Knowledge-Based Systems 150 (2018) 175–197

Contents lists available at ScienceDirect

Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys

Multi-objective spotted hyena optimizer: A Multi-objective

optimization algorithm for engineering problems
Gaurav Dhiman∗, Vijay Kumar
Computer Science and Engineering Department, Thapar University, Patiala, Punjab, India

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a multi-objective version of recently developed Spotted Hyena Optimizer (SHO)
Received 9 August 2017 called Multi-objective Spotted Hyena Optimizer (MOSHO). It is used to optimize the multiple objectives
Revised 1 March 2018
problems. In the proposed algorithm, a fixed-sized archive is employed for storing the non-dominated
Accepted 8 March 2018
Pareto optimal solutions. The roulette wheel mechanism is used to select the effective solutions from
Available online 8 March 2018
archive to simulate the social and hunting behaviors of spotted hyenas. The proposed algorithm is tested
Keywords: on 24 benchmark test functions and compared with six recently developed metaheuristic algorithms. The
Multi-objective optimization proposed algorithm is then applied on six constrained engineering design problems to demonstrate its
Spotted hyena optimizer applicability on real-life problems. The experimental results reveal that the proposed algorithm performs
Constrained optimization better than the others and produces the Pareto optimal solutions with high convergence.
Constrained engineering design problems
© 2018 Elsevier B.V. All rights reserved.

1. Introduction these solution points will distribute uniformly on the front [5]. The
higher dimensional problems can have extremely complex hyper-
In recent years metaheuristic techniques have become a pop- surface as its Pareto front [3,6–8]. Thus, it is more challenging to
ular research area to solve real-life problems. Due to computa- solve such high-dimensional problems.
tionally inexpensiveness and ease to understand, these have been The concept of multi-objective optimization using stochastic
widely used in various engineering design applications [1]. techniques was proposed by David Schaffer [9]. The advantages
The most challenging characteristic of these algorithms is of these techniques are local optima avoidance and gradient-free
multi-objectivity. The multi-objective optimization deals with opti- mechanism that made them applicable to real world problems.
mization of multiple and competing objective functions. The main The application of multi-objective optimization techniques can be
challenge in multi-objective optimization is to address the multi- found in various fields such as civil engineering [10], mechanical
ple objectives which are frequently in conflict. There are two main engineering [11,12], system engineering [13], bioinformatics [14],
approaches for multi-objective optimization such as priori and pos- software engineering [15], artificial intelligence [16], and other
teriori [2,3]. fields [17–19].
In priori approaches, a multi-objective problem is converted to Some of the well-known optimization techniques are: Non-
a single-objective with a set of weights that defines the signifi- dominated Sorting Genetic Algorithm 2 (NSGA-2) [20], Multi-
cance of each objective to solve the problem. A posteriori method objective Particle Swarm Optimization (MOPSO) [21], and
permits to explore the performance of problem and selects one of Multi-objective Evolutionary Algorithm based on Decomposi-
the obtained solutions based on their obligations [4]. tion (MOEA/D) [22]. These are able to estimate the true Pareto
In contrast to single-objective, there is no single solution in optimal solutions. However, these are unable to solve all types of
multi-objective optimization and represents various trade-offs be- optimization problems [23]. The motivation behind this work is
tween the objectives. Therefore, it is more time consuming to ob- to propose a novel multi-objective optimization algorithm called
tain the Pareto fronts because it usually requires to produce many Multi-objective Spotted Hyena Optimizer (MOSHO) which is based
points on the Pareto front for good approximations. Even accu- on Spotted Hyena Optimizer (SHO) [1]. The main contributions of
rate solutions on a Pareto front, there is still no guarantee that this paper are:

• An archive component has been incorporated into SHO that is

∗
Corresponding author. responsible for storing all non-dominated Pareto optimal solu-
E-mail address: [email protected] (G. Dhiman). tions which are obtained so far.

https://doi.org/10.1016/j.knosys.2018.03.011
0950-7051/© 2018 Elsevier B.V. All rights reserved.
176 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

• A group selection approach has been proposed to decide A solution x  X is called Pareto optimal if and only if:
whether the solutions are integrated in group corresponding to
the position of prey from the archive.
y
  X | y ≺ x (5)
• A grid mechanism has been incorporated in SHO to omit one Definition 3. Pareto optimal set.
of the most crowded sections and improves the non-dominated The set of all Pareto optimal solutions including all non-
solutions in the archive. dominated solutions of a problem is called Pareto optimal set if
The performance of the proposed MOSHO has been evaluated and only if:
on seven bi-objective and three tri-objective multi-objective test Ps = {x, y
  X | ∃y
x} (6)
problems taken from the CEC (Congress on Evolutionary Computa-
tion 2009) competition and special session on multi-objective op- Definition 4. Pareto optimal front.
timization algorithms [24]. The proposed algorithm has also been A set containing the objective values corresponding to Pareto
evaluated on five ZDT [25] and nine DTLZ (Scalable test problems) optimal solutions in Pareto optimal set is called Pareto optimal
[26] multi-objective test problems. The results have been com- front. It is defined as:
pared and verified with six recently developed optimization tech- Pf = { f (x ) | x  Ps } (7)
niques such as MOPSO [21], NSGA-2 [20], MOEA/D [22], PESA-2
[27], SPEA-2 [28], and MOACO [29]. The four well-known perfor-
2.2. Related works
mance metrics have been used to assess these algorithms. The re-
sults show that MOSHO can prove to be a very promising meta-
In the last few decades, a wide range of multi-objective tech-
heuristic technique in the field of multi-objective optimization.
niques have been developed. Multi-objective metaheuristic based
The rest of this paper is discussed as follows: Section 2 presents
techniques deal with many difficulties such as infeasible solutions,
the basic concepts of multi-objective optimization and re-
diversity of solutions, and optimum separation [4]. To deal with
lated work done in the field of multi-objective optimization.
these difficulties, two conditions are to be satisfied. First, the infor-
Section 3 describes the concepts of SHO and then proposes
mation is exchanged between the search space and search agents.
the MOSHO algorithm. The results and discussions are given in
Second, the multi-objective techniques should help to estimate the
Section 4. In Section 5, the performance of MOSHO is tested on
whole true Pareto optimal front in a single simulation run.
six constrained multi-objective engineering design problems and
The popular multi-objective metaheuristic is Non-dominated
compared with other well-known optimizers. The conclusion and
Sorting Genetic Algorithm 2 (NSGA-2) [20]. This algorithm uses a
outlines of some future works are drawn in Section 6.
fast non-dominated sorting method, elitist technique, and a nich-
ing operator. The population of NSGA-2 algorithm begins with ran-
2. Background
domized in which each individual is grouped based on the non-
dominated sorting technique. To assist with the selection, muta-
This section provides the basic concepts of multi-objective op-
tion, and recombination operators, the second population is cre-
timization techniques.
ated. Both of these populations create a new one and then sorted
again by the non-dominated sorting technique. The probability to
2.1. Basic concepts of multi-objective optimization
select a new individual depends on the non-domination level for
the final population. Finally, the whole process is run until the sat-
Multi-objective optimization refers to the optimization with
isfactory result is found.
more than one objective (criterion) function of a given problem.
The second popular multi-objective metaheuristic technique is
It can be formulated as [30,31]:
Multi-objective Particle Swarm Optimization (MOPSO) [21]. It uses
Minimize : F (z ) = [ f1 (z ), f2 (z ), . . . , fn (z )] (1) the concept of PSO algorithm. MOPSO uses the external archive for
storing and retrieving the Pareto optimal solutions. Additionally, a
Subject to: mutation operator is integrated in MOPSO to enhance the random-
(2) ness and diversity in the distribution of Pareto optimal solutions.
gi (z ) ≥ 0, i = 1, 2, . . . , m
Multi-objective Evolutionary Algorithm based on Decomposi-
hi (z ) = 0, i = 1, 2, . . . , p (3) tion (MOEA/D) [22] is another popular multi-objective metaheuris-
tics which decomposes a problem into scalar subproblems to op-
]T
where z = [z1 , z2 , . . . , zk is the vector of decision variables, m is timize simultaneously. These subproblems are usually equal to the
the number of inequality constraints, p is the number of equality population size in which each subproblem is assigned a weighting
constraints, gi is the ith inequality constraints, hi is the ith equality vector to combine the objective functions into a single objective
constraints, and obj is the number of objective functions f i : Rob j → function. During search process, two tasks namely co-operation
R. and completion, are performed. In co-operation task, the neigh-
In multi-objective, the solutions in a search space cannot be bouring members can collaborate the subproblem to generate so-
compared by relational operators due to multi-criterion compari- lution. Furthermore, the solution of neighbouring subproblem(s) is
son metrics. The comparison of two solutions was first proposed replaced with the subproblem solution if it is better than the cur-
by Edgeworth [32] and further extended by Pareto [33]. The math- rent one. This task is known as competition task. MOEA/D has less
ematical formulation of Pareto dominance is described as [30]: computational complexity and fast convergence speed than NSGA-
2.
Definition 1. Pareto dominance.
Angus and Woodward [29] proposed the enhanced version of
Assume there are two vectors x=(x1 , x2 , . . . , xr ) and
=(y1 , y2 , . . . , yr ). Vector x is said to dominate vector y Ant Colony Optimization (ACO) to solve multi-objective optimiza-
y  (such
tion problems. A new classification technique is proposed in Multi-
as x≺y) if and only if:
objective Ant Colony Optimization (MOACO) which provides nearby
∀i  {1, 2, . . . , r} : fi (x ) ≤ fi (y ) ∧ ∃i  {1, 2, . . . , r} : fi (x ) < fi (y ) solutions for intended problem domains. MOACO is based on the
basic concepts of ACO. These are selection of pheromone model,
(4)
construction process, solution estimation, and updating process.
Definition 2. Pareto optimality. From an optimization viewpoint, this technique is able to balance
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 177

the exploratory (diversity preserving) behavior with exploitative In spotted hyenas, female members are dominant and live in
(quality enhancing) behavior. their clan. However, male members leave their clan when they be-
Gong et al. [34] proposed a Non-dominated Neighbor Immune come adults and join a new clan. In a new family, they are lowest
Algorithm (NNIA) for multi-objective optimization. NNIA used the ranking members to get their share of meal. A male member who
non-dominated neighbor-based selection, immune inspired opera- has joined the clan always stays with the same members (friends)
tor, heuristic search operators, and elitism. NNIA algorithm might for a long time. Whereas, a female is always assured of a stable
lose the diversity as it refuses all the dominated antibodies when place. An interesting fact about the spotted hyena is that they pro-
non-dominated antibodies are selected for proportional cloning. To duce sound to communicate with each other during the searching
overcome this issue, Yang et al. [35] developed an enhanced ver- of food source.
sion of NNIA named as NNIA2, by integrating an adaptive ranks According to Ilany et al. [46], spotted hyenas usually rely on
clone and K-nearest neighbour approaches to improve the diver- a network of trusted friends that have more than 100 members.
sity. They usually tie up with another spotted hyena that is a friend of
There are different types of Multi-objective algorithms have a friend or linked in some way through kinship rather than any
been proposed in recent years. These are Multi-objective Cat unknown spotted hyena. Spotted hyenas are social animals that
Swarm Optimization (MOCSO) [36], Multi-objective Teaching- can communicate with each other through specialized calls such
Learning based Optimization algorithm (MO-TLBO) [37], as postures and signals. They use multiple sensory procedures to
Multi-objective Artificial Bee Colony algorithm (MOABC) [38], recognize their kin and other individuals. They can also recognize
Multi-objective Gravitational Search Algorithm (MOGSA) [39], third party kin and rank the relationships between their clan mates
Multi-objective network clustering algorithm (GMOEA-net) [40], during social decision making. The spotted hyena track prey by
Multi-objective framework for SAR image segmentation (IMIS) sight, hearing, and smell. Cohesive clusters are helpful for an effi-
[41], Quantum Particle Swarm Optimization (QPSO) [42], Multi- cient co-operation between spotted hyenas. In this work, the hunt-
agent Genetic Algorithm (MAGA) [43], Immune Genetic Algorithm ing technique and social relation of spotted hyenas are mathemat-
(IGA) [44], RDS-NSGA-II [45], Multi-objective Flower Pollination ically modeled to design the multi-objective SHO algorithm.
Algorithm (MOFPA) [5]. All of the above-mentioned algorithms
are not able to solve all types of optimization problems according 3.2. Spotted Hyena Optimizer (SHO)
to NFL theorem [23]. It is probable that a new algorithm should
be able to solve a problem that cannot be solved by the existing To propose the multi-objective version of SHO [1], firstly the
algorithms. A novel multi-objective version of the recently devel- essential concept of this algorithm is discussed. The social relation
oped SHO is proposed in the preceding section for finding optimal and hunting behaviors of spotted hyenas are the main inspiration
solutions of multi-objective problems. of this algorithm. SHO algorithm mimics the cohesive clusters be-
tween the trusted spotted hyenas. The four main steps of SHO are
searching, encircling, hunting, and attacking. In SHO algorithm, the
3. Proposed algorithm
hunting behavior is guided by the group of trusted friends (so far
solutions) towards the best search agent and saves the best opti-
In this section, we first describe the basic concepts of SHO fol-
mal solutions.
lowed by brief description of multi-objective version of SHO.
In order to simulate the encircling behavior of spotted hyenas,
the following equations are used [1]:
3.1. Inspiration  h =| B
 · Pp (x ) − P (x ) |
D (8)
Social relationships are dynamic in nature. These are affected
by the changes in relationship among comprising the network and P (x + 1 ) = Pp (x ) − E · D
h (9)
individual leaving or joining the population. The animal behavior where D  h represents the distance between prey and spotted hyena.
has been classified into three categories [46]: x indicates the current iteration. Pp and P represents the position
vector of prey and spotted hyena, respectively. || and · are the ab-
• The first category includes environmental factors such as re-
 and E are the
solute value and multiplication vector, respectively. B
source availability and competition with other animal species.
co-efficient vectors.
• The second category focuses on social preferences based on in-
B and E are computed as:
dividual behavior.
• The third category has less attention from scientists which in- B 1
 = 2 · rd (10)
cludes the social relations of species itself.

