alavi AH (2011) Mixed variable
including two basic algorithms (i.e. NNA and GWO), four structural optimization using firefly algorithm. Comput Struct
variants of the basic algorithms and seven state-of-the-art 89:2325–2336. https://doi.org/10.1016/j.compstruc.2011.08.002
10. Mirjalili S, Gandomi AH, Mirjalili SZ et al (2017) Salp swarm
algorithms. According to the experimental results, the pro- algorithm: a bio-inspired optimizer for engineering design prob-
posed GNNA is superior or equal to the compared algorithms lems. Adv Eng Softw 114:163–191. https://doi.org/10.1016/j.
for all test functions in terms of mean values. Note that the advengsoft.2017.07.002
proposed GNNA can find global optimal solutions on 17 11. Mirjalili S, Lewis A (2016) The whale optimization algorithm.
Adv Eng Softw 95:51–67. https://doi.org/10.1016/j.advengsoft.
functions including three unimodal functions (i.e. F1, F2, F4), 2016.01.008
11 multimodal functions (i.e. F5, F6, F7, F9, F11, F12, F13, F15, 12. Goldberg DE, Holland JH (1988) Genetic algorithms and
F17, F18, F19) and three rotated functions (i.e. F21, F22 and machine learning. Mach Learn 3:95–99. https://doi.org/10.1023/
F23). Such comparisons suggest that GNNA shows good A:1022602019183
13. Rahnamayan S, Tizhoosh HR, Salama MMA (2008) Opposition-
global search ability and fast convergence speed and is clearly based differential evolution. IEEE Trans Evol Comput 12:64–79.
seen to be more successful in both solution quality and com- https://doi.org/10.1109/TEVC.2007.894200
putational efficiency compared with other algorithms. 14. Simon D (2008) Biogeography-based optimization. IEEE Trans
The future research will focus on the following two Evol Comput 12:702–713. https://doi.org/10.1109/TEVC.2008.
919004
aspects. On the one hand, we intend to apply GNNA to 15. Rao RV, Savsani VJ, Vakharia DP (2011) Teaching–learning-
some practical optimization problems such as multilevel based optimization: a novel method for constrained mechanical
image segmentation and economic dispatch problems. On design optimization problems. Comput Aided Des 43:303–315.
the other hand, we try to improve the performance of other https://doi.org/10.1016/j.cad.2010.12.015
16. Mirjalili S (2016) SCA: a sine cosine algorithm for solving
metaheuristic algorithms by the proposed reinforcement optimization problems. Knowl Based Syst 96:120–133. https://
learning mechanism. doi.org/10.1016/j.knosys.2015.12.022
17. Sadollah A, Sayyaadi H, Yadav A (2018) A dynamic meta-
heuristic optimization model inspired by biological nervous
systems: neural network algorithm. Appl Soft Comput
Compliance with ethical standards 71:747–782. https://doi.org/10.1016/j.asoc.2018.07.039
18. Kirkpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by
simulated annealing. Science 220:671–680. https://doi.org/10.
Conflict of interest The authors declare that they have no conflict of
1126/science.220.4598.671
interest.