The social relation between animals is the inspiration of our

 2 − h
E = 2h · rd (11)
work and correlates this behavior to spotted hyena which is sci-  
entifically named as Crocuta. h = 5 − 5
Iteration ×
Hyenas are large dog-like carnivores. They live in savannas, MaxIterations (12)
grasslands, sub-deserts, and forests of both Africa and Asia. They where, Iteration = 0, 1, 2, . . . , MaxIterations
live 10–12 years in the wild and up to 25 years in imprisonment.
There are four known species of hyena such as spotted, striped, Here, h is linearly decreased from 5 to 0 over the course of iter-
ations. rd and rd  are the random vectors that lie in the range
brown, and aardwolf. These differ in size, behavior, and type of 1 2

diet. All of these species have a bear-like attitude. [0,1]. There are different number of places which can be reached
Spotted hyenas are skillful hunters and largest of three other with respect to the current position by adjusting the value of vec-
tors B and E . This algorithm saves the optimal solution achieved so
hyena species (i.e., striped, brown, and aardwolf). Spotted Hyena
is also known as laughing hyena because its sounds is much sim- far and compels other search agents to update their positions.
ilar to a human laugh. There are spots on their fur reddish brown In order to simulate the hunting behavior of spotted hyenas and
in color with black spots. Spotted hyenas are complicated, intel- find the promising regions of search space, the following equations
ligent, and highly social animals with really dreadful reputation. are defined as:
They have the ability to fight endlessly for territory and food.  h =| B
D  · Ph − Pk | (13)
178 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Fig. 1. Searching Prey (|E | > 1).

Fig. 2. Attacking Prey (|E | < 1).

In order to perform the optimization, SHO algorithm com-

Pk = Ph − E · D
h (14) mences with generating a set of random solutions as the popu-
lation. During optimization, the search agents make a cluster to-
wards the best search agent for updating their positions using
Eqs. (15) - (17). In the meantime, parameters h and E are linearly
Ch = Pk + Pk+1 + . . . + Pk+N (15)
decreased during the course of iterations. Finally, the positions of
where N represents the number of iterations which can be com- search agents that make a cluster are considered as optimal best
puted as: solutions when the termination criteria is satisfied. From [1], it has
been confirmed that SHO is able to solve optimization problems
N = countnos (Ph , Ph+1 , Ph+2 , . . . , (Ph + M
 )) (16) efficiently. The high exploration capability of SHO provides better
results as compared to the existing metaheuristics. Due to this, the
solutions produced by SHO do not stuck in local optima. This moti-
C
P (x + 1 ) = h (17) vates us to propose the multi-objective version of SHO in the pre-
N ceding subsection.
where nos indicates the number of solutions which are far similar
to best optimal solution in the given search space, Ch represents a 3.3. Multi-objective Spotted Hyena Optimizer (MOSHO)
group or cluster of optimal solutions, M  indicates a random vector
in [0.5, 1]. P (x + 1 ) saves N best optimal solutions and updates the To develop a multi-objective version of SHO, we have intro-
positions of other search agents according to the position of best duced two components (see Fig. 8). These components are very
search agent. similar to those of MOPSO [47]. The first component is an archive
The exploration is assurance by vector E with random values which is accountable for storing the best non-dominated Pareto
which are greater than 1 or less than 1 which compel the search optimal solutions. The second component is a group selection ap-
agents to move away from the prey as shown in Figs. 1 and 2. An- proach to choose nearby solutions corresponding to the position
other constituent of SHO is B  which is responsible for exploration.
of prey from the archive. These components are described in the
It contains random values in the range of [0, 5] which provides the preceding subsections.
weight of prey. To demonstrate the effect of distance and random
behavior, assume B >1 precedence than B <1 as seen in Eq. (10) that
will helpful for exploration and local optima avoidance. 3.3.1. Archive
The exploitation of SHO algorithm commences when | E |<1 The archive is a storage part of all non-dominated Pareto opti-
with random values of E are in [−1, 1]. Therefore, Fig. 5 shows the mal solutions obtained so far. This archive component has an abil-
next position of a search agent lies between its current position ity to evenly spread around the Pareto front if it is concave, con-
and the position of the prey which will helpful to meet towards vex, and disconnected. It consists of two main parts such as archive
an estimated position of prey. controller and grid.
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 179

Fig. 3. The adaptive grid mechanism of individuals.

Fig. 5. 3D positions of the spotted hyenas.

Fig. 4. Relocate the individuals if the new element lies outside the boundary of the
grid. Fig. 6. Group selection mechanism.

3.3.1.1. The archive controller. The main function of this controller where f is a constant number that should be greater than one and
is to decide whether the solution should be added or not in the S is the number of Pareto optimal solutions which are obtained
archive. There are some important points regarding the archive up- so far in the kth segment. The roulette-wheel selection method
date mechanism which are given below: is a classic selection operator in the proportional type. The fitness
• The solution should not be allowed to enter the archive when value of each individual in the population corresponds to the range
it is dominated by at least one of archive members. on the roulette wheel proportion. In Eq. (17), it can be seen that
• When a new solution dominates one or more members in the choosing the number of solutions is directly proportional to vector
M . The proposed MOSHO algorithm (see Algorithm 1) includes all
archive, the new solution will be able to enter the archive after
omitting the dominated solution(s) in the archive. the characteristics of SHO algorithm. The main difference between
• The new member should be added in the archive when neither MOSHO and SHO is that MOSHO searches in a set of archive mem-
the new solution nor archive solutions dominate each other. bers, while SHO saves the group of optimal solutions.
• The grid method should be run to omit one of the most The foremost difference between MOSHO and NSGA-2 is
crowded solution section and insert new solutions to improve archive. The motivation behind the use of archive is to reduce the
the diversity corresponding to Pareto optimal front when the possibility of deterioration of non-dominated solutions. MOSHO
archive is full. provides optimal solutions as compared to other archive-based al-
gorithms. The archive-based algorithms usually utilize selection
3.3.1.2. The grid. In this work, the adaptive grid mechanism is used operators such as crossover, mutation, etc. These operators bias
to produce the distributed Pareto fronts [48]. The objective func- the search towards the members of archive. In MOSHO, the vari-
tion space is divided into different regions as shown in Fig. 3. The ables are exchanged among a solution in the search space. This ex-
grid has to be recalculated and relocate each individual if the in- change operation increases the exploration capability of MOSHO.
serted individual into population lies outside the current bounds Whereas, it reduces the convergence property. To overcome this
of the grid [47] (see Fig. 4). The adaptive grid is a space formed by problem, MOSHO uses a group selection approach to select at least
hypercubes and is used to distribute in a uniform way. one non-dominated solution in the given search space.

3.3.1.3. Group selection mechanism. In multi-objective search space, 3.4. Computational complexity
the challenging task is to compare the solutions with archive mem-
bers. To overcome this issue a group selection mechanism is de- In this subsection, the computational complexity of proposed
signed. The group selection strategy chooses the least crowded sec- algorithm is discussed. The time and space complexities of MOSHO
tion of search space and offers one of its non-dominated solutions is described below.
to group of nearby solutions as shown in Fig. 6. The selection is
done on the basis of roulette-wheel method with probability which 3.4.1. Time complexity
is defined as: 1. Initialization of MOSHO population needs O (no × n p ) time
f where no represents the number of objectives and np represents
Hk = (18) the number of population size.
Sk
180 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Algorithm 1 Multi-objective Spotted Hyena Optimizer (MOSHO). 4. Experimental results and discussions
Input: Spotted hyenas population Pi (i = 1, 2, . . . , n)
This section compares the performance of MOSHO with six re-
Output: Archive of non-dominated optimal solutions
cently developed algorithms and validates over 24 benchmark test
1: procedure MOSHO
functions.
2: Initialize the vectors h, B, E, and N
3: Calculate the objective values of each search agent
4.1. Benchmark test functions
4: Find all the non-dominated solutions and initialize these solu-
tions to archive
The twenty four well-known benchmark test functions are
5: Ph = best search agent from archive
applied on MOSHO algorithm to demonstrate its efficiency. The
6: Ch = group or cluster of all far optimal solutions with respect
CEC-2009 special session benchmark test suit [24] is used which
to Ph (archive)
consists of ten unconstrained test functions as described in
7: while (x < Maxiterations ) do
Appendix A. These unconstrained functions are able to solve the
8: for each search agent do
bound constrainted multi-objective problems.
9: Update the position of current search agent by Eqs.
The popular five ZDT [25] and nine DTLZ [26] benchmark test
(13)-(17)
suites are also utilized which are described in the Appendices B
10: end for
and C, respectively. These unconstrained functions involve a feature
11: Update h, B, E, and N
to converge the Pareto optimal front very efficiently.
12: Calculate the objective values of all search agents
In Appendix A, the characteristics of CEC-2009 benchmark test
13: Find the non-dominated solutions from updated search
functions are described. There are ten unconstrained test functions
agents
(UF 1 − UF 10) which are included in this test suite.
14: Update the obtained non-dominated solutions to archive
The detailed characteristics of ZDT and DTLZ unconstrained test
15: if archive is full then
functions are described in Appendices B and C, respectively. There
16: Grid method should be run to omit one of the most
are five ZDT test functions (ZDT 1 − ZDT 6) and nine DTLZ test func-
crowded archive members
tions (DT LZ1 − DT LZ9).
17: Add new solution to the archive
18: end if
4.2. Experimental setup
19: Check if any search agent goes beyond the search space
and then adjust it
To validate the performance of MOSHO algorithm, it is
20: Calculate the objective values of each search agent
compared with six well-known optimization algorithms such
21: Update Ph if there is a better solution than the previous
as Multi-objective Particle Swarm Optimization (MOPSO) [21],
optimal solution from archive
Non-dominated Sorting Genetic Algorithm 2 (NSGA-2) [20],
22: Update the group Ch with respect to Ph (archive)
Multi-objective Evolutionary Algorithm based on Decomposition
23: x = x+1
(MOEA/D) [22], Pareto Envelope-based Selection Algorithm (PESA-
24: end while
2) [49], Strength Pareto Evolutionary Algorithm (SPEA-2) [28], and
25: return archive
Multi-objective Ant Colony Optimization (MOACO) [29].
26: end procedure
The parameter settings of these algorithms are set as they are
recommended in their original papers. The following initial param-
eters for MOPSO are chosen as [21]:
2. The fitness calculation of each search agent requires
O (Maxiterations × no × n p ) time where Maxiterations is the maxi- • φ a = φ b = 2.05
mum number of iterations to simulate the proposed MOSHO • φf = φa + φb
algorithm. 2
• Inertia weight: w = 
3. The algorithm requires O (M ) time to define the group of spot- φ f − 2 + φ 2f − 4φ f
ted hyenas where M indicates the counting value of spotted
hyenas.
• Personal coefficient: c1 = χ ∗ φa
4. It requires O (no × (nns + n p )) time to update the archive of
• Social coefficient: c2 = χ ∗ φb
non-dominated solutions.
• Grid inflation parameter: α = 0.1
5. Repeat Steps 2 to 4 until the termination criteria is satisfactory.
• Leader selection pressure parameter: β = 4
• Number of grids: Gridnumber = 10
Therefore, the overall time complexity of proposed MOSHO algo- For NSGA-2, the following initial parameters are chosen as [34]:
rithm is O (Maxiterations × no × (n p + nns ) × M ).
The computational complexities corresponding to MOPSO, • Population size (X) = 100
NSGA-2, PESA-2, SPEA-2, and MOACO are O (Maxiterations × no × • Cross over probability Pc = 0.8
n2p ), O (Maxiterations × no × n2p ), O (Maxiterations × no × (n p + nns )2 ), • Mutation probability Pm = 0.1
O (Maxiterations × no × n2p ), and O (Maxiterations × no × n2p ), respec- For MOEA/D, the following initial parameters are decided as
tively. Whereas, the computational complexity of MOEA/D is [22]:
O (Maxiterations × no × n p × T ), where T represents the Tchebycheff
approach to decompose the optimization problem into subprob- • Subproblems: N = 100
lems. • Number of neighbours: T = 0.1∗ N
• Updated new child maximal copies: M = 0.01∗ N
• Probability of selecting parents: Pp = 0.9
3.4.2. Space complexity • Mutation rates: Mr = 0.5
The space complexity of MOSHO algorithm is considered dur- • Distribution index: Di = 30
ing its initialization process which requires space at any one
time. Hence, the total space complexity of MOSHO algorithm is The following initial parameters are chosen for PESA-2 and
O ( no × n p ). SPEA-2 [34]:
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 181

Table 1
Archive state in different generations for ZDT1, DTLZ6, and ZDT3 test functions using MOSHO.

Iterations ZDT1 (Concave) DTLZ6 (Convex) ZDT3 (Disconnected)

Archive Objective value Archive Objective value Archive Objective value

x y f1 f2 x y f1 f2 x y f1 f2

1 0.999 0.015 0.861 0.937 0.256 3.616 0.648 0.382 0.999 0.017 0.459 0.045
0.987 0.156 0.005 5.989 0.987 0.170
0.955 0.294 0.005 5.989 0.956 0.290
0.904 0.425 0.030 5.337 0.907 0.421
0.835 0.549 0.030 5.337 0.842 0.535
0.750 0.661 0.005 5.989 0.753 0.655
0.649 0.760 0.090 4.112 0.643 0.751
0.535 0.844 0.030 5.337 0.436 0.840
0.411 0.911 0.256 3.616 0.407 0.900
0.263 0.964 0.030 5.337 0.269 0.960
50 0.842 0.010 0.762 0.753 0.136 2.025 0.590 0.233 0.825 0.011 0.301 0.040
0.862 0.140 0.023 3.234 0.762 0.267
0.823 0.118 0.007 4.023 0.882 0.243
0.843 0.346 0.039 4.237 0.752 0.376
0.705 0.413 0.039 3.007 0.755 0.341
0.639 0.519 0.085 4.462 0.697 0.528
0.556 0.431 0.128 3.096 0.763 0.428
0.448 0.746 0.025 3.039 0.340 0.774
0.401 0.831 0.223 2.219 0.386 0.841
0.248 0.763 0.010 4.414 0.202 0.741
100 0.018 0.900 0.801 0.566 0.999 0.157 0.568 0.967 0.004 0.934 0.284 0.914
0.117 0.670 0.987 0.156 0.032 0.793
0.215 0.541 0.955 0.294 0.055 0.708
0.316 0.442 0.904 0.425 0.082 0.669
0.418 0.359 0.835 0.549 0.209 0.479
0.519 0.285 0.750 0.661 0.240 0.278
0.612 0.218 0.649 0.760 0.415 0.164
0.714 0.157 0.535 0.844 0.439 –0.081
0.818 0.100 0.411 0.911 0.627 –0.273
0.916 0.045 0.263 0.964 0.826 –0.524

Fig. 7. Convergence analysis of proposed archive on (a) Concave, (b) Convex, and (c) Disconnected Pareto fronts.

• Cross over probability Pc = 0.8 Due to stochastic nature of these algorithms, the results are av-
• Distribution index for SBX = 15 eraged over 30 independent runs under 30 different random seeds.
• Mutation probability Pm = 1/n The mean best-of-run solution and standard deviation of best so-
• Polynomial mutation of distribution index = 20 lution in the last iteration are reported in tables.

For MOACO, the following initial parameters are chosen as [50]:

4.3. Performance metrics
• Initial pheromone = 1.0E-06
• Pheromone update constant Q = 20
The four well-known performance metrics are used to inves-
• Exploration constant q0 = 1
tigate the performance of above-mentioned algorithms. These are
• Global pheromone decay rate = 0.9
Hypervolume (HV) [51,52],  p ( p = 1 ) [53–55], Spread [52,56], and
• local pheromone decay rate = 0.5
Epsilon (ε ) [52,57].
• α = 0.5
The best Pareto optimal solutions set which is obtained by each
• β = 2.5
algorithm is compared with the proposed technique. The above-
The population size and maximum number of iterations for mentioned algorithms run 30 times on 24 benchmark test prob-
above-mentioned algorithms are set to 100 and 10 0 0, respectively. lems.
182 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Table 2
Results of the multi-objective algorithms on the CEC-2009 (UF1 - UF10) test functions.

F Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

UF1 Hypervolume 4.69E–02 7.30E–04 4.80E–01 2.04E–02 5.10E–01 4.36E–02 6.61E–01 2.80E–03 4.83E–01 5.97E–02 2.75E–01 1.20E–01 1.46E–01 7.81E–02
p 4.07E–09 8.11E–11 3.21E–04 1.11E–04 4.44E–03 1.35E–03 2.67E–04 1.04E–04 6.42E–03 2.73E–03 1.60E–02 1.69E–03 3.67E–03 2.91E–03
Spread 1.27E–01 3.12E–02 9.99E–01 2.13E–01 1.50E+00 1.24E–01 4.14E–01 1.85E–01 1.41E+00 1.65E–01 1.05E+00 1.05E–01 1.01E+00 2.03E–01
Epsilon 5.90E–03 1.11E–03 2.06E–01 3.10E–02 2.18E–01 5.50E–02 1.19E–02 1.17E–02 2.76E–01 1.04E–01 6.58E–01 1.69E–01 2.83E–01 1.20E–01
UF2 Hypervolume 1.07E–01 1.93E–03 6.12E–01 2.95E–03 6.10E–01 7.45E–03 6.51E–01 3.89E–03 5.94E–01 1.65E–02 5.27E–01 1.07E–01 4.16E–01 2.28E–01
p 4.13E–04 2.10E–04 1.51E–03 3.86E–05 6.37E–03 5.63E–04 3.45E–04 2.75E–04 5.63E–03 5.86E–04 3.21E–03 2.11E–03 3.10E–03 2.98E–03
Spread 4.09E–02 2.93E–02 4.98E–01 1.23E–01 6.04E–01 4.70E–02 5.43E–01 6.76E–02 8.86E–01 8.12E–02 1.04E+00 1.53E–01 2.88E–01 1.49E–01
Epsilon 2.05E–02 5.19E–04 9.23E–02 9.73E–03 1.41E–01 3.68E–02 6.85E–02 1.51E–02 1.79E–01 5.61E–02 3.35E–01 1.50E–01 4.71E–02 3.43E–03
UF3 Hypervolume 5.09E–01 3.70E–02 3.96E–01 3.51E–02 3.24E–01 4.40E–02 6.14E–01 2.24E–02 3.55E–01 3.36E–02 3.00E–01 3.75E–02 4.74E–01 1.71E–02
p 2.63E–03 1.61E–03 5.88E–03 4.31E–04 8.96E–03 1.60E–03 1.97E–03 5.01E–04 1.12E–02 1.89E–03 1.69E–02 1.99E–03 2.88E–02 1.40E–03
Spread 8.02E–01 6.99E–02 6.86E–01 1.05E–01 1.02E+00 5.47E–02 7.03E–01 2.04E–01 1.10E+00 1.03E–01 1.09E+00 9.40E–02 3.48E+00 2.77E–02
Epsilon 2.50E–01 4.11E–02 1.95E–01 2.10E–02 3.84E–01 5.94E–02 1.48E–01 6.18E–02 4.78E–01 6.32E–02 5.55E–01 8.06E–02 3.96E–01 4.48E–02
UF4 Hypervolume 1.00E–01 2.32E–03 2.40E–01 7.71E–03 2.46E–01 3.19E–03 2.37E–01 1.01E–02 2.40E–01 3.13E–03 1.52E–01 6.50E–02 2.41E–01 2.97E–02
p 4.15E–04 2.69E–05 2.66E–03 1.04E–04 1.31E–03 4.45E–05 3.89E–03 5.69E–04 2.39E–03 1.32E–04 5.60E–03 3.64E–03 4.99E–03 2.81E–04
Spread 1.00E–01 3.13E–02 4.01E–01 6.83E–02 5.15E–01 5.11E–02 5.02E–01 9.46E–02 8.24E–01 5.32E–02 1.09E+00 1.20E–01 2.74E–01 4.24E–02
Epsilon 1.10E–02 3.33E–03 7.48E–02 6.03E–03 6.99E–02 9.11E–03 7.80E–02 9.14E–03 8.39E–02 8.61E–03 3.63E–01 3.10E–01 4.71E–02 6.85E–03
UF5 Hypervolume 2.62E–03 1.23E–02 3.51E–03 1.57E–02 2.70E–02 4.06E–02 4.72E–02 5.74E–02 6.20E–02 6.01E–02 3.96E–02 4.77E–02 4.82E–02 2.09E–02
p 1.05E–01 1.08E–02 2.63E–01 1.50E–01 2.89E–01 3.89E–02 1.07E–01 2.91E–02 3.10E–01 1.38E–02 1.35E–01 2.21E–02 2.42E–01 3.85E–02
Spread 1.90E–01 4.78E–02 8.71E–01 9.50E–02 1.26E+00 9.95E–02 1.20E+00 5.02E–02 1.14E+00 1.34E–01 1.13E+00 8.40E–02 1.21E+00 5.72E–02
Epsilon 1.11E–01 1.00E–01 1.28E+00 5.28E–01 6.93E–01 1.89E–01 6.41E–01 1.43E–01 7.82E–01 1.76E–01 9.71E–01 2.21E–01 4.90E–01 3.47E–01
UF6 Hypervolume 1.60E–02 1.38E–02 2.02E–02 2.56E–02 1.96E–01 6.99E–02 2.08E–01 8.26E–02 2.09E–01 6.55E–02 1.15E–01 8.61E–02 2.87E–01 3.75E–02
p 4.31E–03 1.60E–03 2.56E–03 1.56E–03 6.78E–03 2.11E–03 4.14E–03 1.67E–03 1.94E–02 3.79E–03 2.21E–02 2.45E–03 1.71E–02 8.29E–03
Spread 7.09E–01 4.40E–02 9.79E–01 6.74E–02 1.19E+00 1.24E–01 1.14E+00 9.47E–02 1.27E+00 2.03E–01 1.10E+00 8.48E–02 2.11E+00 4.66E–02
Epsilon 2.02E–01 1.20E–01 5.95E–01 1.82E–01 4.22E–01 1.49E–01 3.55E–01 1.48E–01 5.48E–01 1.91E–01 8.70E–01 2.67E–01 5.39E–01 2.09E–01
UF7 Hypervolume 2.33E–01 6.36E–02 3.95E–01 3.57E–02 3.29E–01 8.89E–02 4.91E–01 3.34E–03 2.49E–01 8.82E–02 1.50E–01 8.80E–02 2.86E–01 3.99E–02
p 3.61E–03 1.45E–03 6.22E–03 2.67E–03 7.78E–03 2.06E–03 1.51E–04 2.44E–05 6.32E–02 2.71E–03 3.67E–02 3.30E–03 4.82E–02 1.75E–03
Spread 2.81E–01 2.17E–02 8.52E–01 6.63E–02 1.07E+00 2.11E–01 3.49E–01 1.75E–01 1.18E+00 8.77E–02 1.01E+00 2.58E–02 4.88E–01 2.67E–02
Epsilon 2.05E–02 1.78E–02 1.82E–01 1.07E–01 4.28E–01 2.42E–01 4.40E–02 3.17E–02 6.08E–01 2.19E–01 8.01E–01 1.37E–01 4.73E–01 2.81E–01
UF8 Hypervolume 1.60E–02 2.47E–02 3.05E–02 2.69E–02 1.16E–01 4.07E–02 1.81E–01 1.01E–01 1.93E–02 3.32E–02 1.29E–01 8.96E–02 1.00E–01 2.79E–02
p 3.34E–03 6.11E–04 2.31E–03 2.01E–04 2.81E–03 4.45E–04 3.16E–01 5.17E–03 5.11E–03 4.86E–04 4.87E–03 9.09E–04 5.95E–03 2.86E–03
Spread 6.07E–01 5.13E–01 8.27E–01 8.20E–02 7.77E–01 8.05E–02 4.69E–01 7.90E–01 7.07E–01 1.11E–01 8.62E–01 2.29E–01 5.79E–01 1.58E–01
Epsilon 8.03E–01 3.14E–01 7.69E–01 8.51E–02 6.24E–01 1.11E–01 7.01E–01 9.54E–01 9.70E–01 5.22E–02 8.05E–01 1.31E–01 6.97E–01 3.66E–01
UF9 Hypervolume 4.09E–02 6.33E–01 6.89E–02 5.70E–02 2.54E–01 6.17E–02 2.76E–02 4.47E–01 9.43E–02 3.51E–02 4.57E–01 4.78E–02 3.39E–01 2.11E–02
p 4.39E–03 6.10E–04 2.68E–03 3.27E-04 5.35E–03 2.51E–04 6.90E–03 4.48E–03 4.78E–03 3.31E–04 1.20E–03 1.04E–04 3.23E–03 1.76E–03
Spread 1.01E–01 7.37E–02 8.21E–01 5.05E–02 8.52E–01 6.83E–02 6.63E–01 6.41E–02 7.55E–01 7.45E–02 7.36E–01 6.10E–02 4.59E–01 3.46E–02
Epsilon 3.03E–01 5.15E–02 8.21E–01 1.42E–01 5.14E–01 6.46E–02 6.56E–01 5.77E–02 8.95E–01 4.16E–02 4.81E–01 2.92E–02 6.85E–01 5.68E–02
UF10 Hypervolume 4.60E–02 1.31E–02 0.0 0E+0 0 0.0 0E+0 0 4.53E–03 1.16E–02 6.90E–02 5.11E–01 4.05E–02 2.97E–02 3.38E–02 2.86E–02 4.90E–02 3.00E–02
p 6.78E–03 4.91E–03 3.79E–02 6.73E–03 3.13E–03 3.00E–03 4.01E–03 2.81E–03 1.83E–03 4.18E–04 6.50E–03 4.47E–04 4.22E–03 1.46E–03
Spread 2.62E–01 5.82E–02 6.62E–01 5.07E–02 7.66E–01 6.28E–02 5.21E–01 8.60E–01 9.18E–01 9.13E–02 1.24E+00 2.23E–01 4.53E–01 1.86E–01
Epsilon 4.61E–01 5.53E–01 1.82E+00 2.45E–01 1.05E+00 1.55E–01 5.33E–01 7.17E–01 8.86E–01 1.07E–01 8.68E–01 1.03E–01 6.00E–01 2.85E-01

Table 3
Results of the multi-objective algorithms on the ZDT benchmark (ZDT1 – ZDT6) test functions.

F Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

ZDT1 Hypervolume 6.56E–01 3.82E–04 6.62E–01 1.40E–04 6.60E–01 2.62E–04 6.64E–01 3.56E–04 6.56E–01 8.58E–04 6.60E–01 3.05E–04 6.60E–01 2.46E–04
p 3.34E–04 4.04E–05 2.33E–04 5.00E–07 1.18E–04 3.37E–06 6.32E–05 2.44E–06 4.03E–04 1.44E–04 2.11E–04 3.21E–06 4.66E–04 2.00E–06
Spread 2.30E–02 1.11E–03 6.87E–02 1.04E–02 3.77E–01 2.58E–02 2.85E–01 2.82E–03 6.38E–01 3.78E–02 1.43E–01 1.60E–02 2.31E–01 2.20E–02
Epsilon 5.08E–02 6.16E–02 5.52E–03 1.58E–04 1.34E–02 2.00E–03 3.01E–03 3.82E–04 2.58E–02 1.18E–02 8.90E–03 6.68E–04 7.08E–02 3.41E–04
ZDT2 Hypervolume 3.29E–01 3.71E–05 3.28E–01 8.09E–05 3.26E–01 3.11E–04 3.31E–01 1.84E–04 3.24E–01 6.54E–04 3.26E–01 6.14E–04 3.27E–01 2.09E–04
p 2.01E–05 1.34E–06 1.45E–04 2.57E–06 1.37E–04 3.41E–05 4.98E–05 6.00E–06 3.52E–04 2.35E–05 2.22E–04 4.21E–06 1.30E–04 2.76E–06
Spread 2.02E–01 9.10E–02 6.68E–02 7.85E–03 3.70E–01 4.46E–02 1.45E–01 1.21E–02 6.76E–01 4.20E–02 1.49E–01 1.46E–02 2.56E–01 3.87E–02
Epsilon 1.20E–03 2.19E–04 5.46E–03 1.37E–04 1.27E–02 2.30E–03 2.68E–03 3.62E–04 3.29E–02 1.88E–02 8.49E–03 6.16E–04 3.07E–03 8.73E–04
ZDT3 Hypervolume 5.11E–01 2.22E–04 5.15E–01 3.87E–04 5.15E–01 1.58E–04 5.16E–01 3.09E–05 5.10E–01 2.80E–03 5.14E–01 2.31E–04 5.13E–01 5.80E–04
p 4.13E–05 3.31E–06 2.57E–04 2.08E–06 1.33E–04 4.49E–06 4.14E–05 5.55E–06 1.00E–03 5.29E–04 3.21E–04 2.88E–05 2.40E–04 4.99E–05
Spread 1.00E–01 2.17E–03 7.05E–01 3.38E–03 7.42E–01 1.32E–02 1.03E+00 2.47E–03 9.02E–01 2.62E–02 7.09E–01 4.74E–03 6.06E–01 2.53E–03
Epsilon 4.91E–03 6.74E–04 5.71E–03 5.49E–04 8.87E–03 1.56E–03 5.33E–03 1.27E–04 1.30E–01 1.54E–01 9.60E–03 1.14E–03 5.94E–03 2.48E–03
ZDT4 Hypervolume 6.60E–01 4.44E–04 0.0 0E+0 0 0.0 0E+0 0 6.54E–01 3.53E–03 6.65E–01 4.95E–05 6.50E–01 4.64E–03 6.49E–01 8.60E–03 6.45E–01 3.46E–02
p 4.30E–04 1.80E–04 3.41E–01 2.38E–02 3.78E–04 5.99E–05 4.18E–05 3.04E–05 3.10E–04 1.10E–04 4.71E–03 1.20E–03 1.69E–04 1.00E–04
Spread 2.02E–01 1.98E–03 8.86E–01 7.18E–02 3.98E–01 3.61E–02 2.85E–01 2.00E–03 9.10E–01 2.88E–01 3.06E–01 1.36E–01 2.01E–01 1.07E–02
Epsilon 2.09E–03 1.00E–04 5.14E+00 1.91E+00 1.76E–02 8.12E–03 2.94E–03 1.03E–04 2.87E–02 1.06E–02 6.37E–02 5.29E–02 3.28E–02 2.02E–02
ZDT6 Hypervolume 3.19E–01 4.33E–05 4.01E–01 5.40E–05 3.88E–01 1.46E–03 4.05E–01 6.11E–08 3.91E–01 1.34E–03 3.78E–01 2.70E–03 2.26E–01 1.86E–03
p 2.70E–04 4.11E–05 2.31E–04 2.09E–06 4.81E–04 7.07E–05 3.61E–05 6.41E–08 3.07E–04 3.80E–05 4.95E–04 3.37E–05 3.27E–04 1.30E–05
Spread 1.11E–01 1.24E–04 8.48E–01 4.79E–01 3.66E–01 3.99E–02 1.50E–01 1.86E–04 7.59E–01 1.93E–01 2.38E–01 3.08E–02 3.47E–01 1.94E–02
Epsilon 1.20E–03 3.13E–04 4.86E–03 5.16E–04 1.58E–02 2.06E–03 1.64E–03 1.43E–01 1.89E–02 3.66E–03 2.56E–02 5.35E–06 3.00E–02 2.17E–04
G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 183

Fig. 8. Flowchart of the proposed MOSHO algorithm.

184 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Fig. 9. Convergence analysis of selection approaches on (a) ZDT1, (b) ZDT3, and (c) ZDT6 test functions.

Table 4
Results of the multi-objective algorithms on the DTLZ benchmark (DTLZ1 - DTLZ9) test functions.

F Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

DTLZ1 Hypervolume 4.60E–01 4.33E–02 3.63E–01 5.56E–02 6.62E–01 3.95E–01 3.56E–01 5.94E–03 5.93E–01 3.73E–02 7.56E–01 1.97E–02 4.11E–01 2.00E–02
p 3.45E–04 1.11E–05 2.53E–03 2.49E–04 3.78E–03 3.81E–04 4.24E–04 5.40E–05 5.00E–03 2.00E–04 4.46E–04 7.21E–06 5.79E–04 4.11E–04
Spread 1.33E–01 3.17E–01 5.95E–01 2.93E–01 2.02E+00 2.66E–01 2.85E–01 4.24E–01 3.72E+00 3.89E–01 9.15E–01 2.97E–01 4.97E–01 1.08E–01
Epsilon 3.60E–03 5.83E–02 3.96E–02 3.93E–02 2.10E–02 5.23E–01 4.35E–02 3.73E–02 2.04E–03 2.67E–01 5.39E–02 1.49E–02 2.07E–02 2.00E–02
DTLZ2 Hypervolume 1.09E–02 1.07E–03 1.94E–01 2.63E–03 2.63E–02 3.96E–03 3.08E–01 5.23E–02 5.70E–02 1.12E–02 4.04E–01 1.87E–03 3.97E–01 2.05E–03
p 3.54E–04 3.11E–05 3.48E–03 8.06E–04 3.67E–04 2.00E–05 6.90E–04 6.57E–05 3.19E–02 5.39E–03 4.73E–04 4.11E–05 4.95E–04 4.06E–05
Spread 3.30E–01 5.78E–03 6.71E–01 2.73E–01 4.83E–01 7.54E–03 7.40E–01 3.30E–02 7.56E–01 5.73E–01 5.46E–01 3.05E–02 4.83E–01 6.28E–02
Epsilon 3.69E–02 6.16E–02 7.73E–02 8.33E–03 2.37E–01 8.08E–02 5.05E–02 3.57E–01 3.73E–01 4.73E–01 8.41E–02 1.13E–02 6.05E–02 4.74E–02
DTLZ3 Hypervolume 2.09E–02 8.37E–03 3.63E–01 2.63E–02 6.24E–02 5.25E–04 3.74E–01 2.74E–02 5.23E–01 2.35E–03 0.0 0E+0 0 0.0 0E+0 0 4.03E–01 4.46E–03
p 3.47E–04 2.10E–04 5.22E–03 3.14E–03 5.05E–04 3.33E–04 2.49E–03 6.41E–04 3.03E–01 3.90E–02 1.00E–01 4.59E–02 3.04E–02 2.41E–02
Spread 1.39E–02 3.33E–02 5.73E–01 3.73E–01 1.32E+00 6.34E–02 4.63E–01 2.63E–02 2.43E–02 4.03E–01 1.20E+00 1.09E–01 2.92E–01 1.63E–01
Epsilon 2.09E–02 3.56E–02 3.74E–01 4.38E–02 5.73E–01 5.77E–02 3.84E–02 6.63E–02 3.22E–01 8.74E–02 6.37E+00 2.70E+00 3.20E–01 2.83E–02
DTLZ4 Hypervolume 2.60E–02 1.31E–02 2.74E–01 3.73E–03 3.45E–02 6.89E–02 2.74E–01 3.63E–02 4.53E–03 5.23E–03 3.12E–01 9.77E–02 4.96E–02 2.38E–02
p 2.33E–04 4.11E–05 4.61E–02 2.00E–04 3.69E–02 5.31E–03 2.14E–04 1.09E–04 4.79E–02 6.49E–03 4.35E–03 3.20E–03 3.16E–02 4.92E–03
Spread 2.78E–02 5.11E–02 3.08E–01 3.53E–02 5.73E–02 6.63E–01 6.64E–01 3.21E–02 4.23E–03 6.66E–02 4.68E–01 1.29E–01 5.05E–02 4.80E–02
Epsilon 5.05E–02 8.70E–02 5.13E–02 6.63E–02 6.73E–03 9.14E–03 3.63E–02 9.74E–02 3.53E–01 8.33E–03 3.23E–01 2.86E–01 4.60E–02 3.94E–02
DTLZ5 Hypervolume 1.49E–03 1.32E–04 1.63E–03 2.83E–04 3.13E–02 5.06E–02 6.05E–01 3.83E–02 5.93E–02 5.23E–02 1.96E–01 2.47E–02 2.42E–02 1.99E–02
p 1.46E–03 2.81E–04 1.35E–01 7.36E–02 3.21E–01 1.73E–02 5.64E–03 6.88E–04 4.73E–02 2.32E–02 3.77E–01 2.31E–02 2.90E–02 1.04E–03
Spread 3.88E–01 5.00E–02 5.63E–01 5.73E–02 1.36E+00 7.96E–03 6.27E–01 5.22E–02 1.04E+00 4.14E–04 2.43E–02 5.47E–02 8.33E–01 4.29E–02
Epsilon 2.03E–01 2.78E–01 3.73E+00 5.28E–01 5.23E–01 1.69E–01 6.94E–01 5.92E–01 5.82E–01 1.02E–01 7.75E–01 3.20E–01 4.06E–01 1.95E–01
DTLZ6 Hypervolume 2.63E–03 1.11E–03 3.02E–03 2.56E–02 1.92E–01 7.83E–02 5.04E–02 3.16E–03 3.09E–01 5.55E–02 1.05E–01 6.61E–02 2.53E–02 2.00E–02
p 3.00E–03 2.67E–03 2.33E–02 3.66E–03 3.50E–02 5.12E–03 5.21E–03 2.51E–03 3.67E–02 6.36E–03 3.11E–02 3.10E–03 2.10E–02 4.96E–03
Spread 3.45E–01 6.77E–02 6.09E–01 3.32E–02 1.53E+00 1.94E–01 4.10E–01 5.07E–01 1.26E–01 2.63E–01 2.10E+00 7.18E–02 3.53E–01 4.75E–02
Epsilon 2.47E–03 7.16E–03 6.95E–01 1.82E–03 4.63E–03 3.63E–01 3.65E–02 3.48E–04 3.43E–01 6.36E–01 7.72E–01 2.73E–01 4.34E–02 6.80E–03
DTLZ7 Hypervolume 2.09E–02 1.32E–02 3.53E–01 3.63E–02 6.29E–03 3.09E–04 4.11E–01 3.83E–03 3.49E–03 5.82E–01 2.51E–01 8.38E–02 5.03E–02 7.72E–03
p 4.10E–03 3.77E–04 5.50E–03 1.13E–03 6.98E–03 3.48E–04 3.38E–03 3.27E–03 4.91E–02 6.94E–03 1.89E–02 2.81E–03 6.42E–03 3.92E–03
Spread 4.67E–02 7.17E–02 6.31E–01 8.04E–02 1.83E+00 2.53E–03 4.35E–01 3.37E–01 3.73E–03 3.27E–04 1.03E–01 3.91E–02 2.28E–02 5.40E–03
Epsilon 7.78E–02 5.18E–02 8.63E–01 1.07E–01 3.28E–02 3.74E–04 3.63E–02 3.35E–02 3.08E–01 3.33E–02 4.04E–01 1.93E–01 2.28E–01 4.32E–02
DTLZ8 Hypervolume 7.60E–03 2.47E–02 4.50E–03 2.69E–04 1.84E–01 3.94E–02 4.88E–03 2.01E–02 2.03E–02 4.23E–02 2.98E–01 6.93E–02 1.39E–01 2.76E–03
p 6.46E–03 2.81E–04 3.18E–04 1.00E–03 5.03E–03 2.68E–03 4.75E–02 2.30E–03 4.69E–03 6.68E–04 3.11E–03 1.47E–03 4.53E–03 8.95E–04
Spread 3.87E–02 7.77E–02 3.27E–01 3.20E–02 2.37E–03 8.53E–03 3.39E–01 3.90E–01 5.07E–02 3.14E–02 5.02E–02 2.99E–02 1.37E–02 4.05E–03
Epsilon 1.01E–02 6.19E–02 3.69E–01 3.51E–02 6.25E–03 1.24E–03 3.01E–01 4.23E–01 3.24E–02 2.22E–02 6.09E–01 1.82E–01 2.23E–02 6.97E–03
DTLZ9 Hypervolume 1.60E–03 2.36E–02 6.89E–02 5.70E–02 2.35E–03 6.17E–02 3.76E–02 4.47E–01 5.43E–02 3.51E–02 3.83E–01 2.94E–02 1.98E–01 8.76E–02
p 1.47E–03 3.62E–04 5.47E–03 3.58E–03 6.77E–02 4.33E–03 3.24E–03 2.87E–03 6.89E–03 4.30E–04 5.47E–03 3.91E–04 7.04E–03 9.43E–04
Spread 9.87E–02 3.78E–02 4.21E–01 8.88E–02 4.52E–03 3.83E–01 6.76E–01 6.31E–02 3.55E–02 8.45E–03 4.16E–01 4.94E–01 5.09E–02 9.81E–03
Epsilon 9.58E–02 4.99E–02 4.21E–01 8.42E–01 7.73E–03 6.93E–03 6.06E–01 7.77E–02 7.67E–02 6.96E–01 3.41E–02 2.44E–02 4.38E–02 3.79E–02

4.4. Performance evaluation UF5, and UF6 test functions. For UF3 test function, MOPSO pro-
vides better value of Spread than the others. MOEA/D provides
This subsection discuss the results of proposed MOSHO algo- better results than the other competitive algorithms in terms of
rithm on CEC-2009, ZDT, and DTLZ benchmark test suites. p and Epsilon. For UF7, UF9, and UF10 test functions, MOSHO
provides competitive results in comparison with the others. For
4.4.1. Results on CEC-2009 (UF 1 − UF 10) benchmark test functions UF8 test function, NSGA-2 gives the better value of p and Ep-
Table 2 shows the performance comparison of the proposed silon as compared with other algorithms. The Pareto optimal so-
MOSHO with the above-mentioned existing techniques for CEC- lutions of MOSHO on UF benchmark problems are illustrated in
2009 benchmark test problems. The results reveal that MOSHO Fig. 10.
is able to determine the optimal solutions for UF1, UF2, UF4,
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 185

Fig. 10. Best Pareto optimal solutions obtained by the proposed MOSHO algorithm on CEC–2009 benchmark test problems.