123
 Neural Computing and Applications

19. Wolpert DH, Macready WG (1997) No free lunch theorems for 32. Kaelbling LP, Littman ML, Moore AP (1996) Reinforcement
optimization. IEEE Trans Evol Comput 1:67–82. https://doi.org/ learning: a survey. J Artif Intell Res 4:237–285
10.1109/4235.585893 33. Salgotra R, Singh U, Saha S (2018) New cuckoo search algo-
20. Zang H, Zhang S, Hapeshi K (2010) A review of nature-inspired rithms with enhanced exploration and exploitation properties.
algorithms. J Bionic Eng 7:S232–S237. https://doi.org/10.1016/ Expert Syst Appl 95:384–420. https://doi.org/10.1016/j.eswa.
S1672-6529(09)60240-7 2017.11.044
21. Garg H (2019) A hybrid GSA-GA algorithm for constrained 34. Sun Y, Wang X, Chen Y, Liu Z (2018) A modified whale opti-
optimization problems. Inf Sci 478:499–523. https://doi.org/10. mization algorithm for large-scale global optimization problems.
1016/j.ins.2018.11.041 Expert Syst Appl 114:563–577. https://doi.org/10.1016/j.eswa.
22. Xiong G, Zhang J, Yuan X et al (2018) Parameter extraction of 2018.08.027
solar photovoltaic models by means of a hybrid differential 35. Long W, Jiao J, Liang X, Tang M (2018) An exploration-en-
evolution with whale optimization algorithm. Sol Energy hanced grey wolf optimizer to solve high-dimensional numerical
176:742–761. https://doi.org/10.1016/j.solener.2018.10.050 optimization. Eng Appl Artif Intell 68:63–80. https://doi.org/10.
23. Le DT, Bui D-K, Ngo TD et al (2019) A novel hybrid method 1016/j.engappai.2017.10.024
combining electromagnetism-like mechanism and firefly algo- 36. Wang H, Wu Z, Rahnamayan S et al (2011) Enhancing particle
rithms for constrained design optimization of discrete truss swarm optimization using generalized opposition-based learning.
structures. Comput Struct 212:20–42. https://doi.org/10.1016/j. Spec Issue Interpret Fuzzy Syst 181:4699–4714. https://doi.org/
compstruc.2018.10.017 10.1016/j.ins.2011.03.016
24. Qais MH, Hasanien HM, Alghuwainem S (2018) Augmented 37. Derrac J, Garcı́a S, Molina D, Herrera F (2011) A practical
grey wolf optimizer for grid-connected PMSG-based wind energy tutorial on the use of nonparametric statistical tests as a
conversion systems. Appl Soft Comput 69:504–515. https://doi. methodology for comparing evolutionary and swarm intelligence
org/10.1016/j.asoc.2018.05.006 algorithms. Swarm Evol Comput 1:3–18. https://doi.org/10.1016/
25. Long W, Liang X, Cai S et al (2017) A modified augmented j.swevo.2011.02.002
Lagrangian with improved grey wolf optimization to constrained 38. Mafarja M, Aljarah I, Heidari AA et al (2018) Binary dragonfly
optimization problems. Neural Comput Appl 28:421–438. https:// optimization for feature selection using time-varying transfer
doi.org/10.1007/s00521-016-2357-x functions. Knowl Based Syst 161:185–204. https://doi.org/10.
26. Khairuzzaman AKM, Chaudhury S (2017) Multilevel threshold- 1016/j.knosys.2018.08.003
ing using grey wolf optimizer for image segmentation. Expert 39. Sun G, Ma P, Ren J et al (2018) A stability constrained adaptive
Syst Appl 86:64–76. https://doi.org/10.1016/j.eswa.2017.04.029 alpha for gravitational search algorithm. Knowl Based Syst
27. Sahoo A, Chandra S (2017) Multi-objective grey wolf optimizer 139:200–213. https://doi.org/10.1016/j.knosys.2017.10.018
for improved cervix lesion classification. Appl Soft Comput 40. Martı́nez-Peñaloza M-G, Mezura-Montes E (2018) Immune
52:64–80. https://doi.org/10.1016/j.asoc.2016.12.022 generalized differential evolution for dynamic multi-objective
28. Zhang X, Kang Q, Cheng J, Wang X (2018) A novel hybrid environments: an empirical study. Knowl Based Syst
algorithm based on biogeography-based optimization and grey 142:192–219. https://doi.org/10.1016/j.knosys.2017.11.037
wolf optimizer. Appl Soft Comput 67:197–214. https://doi.org/ 41. Yi J, Gao L, Li X et al (2019) An on-line variable-fidelity sur-
10.1016/j.asoc.2018.02.049 rogate-assisted harmony search algorithm with multi-level
29. Lu C, Gao L, Li X, Xiao S (2017) A hybrid multi-objective grey screening strategy for expensive engineering design optimization.
wolf optimizer for dynamic scheduling in a real-world welding Knowl Based Syst 170:1–19. https://doi.org/10.1016/j.knosys.
industry. Eng Appl Artif Intell 57:61–79 2019.01.004
30. Emary E, Zawbaa HM, Grosan C (2018) Experienced gray wolf
optimization through reinforcement learning and neural net- Publisher’s Note Springer Nature remains neutral with regard to
works. IEEE Trans Neural Netw Learn Syst 29:681–694. https:// jurisdictional claims in published maps and institutional affiliations.
doi.org/10.1109/TNNLS.2016.2634548
31. Rakhshani H, Rahati A (2017) Snap-drift cuckoo search: a novel
cuckoo search optimization algorithm. Appl Soft Comput
52:771–794. https://doi.org/10.1016/j.asoc.2016.09.048

123