4.4.2. Results on ZDT (ZDT 1 − ZDT 6) benchmark test functions form poorer results on ZDT2 test problem. MOSHO is able to pro-
The performance comparison of MOSHO with various algo- vide efficient results on ZDT2, ZDT3, and ZDT6 test problems while
rithms on ZDT benchmark test problems are tabulated in Table 3. It all of the other algorithms fail to achieve Pareto optimal solutions.
has been observed from table that MOSHO outperforms the other Fig. 12 shows the boxplot of MOSHO is considerably narrower than
algorithms on the majority of ZDT test functions in terms of p , those of MOPSO, NSGA-2, MOEA/D, PESA-2, SPEA-2, and MOACO.
Spread, and Epsilon. The best Pareto optimal solutions obtained
from the above-mentioned algorithms is shown in Fig. 11. This fig- 4.4.3. Results on DTLZ (DT LZ1 − DT LZ9) benchmark test functions
ure illustrates that the MOSHO provides better true Pareto optimal The statistical results of MOPSO, NSGA-2, MOEA/D, PESA-2,
front. It is also observed that NSGA-2, MOEA/D, and SPEA-2 per- SPEA-2, and MOACO on DTLZ benchmark problems are shown in
186 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Fig. 11. Best Pareto optimal solutions obtained by the multi–objective algorithms on (a) ZDT1, (b) ZDT2, (c) ZD3, (d) ZDT4, and (e) ZDT6 test problems.

Table 4. For DTLZ1, DTLZ2, DTLZ3, DTLZ5, DTLZ6, and DTLZ9 test problems has been fixed at 10. The detailed convergence analysis is
functions, MOSHO provides better statistical results than the oth- depicted in Fig. 7. It has been observed that the proposed MOSHO
ers for most of the performance measures. For DTLZ4 test func- achieves optimal values of these test functions during the course
tion, MOEA/D and NSGA-2 give better value of p and Epsilon of iterations.
over other algorithms. PESA-2 provides better value of Hypervol-
ume and Spread than the others. 4.5.2. Effect of selection mechanism on proposed algorithm
The results show that the proposed algorithm has better con- The roulette wheel and tournament selection approaches are
vergence ability for most of the DTLZ benchmark test functions. used to analyze the performance of proposed MOSHO. For experi-
Fig. 13 shows the shapes of the obtained Pareto optimal fronts of mentation, ZDT1, ZDT3, and ZDT6 test functions are taken into the
MOSHO which are closer to the true Pareto optimal front. However, consideration. Fig. 9 shows the convergence analysis of roulette
these results prove that the durability of the proposed MOSHO al- wheel and tournament selection approaches. From Fig. 9, it can
gorithm is considerably high on these unconstrained benchmark be concluded that the convergence of roulette wheel selection ap-
test problems. Therefore, MOSHO shows superior convergence on proach is better than the tournament selection approach. It is also
all of the above benchmark test problems. observed that the roulette wheel selection approach is able to con-
verge towards the optimal solutions.
4.5. Sensitivity analysis
5. MOSHO for engineering design problems
4.5.1. Effect of archive on proposed algorithm
The behavior of archive during consecutive generations is il- In order to investigate the performance of proposed MOSHO,
lustrated in Table 1. To show the effect of archive on proposed it has been applied on six real-life constrained engineering design
MOSHO, it has been tested on ZDT1, DTLZ6, and ZDT3 benchmark problems such as welded beam design, multiple-disk clutch brake
test functions. ZDT1, DTLZ6, and ZDT3 have concave, convex, and design, pressure vessel design, speed reducer design, gear train de-
disconnected properties respectively. The archive size of these test sign, and 25-bar truss design. There are different types of penalty
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 187

Fig. 12. Boxplot of the statistical results for Hypervolume, p , Spread, and Epsilon on ZDT benchmark test problems.

functions to handle these multi-constraint problems such as Static rithm is equipped with death penalty function to handle multiple
penalty, Dynamic penalty, Annealing penalty, Adaptive penalty, Co- constraints.
evolutionary penalty, and Death penalty [58]. The proposed MOSHO is compared with six well-known al-
However, death penalty function is used to discard the infeasi- gorithms such as MOPSO, NSGA-2, MOEA/D, PESA-2, SPEA-2, and
ble solutions and does not employ the information of such solu- MOACO. The algorithm’s parameter settings are same as used in
tions which are helpful to solve the dominated infeasible regions. Section 4.2.
Due to low computational cost and its simplicity, MOSHO algo-
188 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Fig. 13. Best Pareto optimal solutions obtained by the proposed MOSHO algorithm on DTLZ benchmark test problems.

Table 5
Comparison of best solution obtained from different algorithms for welded beam design problem.

Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

Hypervolume 5.18E–01 3.35E–01 8.09E–01 6.93E–01 7.17E–01 5.34E–01 7.94E–01 6.97E–01 9.93E–01 8.96E–01 6.40E–01 4.87E–01 8.59E–01 5.70E–01
p 3.71E–02 5.07E–03 2.47E–01 6.55E–02 1.69E–01 2.87E–02 4.45E–02 2.30E–02 6.84E–01 6.87E–02 7.92E–02 5.50E–02 4.34E–02 3.48E–02
Spread 1.37E–01 5.96E–02 2.14E–01 8.29E–02 9.90E–01 1.78E–01 1.78E+00 1.01E+00 1.81E–01 7.20E–02 6.59E–01 2.18E–01 3.66E–01 2.97E–01
Epsilon 9.64E–03 2.41E–03 1.04E–01 9.64E–02 9.64E–02 1.36E–02 4.22E–02 9.75E–03 1.15E–01 9.63E–02 1.52E–02 8.76E–03 4.68E–02 3.05E–03

5.1. Welded beam design • The first constraint confirms that the shear stress of beam is
smaller than the shear strength of material.
The main objective of this problem is to minimize the fabri- • The normal stress is smaller than the strength of the material
cation cost and simultaneously minimize the vertical deflection at (i.e., 30,0 0 0 psi).
the end of the welded beam as shown in Fig. 14. There are four • The third constraint confirms that the thickness of beam is not
optimization variables of this problem. smaller than the thickness of weld.
• The fourth constraint confirms that the buckling load is more
• Thickness of weld (h).
than the applied load of beam.
• Length of the clamped bar (l).
• Height of the bar (t).
• Thickness of the bar (b).

The constraints of welded beam problem are: The mathematical formulation of this problem is given in
Appendix D. Table 5 shows the comparisons of different algorithms
• Shear stress (τ ) and bending stress (θ ) in the beam.
for the best obtained solutions. The results reveal that MOSHO per-
• Buckling load (Pc ) on the bar.
forms better than the other existing algorithms.
• End deflection (δ ) of the beam.
By observing Fig. 15, MOSHO achieves the near optimal solu-
There are some important points that are considered during the tion in the initial steps of iterations and durability of the proposed
optimization of problem. MOSHO algorithm is high on this constrained engineering problem.
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 189

Table 6
Comparison of best solution obtained from different algorithms for multiple-disk clutch brake design problem.

Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

Hypervolume 4.53E–01 2.49E–01 7.86E–01 5.32E–01 7.19E–01 5.63E–01 9.64E–01 7.63E–01 9.03E–01 8.14E–01 9.64E–01 7.96E–01 6.40E–01 4.02E–01
p 2.73E–02 1.48E–03 5.16E–02 5.99E–03 2.13E–01 2.52E–02 7.43E–02 6.03E–02 2.21E–01 9.91E–02 9.72E–02 3.17E–02 5.13E–02 4.52E–02
Spread 1.96E–01 1.10E–01 2.36E–01 1.35E–01 8.64E–01 5.24E–01 1.54E+00 1.35E+00 9.52E–01 5.21E–01 8.30E–01 4.40E–01 4.42E–01 2.38E–01
Epsilon 1.76E–02 5.24E–03 1.14E–01 9.84E–02 7.52E–02 2.44E–02 1.39E–01 8.32E–02 1.10E–01 7.30E–02 2.93E–02 1.31E–02 2.08E–02 1.58E–02

Table 7
Comparison of best solution obtained from different algorithms for pressure vessel design problem.

Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

Hypervolume 3.03E–01 2.00E–01 7.30E–01 4.23E–01 6.25E–01 4.51E–01 8.95E–01 7.13E–01 8.36E–01 7.01E–01 9.60E–01 8.95E–01 4.00E–01 3.79E–01
p 2.67E–03 1.00E–03 6.16E–01 2.19E–02 4.31E–02 4.45E–03 2.59E–01 5.98E–02 4.00E–02 3.22E–02 3.67E–01 9.90E–02 2.28E–02 1.63E–02
Spread 1.77E–01 1.20E–01 2.01E–01 1.56E–01 5.21E–01 4.03E–01 1.44E+00 1.04E+00 8.45E–01 6.38E–01 9.63E–01 6.35E–01 3.07E–01 2.89E–01
Epsilon 1.19E–02 4.21E–03 1.04E–01 8.43E–02 9.52E–02 7.84E–02 1.54E–01 1.35E–01 1.40E–01 1.04E–01 3.00E–02 1.93E–02 2.58E–02 2.07E–02

Fig. 14. Schematic view of welded beam problem.

Fig. 16. Schematic view of multiple-disk clutch brake design problem.

Fig. 15. Best Pareto optimal front obtained by MOSHO, MOPSO, and NSGA-2 on
welded beam design problem.

5.2. Multiple-disk clutch brake design problem

The objective of this problem is to minimize the stopping time

(f1 ) as well as minimize the mass of the brake system (f2 ) as
shown in Fig. 16. There are five decision variables of this problem.
Fig. 17. Best Pareto optimal front obtained by MOSHO, MOPSO, and NSGA-2 on
• Inner radius in mm (ri ). multiple-disk clutch brake design problem.
• Outer radius in mm (ro ).
190 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Fig. 18. Schematic view of pressure vessel problem.

Fig. 21. Best Pareto optimal front obtained by MOSHO, MOPSO, and NSGA-2 on
speed reducer problem.

Fig. 22. Schematic view of gear train problem.

Fig. 19. Best Pareto optimal front obtained by MOSHO, MOPSO, and NSGA-2 on
pressure vessel problem.

Fig. 20. Schematic view of speed reducer problem.

• Thickness of discs in mm (t).

• Actuating force (F).
• Number of friction surfaces (Z).
Fig. 23. Best Pareto optimal front obtained by MOSHO, MOPSO, and NSGA-2 on
gear train design problem.
The statistical results of best optimal solution among several al-
gorithms are given in Table 6. Among these algorithms, the pro-
posed MOSHO gives optimal value of decision variables. It is worth 5.3. Pressure vessel design
mentioning that MOSHO provides better statistical results in terms
of Hypervolume, p , Spread, and Epsilon. This problem was proposed by Kannan and Kramer [59] to min-
Fig. 17 shows the best Pareto optimal fronts obtained from imize the total cost (f1 ) including cost of material, forming and
MOSHO, MOPSO, and NSGA-2. From Fig. 17, it can be concluded welding of cylindrical vessel which are capped at both ends by
that the performances of MOSHO, MOPSO, and NSGA-2 are com- hemispherical heads, and to maximize the storage capacity (f2 ) as
parable to each other. shown in Fig. 18.
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 191

Table 8
Comparison of best solution obtained from different algorithms for speed reducer design problem.

Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

Hypervolume 3.22E–01 2.20E–01 6.97E–01 4.35E–01 6.00E–01 5.62E–01 8.63E–01 6.69E–01 9.95E–01 8.03E–01 8.64E–01 6.14E–01 5.11E–01 4.00E–01
p 4.13E–03 1.57E–03 5.07E–02 4.84E–03 3.37E–02 3.03E–02 7.68E–02 2.00E–02 5.00E–01 8.93E–02 6.42E–03 4.66E–03 5.99E–02 3.57E–03
Spread 1.69E–01 1.22E–01 3.36E–01 2.15E–01 7.63E–01 4.32E–01 1.97E+00 1.45E+00 4.52E–01 6.32E–01 6.95E–01 4.44E–01 3.11E–01 2.52E–01
Epsilon 1.00E–02 5.09E–03 1.09E–01 1.02E–01 1.39E–01 8.62E–02 1.96E–01 7.63E–02 6.53E–02 4.25E–02 1.34E–01 1.96E–01 2.75E–01 4.54E–02

Table 9
Comparison of best solution obtained from different algorithms for gear train design problem.

Performance MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Metrics Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

Hypervolume 3.64E–01 2.86E–01 9.53E–01 6.96E–01 6.11E–01 5.60E–01 6.39E–01 7.97E–01 8.62E–01 7.14E–01 9.86E–01 8.91E–01 6.48E–01 4.34E–01
p 2.62E–02 1.41E–02 5.40E–01 2.89E–02 4.58E–02 5.65E–03 6.79E–01 2.33E–02 4.45E–01 4.08E–02 7.00E–02 4.66E–02 4.50E–02 3.32E–02
Spread 1.26E–01 7.63E–02 2.04E–01 1.03E–01 5.39E–01 3.12E–01 1.25E+00 1.02E+00 7.93E–01 5.00E–01 9.64E–01 5.32E–01 5.20E–01 3.47E–01
Epsilon 3.45E–03 1.51E–02 1.11E–01 7.62E–02 1.30E–01 1.04E–01 1.20E–01 9.65E–02 9.46E–02 3.49E–02 1.27E–01 1.25E–01 4.62E–02 2.22E–02

Table 10
Member stress limitations for the 25-bar truss design problem.

Element group Compressive stress limitations Ksi (MPa) Tensile stress limitations Ksi (MPa)

Group 1 35.092 (241.96) 40.0 (275.80)

Group 2 11.590 (79.913) 40.0 (275.80)
Group 3 17.305 (119.31) 40.0 (275.80)
Group 4 35.092 (241.96) 40.0 (275.80)
Group 5 35.092 (241.96) 40.0 (275.80)
Group 6 6.759 (46.603) 40.0 (275.80)
Group 7 6.959 (47.982) 40.0 (275.80)
Group 8 11.082 (76.410) 40.0 (275.80)

Fig. 24. Schematic view of 25-bar truss design problem.

There are four design variables (x1 − x4 ) of this problem. MOSHO is able to find optimal design with minimum cost and
maximum storage.
• Ts (x1 , thickness of the shell).
The shape of Pareto optimal fronts obtained from MOSHO,
• Th (x2 , thickness of the head).
MOPSO, and NSGA-2 are shown in Fig. 19. MOSHO is also able to
• R (x3 , inner radius).
find a good distribution of points similar to that of NSGA-2 and
• L (x4 , length of the cylindrical section without considering the
MOPSO.
head).

where R and L are continuous variables, while Ts and Th are in- 5.4. Speed reducer design problem
teger values which are multiples of 0.0625 in. The mathematical
formulation of this problem is formulated in Appendix D. The design of speed reducer is a most challenging test problem
Table 7 shows the comparisons of best optimal solution for which is associated with seven design variables [60]. In this op-
MOSHO and other reported algorithms. According to this table, timization problem (see Fig. 20), the weight of speed reducer (f1 )
192 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

Table 11
Two loading conditions for the 25-bar truss design problem.

Node Case 1 Case 2

Px Kips(kN) Py Kips(kN) Pz Kips(kN) Px Kips(kN) Py Kips(kN) Pz Kips(kN)

1 0.0 20.0 (89) –5.0 (22.25) 1.0 (4.45) 10.0 (44.5) –5.0 (22.25)
2 0.0 –20.0 (89) –5.0 (22.25) 0.0 10.0 (44.5) –5.0 (22.25)
3 0.0 0.0 0.0 0.5 (2.22) 0.0 0.0
6 0.0 0.0 0.0 0.5 (2.22) 0.0 0.0

Table 12
Comparison of best solution obtained from different algorithms for 25-bar truss design problem.

Performance Metrics MOSHO MOPSO NSGA-2 MOEA/D PESA-2 SPEA-2 MOACO

Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std Ave Std

Hypervolume 2.14E–01 1.16E–01 8.59E–01 5.79E–01 4.35E–01 2.42E–01 7.48E–01 5.00E–01 4.05E–01 2.71E–01 6.76E–01 5.85E–01 3.05E-01 2.31E-01
p 3.22E–02 1.42E–02 2.56E–01 1.72E–01 3.79E–02 2.47E–02 2.00E–01 1.34E–01 2.68E–01 4.00E–02 3.27E–02 2.86E–02 5.38E–02 3.59E–02
Spread 2.17E–01 4.33E–02 3.59E–01 3.43E–01 4.30E–01 3.27E–01 2.49E–01 1.19E–01 5.07E–01 4.04E–01 5.38E–01 2.86E–01 3.34E-01 2.02E-01
Epsilon 2.88E–02 1.72E–02 1.91E–01 3.00E–02 2.80E–01 1.44E–01 1.59E–01 2.51E–02 3.87E–02 2.76E–02 2.42E–01 2.00E–01 3.66E-01 2.58E–01

The best Pareto optimal solutions obtained from MOSHO,

MOPSO, and NSGA-2 algorithms are shown in Fig. 21. MOSHO ob-
tains best Pareto optimal solution for speed reducer problem dur-
ing course of iterations. The MOSHO algorithm achieves better re-
sults and verified by finding a set of Pareto solutions.

5.5. Gear train design problem

The objective of this problem is to find the number of teeth

for each four gears so as to minimize the error between obtained
and required gear ratio and minimize the size of four gears as
shown in Fig. 22. There are four integer decision variables of this
problem.

• Number of teeth on the driving gear (Td ).

• Number of teeth on the driven gear (Ta ).
• Number of teeth on the gear attached to the driven gear (Tb ).
• Number of teeth on the following gear (Tf ).

The mathematical formulation of this problem is formulated in

Appendix D.
Fig. 25. Best Pareto optimal front obtained by MOSHO, MOPSO, and NSGA-2 on 25-
Table 9 shows the comparison for best solution given by
bar truss design problem.
MOSHO and other algorithms. The proposed algorithm provides
best optimal solution over the other optimizers.
The best Pareto optimal fronts obtained from MOSHO, MOPSO,
and stress of a speed reducer (f2 ) should be minimized with eleven
and NSGA-2 are shown in Fig. 23. MOSHO is able to find a set of
constraints. There are seven optimization variables (x1 − x7 ) of this
well-distributed points on the entire Pareto optimal front.
problem.

• Face width (b). 5.6. 25-Bar truss design

• Module of teeth (m).
• Number of teeth in the pinion (z). The truss design problem is very popular optimization prob-
• Length of the first shaft between bearings (l1 ). lem in the literature [61,62]. As shown in Fig. 24, there are ten
• Length of the second shaft between bearings (l2 ). nodes which are fixed and twenty five bars cross-sectional mem-
• Diameter of first (d1 ) shafts. bers which are grouped into eight categories:
• The diameter of second shafts (d2 ). • Group 1: A1
• Group 2: A2 , A3 , A4 , A5
The mathematical formulation of this problem is formulated in • Group 3: A6 , A7 , A8 , A9
Appendix D. • Group 4: A10 , A11
Table 8 shows the comparison of best solutions for six opti-
mization algorithms and proposed MOPSO in terms of function val- • Group 5: A12 , A13
ues. • Group 6: A14 , A15 , A17
From Table 8, it is evident that MOSHO produces best solution • Group 7: A18 , A19 , A20 , A21
with considerable improvement as compared to other algorithms. • Group 8: A22 , A23 , A24 , A25
MOSHO algorithm is able to find a design with minimum weight
and stress, simultaneously. The other variables which effects on this problem are as:
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 193

• p = 0.0272 N/cm3 (0.1 lb/in.3 ) 6. Conclusions

• E = 68947 MPa (10 0 0 0 Ksi)
• Displacement limitation = 0.35 in. This paper presented a multi-objective swarm-based optimiza-
• Maximum displacement = 0.3504 in. tion technique called Multi-objective Spotted Hyena Optimizer
• Design variable set = {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, (MOSHO). The two components are introduced to solve multi-
1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0, 2.1, 2.2, 2.3, 2.4, 2.6, 2.8, objective problem. The first component is an archive for retrieving
3.0, 3.2, 3.4 } and storing the best non-dominated solutions which are obtained
so far and the second component is a group selection mechanism
The member stress limitations for this truss is shown in of nearby optimal solutions from the archive. MOSHO has been
Table 10. tested on twenty four well-known benchmark test functions. The
The leading conditions of 25-bar truss is tabulated in Table 11. results reveal that the performance of MOSHO algorithm is better
Table 12 shows the best optimal weight obtained by MOSHO is than the MOPSO, NSGA-2, MOEA/D, PESA-2, SPEA-2, and MOACO.
better than other competitor algorithms. The statistical results (i.e., The statistical results demonstrate that the proposed method is
average and standard deviation) also show that MOSHO outper- able to handle many constraints and offers better solutions than
forms the other algorithms and converge very efficiently to opti- other optimizers. The proposed method is applied on six real-
mize this problem as shown in Fig. 25. life multi-objective optimization problems. The results reveal that
In summary, the experimental results reveal that MOSHO MOSHO provides better results than the others in terms of compu-
is the obvious choice for high performance to solve challeng- tational cost.
ing problems. The optimization results point out that MOSHO For future work, we are planning to extend this algorithm to
has the capability to handle various multi-objective optimization solve many-objective problems in the field of bio-informatics and
problems. further its applicability on data clustering techniques.

Appendix A. Unconstrained multi-objective test problems

Name Mathematical formulation Properties

2  jπ 2
 
UF1 f 1 = x1 + [x − sin 6π x1 + ] Bi-objective
| J1 | j J1 j  n 
√ 2  jπ 2
f2 = 1 − x + [x − sin 6π x1 + ]
| J2 | j J2 j n
J1 = { j | j is odd and 2 ≤ j ≤ n}, J2 = { j | j is even and 2 ≤ j ≤ n}
2 
UF2 f 1 = x1 + j J1 y2j Bi-objective
| J1 |
√ 2 
f2 = 1 − x + y2
| J2 | j J2 j
J1 = { j | j is odd and 2 ≤ j ≤ n}, J2 = { j | j is even and 2 ≤ j ≤ n}
⎧    
⎨x j − [0.3x21 cos 24π x1 + 4 jπ + 0.6x1 ]cos 6π x1 + jπ , i f j J1
yj =  n   n 
⎩x j − [0.3x2 cos 24π x1 + 4 jπ + 0.6x1 ]sin 6π x1 + jπ , i f j J2
1
n n
2
  
 20y π  
UF3 f 1 = x1 + 4 j J1 y2j − 2 j J1 cos j +2 Bi-objective
| J1 | j
√ 2
  
 20y π  
f 2 = x1 + 4 j J1 y2j − 2 j J2 cos j +2
| J2 | j
J1 = { j | j is odd and 2 ≤ j ≤ n}, J2 = { j | j is even and 2 ≤ j ≤ n}
3( j−2 )
0.5(1.0+ )
n−2
y j = x j − x1 , j = 2, 3, . . . , n
2 
UF4 f 1 = x1 + h (y j ) Bi-objective
| J1 | j J1
2 
f 2 = 1 − x2 + h (y j )
| J2 | j J2
J1 = { j | j is odd and 2 ≤ j ≤ n}, J2 = { j | j is even and 2 ≤ j ≤ n}
 jπ
 |t |
y j = x j − sin 6π x1 + , j = 2, 3, . . . , n, h (t ) =
n 1 + e2|t |
 1  2 
UF5 f 1 = x1 + +  | sin(2Nπ x1 ) | + h ( yi ) Bi-objective
2N  | J1 | j J1
1 2 
f 2 = 1 − x1 + +  | sin(2Nπ x1 ) | + h ( yi )
2N | J2 | j J2
J1 = { j | j is odd and 2 ≤ j ≤ n}, J2 = { j | j is even and 2 ≤ j ≤ n}
 jπ

 > 0, y j = x j − sin 6π x1 + , j = 2, 3, . . . , n, h(t ) = 2t 2 − cos(4π t ) + 1
n
 1  2
  
 20y π 
UF6 f 1 = x1 + max{0, 2 +  sin(2Nπ x1 )} + 4 j J1 y2j − 2 j J1 cos j + 1) Bi-objective
2N | J1 | j
 1  2
  
 20 y π 
f 2 = 1 − x1 + max{0, 2 +  sin(2Nπ x1 )} + 4 j J2 y2j − 2 j J2 cos j + 1)
2N | J2 | j
J1 = { j | j is odd and 2 ≤ j ≤ n}, J2 = { j | j is even and 2 ≤ j ≤ n}
 jπ

 > 0, y j = x j − sin 6π x1 + , j = 2, 3, . . . , n
n
√ 2 
UF7 f 1 = 5 x1 + y2 Bi-objective
| J1 | j J1 j
√ 2 
f 2 = 1 − 5 x1 + y2
| J2 | j J2 j
(continued on next page)
194 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

(continued)

Name Mathematical formulation Properties

J1 = { j | j is odd and 2 ≤ j ≤ n}, J2 = { j | j is even and 2 ≤ j ≤ n}

 jπ

 > 0, y j = x j − sin 6π x1 + , j = 2, 3, . . . , n
n
2 
   
jπ 2
UF8 f 1 = cos(0.5x1 π )cos(0.5x2 π ) + j J1 x j − 2x2 sin 2π x1 + Tri-objective
| J1 |   n
 
2  jπ 2
f 2 = cos(0.5x1 π )sin(0.5x2 π ) + j J2 x j − 2x2 sin 2π x1 +

2 
| J2 |  jπ 2
  n
f 3 = sin(0.5x1 π ) + j J3 x j − 2x2 sin 2π x1 +
| J3 | n
J1 = { j | 3 ≤ j ≤ n, and j − 1 is a multipl ication o f 3}
J2 = { j | 3 ≤ j ≤ n, and j − 2 is a multipl ication o f 3}
J3 = { j | 3 ≤ j ≤ n, and j is a multipl ication o f 3}
2 
jπ 2
  
UF9 f 1 = 0.5[max{0, (1 +  )(1 − 4(2x1 − 1 )2 )} + 2x1 ]x2 + j J1 x j − 2x2 sin 2π x1 + Tri-objective
| J1 |  n
 
2  jπ 2
f 2 = 0.5[max{0, (1 +  )(1 − 4(2x1 − 1 )2 )} + 2x1 ]x2 + j J2 x j − 2x2 sin 2π x1 +
    | J2 | n
2  jπ 2
f 3 = 1 − x2 + x − 2x2 sin 2π x1 +
| J3 | j J3 j n
J1 = { j | 3 ≤ j ≤ n, and j − 1 is a multipl ication o f 3}
J2 = { j | 3 ≤ j ≤ n, and j − 2 is a multipl ication o f 3}
J3 = { j | 3 ≤ j ≤ n, and j is a multipl ication o f 3},  = 0.1
2 
UF10 f 1 = cos(0.5x1 π )cos(0.5x2 π ) + 2
j J1 [4y j − cos(8π y j ) + 1] Tri-objective
| J1 |
2 
f 2 = cos(0.5x1 π )sin(0.5x2 π ) + [4y2 − cos(8π y j ) + 1]
| J2 | j J2 j
2 
f 3 = sin(0.5x1 π ) + [4y2 − cos(8π y j ) + 1]
| J3 | j J3 j
J1 = { j | 3 ≤ j ≤ n, and j − 1 is a multipl ication o f 3}
J2 = { j | 3 ≤ j ≤ n, and j − 2 is a multipl ication o f 3}
J3 = { j | 3 ≤ j ≤ n, and j is a multipl ication o f 3}

Appendix B. Unconstrained multi-objective test problems 0 ≤ x1 ≤ 1, −5 ≤ xi ≤ 5, i = 1, 2, . . . , n

ZDT6:
ZDT1: Minimize: f1 (x ) = 1 − e−4x1 × sin6 (6π x1 )
 f ( x ) 2
Minimize: f 1 ( x ) = x1
Minimize: f 2 (x ) = 1 − 1
Minimize: f2 (x ) = g(x ) × h( f1 (x ), g(x )) g( x )
where, where,
9 N ⎡  ⎤0.25
g( x ) = 1 + xi n
i=2 xi
N − 1 i=2 g( x ) = 1 + 9 ⎣ ⎦
f 1 (x ) (n − 1 )
h( f1 (x ), g(x )) = 1 −
g( x )
0 ≤ xi ≤ 1, i = 1, 2, . . . , n
0 ≤ xi ≤ 1, 1 ≤ i ≤ 30
ZDT2:
Minimize: f 1 ( x ) = x1
Minimize: f2 (x ) = g(x ) × h( f1 (x ), g(x )) Appendix C. Unconstrained multi-objective test problems
where,
9 N
g( x ) = 1 + x DTLZ1:
N − 1 i=2 i 1
 2 Minimize: f1 (x ) = x1 (1 + g(x ))
f 1 (x ) 2
h( f1 (x ), g(x )) = 1 − 1
g( x ) Minimize: f2 (x ) = (1 − x1 )(1 + g(x ))
2
0 ≤ xi ≤ 1, 1 ≤ i ≤ 30 where,  
ZDT3: g(x ) = 100 | x | + xi  x (x1 − 0.5 )2 −
Minimize: f 1 ( x ) = x1 
Minimize: f2 (x ) = g(x ) × h( f1 (x ), g(x )) cos(20π (xi − 0.5 ))
where,
9 N 0 ≤ xi ≤ 1, i = 1, 2, . . . , n
g( x ) = 1 + x DTLZ2:  π
29 i=2 i
f 1 (x ) Minimize: f1 (x ) = (1 + g(x ))cos x1
h( f1 (x ), g(x )) = 1 −  π2 
g( x ) f2 (x ) = (1 + g(x ))sin x1
  Minimize:
2
f (x )
− 1 sin(10π f1 (x )) where,
g( x ) 
g(x ) = x  x (xi − 0.5 )2
0 ≤ xi ≤ 1, 1 ≤ i ≤ 30 i
0 ≤ xi ≤ 1, i = 1, 2, . . . , n
ZDT4: DTLZ3:  π
Minimize: f 1 ( x ) = x1
Minimize: f1 (x ) = (1 + g(x ))cos x1
Minimize: f2 (x ) = g(x ) × [1 − (x1 /g(x ))2 ]  π2 
where,
n Minimize: f2 (x ) = (1 + g(x ))sin x1
g(x ) = 1 + 10(n − 1 ) + i=2 (xi
2 − 10cos(4π xi )) 2
where,
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 195

 
g(x ) = 100 | x | + D2. Multiple-disk clutch brake design problem
xi  x (xi − 0.5 )2 −

cos(20π (xi − 0.5 ))
0 ≤ xi ≤ 1, i = 1, 2, . . . , n Minimize f1 (z ) = M = π (ro2 − ri2 )t (Z + 1 ) pm ,
DTLZ4:  π Iz w
Minimize: f1 (x ) = (1 + g(x ))cos xα Minimize f2 (z ) = T = ,
1 Mh + M f
 2
π
Minimize: f2 (x ) = (1 + g(x ))sin xα
1
Subject to
2
where, g1 (z ) = ro − ri − R ≥ 0,

g(x ) = x  x (xi − 0.5 )2
i
g2 (z ) = Lmax − (Z + 1 )(t + δ ) ≥ 0,
α = 100 g3 (z ) = pmax − prz ≥ 0,
0 ≤ xi ≤ 1, i = 1, 2, . . . , n
DTLZ5: g4 (z ) = pmaxVsr,max − przVsr ≥ 0,
 1 + 2g(x )x π
Minimize: f1 (x ) = (1 + g(x ))cos 1
× g5 (z ) = Vsr,max − Vsr ≥ 0,
 14(+1 2+g(gx()xx)) 2

g6 (z ) = Mh − sMs ≥ 0,
π
Minimize: f2 (x ) = (1 + g(x ))sin 1
× g7 (z ) = T ≥ 0,
4(1 + g(x )) 2
where,
 g8 (z ) = Tmax − T ≥ 0,
g(x ) = x  x (xi − 0.5 )2
i 60 ≤ ri ≤ 80 mm,
0 ≤ xi ≤ 1, i = 2, 3, . . . , n
DTLZ6: 90 ≤ ro ≤ 110 mm,
 1 + 2g(x )x π
Minimize: f1 (x ) = (1 + g(x ))cos 1
× 1.5 ≤ t ≤ 3 mm,
 14(+1 2+g(gx()xx)) 2

π 0 ≤ F ≤ 10 0 0 N,
Minimize: f2 (x ) = (1 + g(x ))sin 1
×
4(1 + g(x )) 2 2≤Z≤9
where,

g(x ) = x  x x0i .1
i
0 ≤ xi ≤ 1, i = 2, 3, . . . , n where
DTLZ7: 3
pm = 0.0 0 0 0 078 kg/mm , pmax = 1 MPa,
Minimize: f1 (x ) = x1 μ = 0.5, Vsr,max = 10 m/s,
Minimize: f2 (x ) = (1 + g(x ))h( f1 (x ), g(x )) s = 1.5, Tmax = 15 s, n = 250 rpm, Ms = 40 Nm, M f = 3 Nm,
where, Iz = 55 kg-m , δ = 0.5 mm, R = 20 mm, Lmax = 30 mm,
2
9 
g(x ) = 1 +
| x | xi  x xi 2 ro3 − ri3 πn 2 ro3 − ri3
M h = μF Z 2 N-mm, w = rad/s, Rsr = mm
f1 (x ) 3 ro − ri 2 30 3 ro2 − ri2
h( f1 (x ), g(x )) = M − (1 + sin(3π f1 (x ))) F π Rsr n
1 + g(x ) A = π (ro2 − ri2 ) mm2 , prz =
2
N/mm , Vsr = mm/s,
0 ≤ xi ≤ 1, 1 ≤ i ≤ n A 30
D3. Pressure vessel design problem

Appendix D. Constrained engineering design problems

D1. Welded beam design problem Minimize f1 (z ) = 0.6224Ts LR + 1.7781Th R2 + 3.1661Ts2 L
+19.84Ts2 R,
Minimize f2 (z ) = −(π R2 L + 1.333π R3 ),
Subject to
Minimize f1 (z ) = C = 1.10471h2 l + 0.04811tb(14.0 + l ), g1 (z ) = 0.0193R − Ts ≤ 0,
2.1952 g2 (z ) = 0.00954R − Th ≤ 0,
Minimize f2 (z ) = D = ,
t 3b g3 (z ) = 0.0625 − Ts ≤ 0,
Subject to g4 (z ) = Ts − 5 ≤ 0,
g1 (z ) = 13, 600 − τ (z ) ≥ 0, g5 (z ) = 0.0625 − Th ≤ 0,
g2 (z ) = 30, 0 0 0 − σ (z ) ≥ 0, g6 (z ) = Th − 5 ≤ 0,
g3 (z ) = b − h ≥ 0, g7 (z ) = 10 − R ≤ 0,
g4 (z ) = Pc (z ) − 60 0 0 ≥ 0, g8 (z ) = R − 200 ≤ 0,
g9 (z ) = 10 − L ≤ 0,
Variable range g10 (z ) = L − 240 ≤ 0,
0.125 ≤ h, b ≤ 5.0in.,
D4. Speed reducer design problem
0.1 ≤ l, t ≤ 10.0in.,

where
 Minimize f1 (z ) = 0.7854z1 z22 (3.3333z32 + 14.9334z3 − 43.0934 )

τ (z ) = (τ ) + (τ ) + (l τ τ )/ 0.25( + (h + t ) ),
 2  2  
l2 2 −1.508z1 (z62 + z72 ) + 7.4777(z63 + z73 ) + 0.7854(z4 z62 + z5 z72 ),
 6, 0 0 0 504, 0 0 0
τ = √ , σ (z ) = , 745z4 2
2hl 2
t b + 1.69 × 107
z2 z3
 6, 0 0 0(14 + 0.5l ) 0.25(l 2 + (h + t )2 ) Minimize f2 (z ) = ,
τ = , 0.1z63
2[0.707hl (l 2 /12 + 0.25(h + t )2 )]
Pc (z ) = 64, 746.022(1 − 0.0282346t )t b . 3 Subject to
196 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197

27 [11] T. Kipouros, D.M. Jaeggi, W.N. Dawes, G.T. Parks, A.M. Savill, P.J. Clarkson, Biob-
g1 (z ) = − 1 ≤ 0, jective design optimization for axial compressors using tabu search, AIAA J. 46
z1 z22 z3 (3) (2008) 701–711, doi:10.2514/1.32794.
397.5 [12] G. Dhiman, V. Kumar, Spotted hyena optimizer for solving complex and non–
g2 (z ) = − 1 ≤ 0, linear constrained engineering problems, Advances in Intelligent Systems and
z1 z22 z32 Computing, Springer, 2018, In press.
[13] M. Chen, O. Hammami, A System Engineering Conception of Multi-objective
1.93z43
g3 (z ) = − 1 ≤ 0, Optimization for Multi-physics System, Springer International Publishing,
z2 z64 z3 Cham, pp. 299–306. doi:10.1007/978- 3- 319- 14532- 7_31.
[14] J. Handl, D.B. Kell, J. Knowles, Multiobjective optimization in bioinformatics
1.93z53 and computational biology, IEEE/ACM Trans. Comput. Biol. Bioinformatics 4 (2)
g4 (z ) = − 1 ≤ 0, (2007) 279–292, doi:10.1109/TCBB.2007.070203.
z2 z74 z3 [15] A. Kaur, G. Dhiman, A review on search based tools and techniques to identify
[(745(z4 /z2 z3 ))2 + 16.9 × 106 ]1/2 bad code smells in object oriented systems, Advances in Intelligent Systems
g5 (z ) = − 1 ≤ 0, and Computing, Springer, 2018, In press.
110z63 [16] R. Maini, G. Dhiman, Impacts of artificial intelligence on real-life problems, In-
ternation Journal Of Advance Research And Innovative Ideas In Education 4 (1)
[(745(z5 /z2 z3 ))2 + 157.5 × 106 ]1/2 (2018) 291–295.
g6 (z ) = − 1 ≤ 0, [17] C.A.C. Coello, G.B. Lamont, D.A.V. Veldhuizen, Evolutionary Algorithms for
85z73
Solving Multi-Objective Problems (Genetic and Evolutionary Computation),
z2 z3 Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2006.
g7 (z ) = − 1 ≤ 0, [18] R.K. Chandrawat, R. Kumar, B.P. Garg, G. Dhiman, S. Kumar, An Analysis of
40
Modeling and Optimization Production Cost Through Fuzzy Linear Program-
5z2
g8 (z ) = − 1 ≤ 0, ming Problem with Symmetric and Right Angle Triangular Fuzzy Number,
z1 Springer Singapore, Singapore, pp. 197–211. doi:10.1007/978- 981- 10- 3322- 3_
z1 18.
g9 (z ) = − 1 ≤ 0, [19] P. Singh, G. Dhiman, A fuzzy-lp approach in time series forecasting, in:
12z2 B.U. Shankar, K. Ghosh, D.P. Mandal, S.S. Ray, D. Zhang, S.K. Pal (Eds.), Pattern
1.5z6 + 1.9 Recognition and Machine Intelligence, Springer International Publishing, Cham,
g10 (z ) = − 1 ≤ 0, 2017, pp. 243–253.
z4 [20] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective
1.1z7 + 1.9 genetic algorithm: NSGA-ii, IEEE Trans. Evol. Comput. 6 (2) (2002) 182–197,
g11 (z ) = − 1 ≤ 0, doi:10.1109/4235.996017.
z5 [21] C.A. Coello Coello, M.S. Lechuga, Mopso: A proposal for multiple objective par-
where ticle swarm optimization, in: Proceedings of the Evolutionary Computation on
2002. CEC ’02. Proceedings of the 2002 Congress - Volume 02, in: CEC ’02,
2.6 ≤ z1 ≤ 3.6, 0.7 ≤ z2 ≤ 0.8, 17 ≤ z3 ≤ 28, 7.3 ≤ z4 ≤ 8.3, IEEE Computer Society, Washington, DC, USA, 2002, pp. 1051–1056.
[22] Q. Zhang, H. Li, MOEA/D: A multiobjective evolutionary algorithm based on
7.3 ≤ z5 ≤ 8.3, 2.9 ≤ z6 ≤ 3.9, 5.0 ≤ z7 ≤ 5.5 decomposition, IEEE Trans. Evol. Comput. 11 (6) (2007) 712–731, doi:10.1109/
TEVC.2007.892759.
[23] D.H. Wolpert, W.G. Macready, No free lunch theorems for optimization, Trans.
D5. Gear train design problem Evol. Comp 1 (1) (1997) 67–82, doi:10.1109/4235.585893.
[24] Q. Zhang, A. Zhou, S. Zhao, P.N. Suganthan, W. Liu, S. Tiwari, Multiobjective op-
timization test instances for the cec 2009 special session and competition, Uni-
versity of Essex, Colchester, UK and Nanyang technological University, Singa-
 
 Ta T f  pore, special session on performance assessment of multi-objective optimiza-
Minimize f1 (T ) = 6.931 −  tion algorithms, technical report 264 (2008).
 Td Tb  [25] E. Zitzler, K. Deb, L. Thiele, Comparison of multiobjective evolutionary algo-
Minimize f2 (T ) = max Td , Tb , Ta , T f rithms: empirical results, Evol. Comput. 8 (2) (20 0 0) 173–195, doi:10.1162/
106365600568202.
Subject to [26] K. Deb, L. Thiele, M. Laumanns, E. Zitzler, Scalable Test Problems for Evolu-
f 1 (T ) tionary Multiobjective Optimization, Springer London, London, pp. 105–145.
≤ 0.5, doi:10.1007/1- 84628- 137- 7_6.
6.931 [27] D.W. Corne, N.R. Jerram, J.D. Knowles, M.J. Oates, M. J, Pesa-ii: Region-based
12 ≤ Td , Tb , Ta , T f ≤ 60, selection in evolutionary multiobjective optimization, in: Proceedings of the
Td , Tb , Ta , T f are integers Genetic and Evolutionary Computation Conference (GECCO’2001, Morgan Kauf-
mann Publishers, 2001, pp. 283–290.
[28] E. Zitzler, M. Laumanns, L. Thiele, SPEA2: Improving the Strength Pareto Evo-
References lutionary Algorithm, Technical Report, 2001.
[29] D. Angus, C. Woodward, Multiple objective ant colony optimisation, Swarm In-
[1] G. Dhiman, V. Kumar, Spotted hyena optimizer: a novel bio-inspired based tell. 3 (1) (2009) 69–85, doi:10.1007/s11721- 008- 0022- 4.
metaheuristic technique for engineering applications, Adv. Eng. Software 114 [30] C.A. Coello Coello, Evolutionary multi-objective optimization: some current re-
(2017) 48–70, doi:10.1016/j.advengsoft.2017.05.014. search trends and topics that remain to be explored, Front. Comput. Sci. China
[2] J. Branke, K. Deb, H. Dierolf, M. Osswald, Finding Knees in Multi-objective Op- 3 (1) (2009) 18–30, doi:10.1007/s11704- 009- 0005- 7.
timization, Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 722–731. doi:10. [31] O. Schütze, M. Laumanns, C.A. Coello Coello, M. Dellnitz, E.-G. Talbi, Conver-
1007/978- 3- 540- 30217- 9_73. gence of stochastic search algorithms to finite size pareto set approximations,
[3] R. Marler, J. Arora, Survey of multi-objective optimization methods for en- J. Global Optim. 41 (4) (2008) 559–577, doi:10.1007/s10898- 007- 9265- 7.
gineering, Struct. Multidiscip. Optim. 26 (6) (2004) 369–395, doi:10.1007/ [32] F. Edgeworth, Mathematical physics: P.London, England: Keagan (1881).
s0 0158-0 03-0368-6. [33] V. Pareto, Cours d’economie politique: Librairie Droz (1964).
[4] K. Deb, Advances in Evolutionary Multi-objective Optimization, Springer Berlin [34] M. Gong, L. Jiao, H. Du, L. Bo, Multiobjective immune algorithm with non-
Heidelberg, Berlin, Heidelberg, pp. 1–26. doi:10.1007/978- 3- 642- 33119- 0_1. dominated neighbor-based selection, Evol. Comput. 16 (2) (2008) 225–255,
[5] X.-S. Yang, M. Karamanoglu, X. He, Flower pollination algorithm: a novel ap- doi:10.1162/evco.2008.16.2.225.
proach for multiobjective optimization, Eng. Optim. 46 (9) (2014) 1222–1237. [35] D. Yang, L. Jiao, M. Gong, J. Feng, Adaptive ranks clone and k-nearest neighbor
[6] X.-S. Yang, A.H. Gandomi, Bat algorithm: a novel approach for global engineer- list-based immune multi-objective optimization, Comput. Intell. 26 (4) (2010)
ing optimization, Eng. Comput. (Swansea) 29 (5) (2012) 464–483. 359–385, doi:10.1111/j.1467-8640.2010.00363.x.
[7] X.-S. Yang, Engineering optimization: An introduction with metaheuristic ap- [36] P.M. Pradhan, G. Panda, Solving multiobjective problems using cat swarm op-
plications, John Wiley & Sons, 2010. timization, Expert Syst Appl 39 (3) (2012) 2956–2964. https://doi.org/10.1016/
[8] N.K. Madavan, Multiobjective optimization using a pareto differential evolu- j.eswa.2011.08.157.
tion approach, in: Evolutionary Computation, 2002. CEC ’02. Proceedings of the [37] W. Lin, D. Yu, S. Wang, C. Zhang, S. Zhang, H. Tian, M. Luo, S. Liu, Multi-
2002 Congress on, 2, 2002, pp. 1145–1150, doi:10.1109/CEC.2002.1004404. objective teaching-learning-based optimization algorithm for reducing carbon
[9] C.A.C. Coello, Evolutionary multi-objective optimization: a historical view of emissions and operation time in turning operations, Eng. Optim. 47 (7) (2015)
the field, IEEE Comput. Intell. Mag. 1 (1) (2006) 28–36, doi:10.1109/MCI.2006. 994–1007, doi:10.1080/0305215X.2014.928818.
1597059. [38] E. Hancer, B. Xue, M. Zhang, D. Karaboga, B. Akay, A multi-objective artificial
[10] G.-C. Luh, C.-H. Chueh, Multi-objective optimal design of truss structure with bee colony approach to feature selection using fuzzy mutual information, in:
immune algorithm, Comput. Struct. 82 (11–12) (2004) 829–844. https://doi. 2015 IEEE Congress on Evolutionary Computation (CEC), 2015, pp. 2420–2427,
org/10.1016/j.compstruc.20 04.03.0 03. doi:10.1109/CEC.2015.7257185.
 G. Dhiman, V. Kumar / Knowledge-Based Systems 150 (2018) 175–197 197

[39] H. Hemmatian, A. Fereidoon, E. Assareh, Optimization of hybrid laminated [51] E. Zitzler, L. Thiele, Multiobjective evolutionary algorithms: a comparative case
composites using the multi-objective gravitational search algorithm (mogsa), study and the strength pareto approach, IEEE Trans. Evol. Comput. 3 (4) (1999)
Eng. Optim. 46 (9) (2014) 1169–1182, doi:10.1080/0305215X.2013.832234. 257–271, doi:10.1109/4235.797969.
[40] R. Shang, H. Liu, L. Jiao, Multi-objective clustering technique based on k-nodes [52] C.A. Coello Coello, C. Dhaenens, L. Jourdan, Multi-Objective Combinatorial Op-
update policy and similarity matrix for mining communities in social net- timization: Problematic and Context, Springer Berlin Heidelberg, Berlin, Hei-
works, Physica A 486 (Supplement C) (2017) 1–24, doi:10.1016/j.physa.2017.05. delberg, pp. 1–21. doi:10.1007/978- 3- 642- 11218- 8_1.
026. [53] G. Rudolph, O. Schütze, C. Grimme, C. Domínguez-Medina, H. Trautmann, Op-
[41] D. Yang, L. Jiao, M. Gong, F. Liu, Artificial immune multi-objective sar im- timal averaged hausdorff archives for bi-objective problems: theoretical and
age segmentation with fused complementary features, Inf. Sci. (Ny) 181 (13) numerical results, Comput. Optim. Appl. 64 (2) (2016) 589–618, doi:10.1007/
(2011) 2797–2812, doi:10.1016/j.ins.2011.02.025. Including Special Section on s10589-015-9815-8.
Databases and Software Engineering [54] O. Schütze, X. Esquivel, A. Lara, C.A.C. Coello, Using the averaged hausdorff
[42] S. Yang, M. Wang, L. jiao, A quantum particle swarm optimization, in: distance as a performance measure in evolutionary multiobjective optimiza-
Proceedings of the 2004 Congress on Evolutionary Computation (IEEE Cat. tion, IEEE Trans. Evol. Comput. 16 (4) (2012) 504–522, doi:10.1109/TEVC.2011.
No.04TH8753), 1, 2004, pp. 320–324Vol.1, doi:10.1109/CEC.2004.1330874. 2161872.
[43] W. Zhong, J. Liu, M. Xue, L. Jiao, A multiagent genetic algorithm for global nu- [55] O. Schütze, M. Laumanns, E. Tantar, C.A.C. Coello, E.G. Talbi, Computing gap
merical optimization, IEEE Trans. Syst„ Man, Cybern., Part B (Cybernetics) 34 free pareto front approximations with stochastic search algorithms, Evol. Com-
(2) (2004) 1128–1141, doi:10.1109/TSMCB.2003.821456. put. 18 (1) (2010) 65–96, doi:10.1162/evco.2010.18.1.18103.
[44] L. Jiao, S. Member, L. Wang, A novel genetic algorithm based on immunity, [56] M. Li, J. Zheng, Spread Assessment for Evolutionary Multi-Objective Optimiza-
IEEE Trans. on Systems, Man, and Cybernetics-Part A: Systems and Humans 30 tion, Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 216–230. doi:10.1007/
(20 0 0) 552–561. 978- 3- 642- 01020- 0_20.
[45] J.A.H. Mejia, O. Schütze, O. Cuate, A. Lara, K. Deb, Rds-nsga-ii: a memetic algo- [57] E. Zitzler, L. Thiele, M. Laumanns, C.M. Fonseca, V.G. da Fonseca, Performance
rithm for reference point based multi-objective optimization, Eng. Optim. 49 assessment of multiobjective optimizers: an analysis and review, Trans. Evol.
(5) (2017) 828–845, doi:10.1080/0305215X.2016.1211127. Comp 7 (2) (2003) 117–132, doi:10.1109/TEVC.2003.810758.
[46] A. Ilany, A.S. Booms, K.E. Holekamp, Topological effects of network structure [58] C.A.C. Coello, Theoretical and numerical constraint-handling techniques used
on long-term social network dynamics in a wild mammal, Ecol. Lett. 18 (7) with evolutionary algorithms: a survey of the state of the art, Comput. Meth-
(2015) 687–695, doi:10.1111/ele.12447. ods Appl. Mech. Eng. 191 (11–12) (2002) 1245–1287. https://doi.org/10.1016/
[47] C.A. Coello, G.T. Pulido, M.S. Lechuga, Handling multiple objectives with parti- S0 045-7825(01)0 0323-1.
cle swarm optimization, Trans. Evol. Comp 8 (3) (2004) 256–279, doi:10.1109/ [59] B. Kannan, S.N. Kramer, An augmented lagrange multiplier based method for
TEVC.2004.826067. mixed integer discrete continuous optimization and its applications to me-
[48] J.D. Knowles, D.W. Corne, Approximating the nondominated front using the chanical design, J. Mech. Des. 116 (2) (1994) 405–411.
pareto archived evolution strategy, Evol. Comput. 8 (2) (20 0 0) 149–172, doi:10. [60] A.H. Gandomi, X.-S. Yang, Benchmark problems in structural optimization,
1162/106365600568167. Springer Berlin Heidelberg (2011) 259–281, doi:10.1007/978- 3- 642- 20859- 1_
[49] D.W. Corne, N.R. Jerram, J.D. Knowles, M.J. Oates, Pesa-ii: Region-based selec- 12.
tion in evolutionary multiobjective optimization, in: Proceedings of the 3rd [61] G. Dhiman, A. Kaur, A hybrid algorithm based on particle swarm and spotted
Annual Conference on Genetic and Evolutionary Computation, Morgan Kauf- hyena optimizer for global optimization, Advances in Intelligent Systems and
mann Publishers Inc., 2001, pp. 283–290. Computing, Springer, 2018. In press.
[50] M.N. Ab Wahab, S. Nefti-Meziani, A. Atyabi, A comprehensive review of swarm [62] G. Dhiman, A. Kaur, Spotted hyena optimizer for solving engineering design
optimization algorithms, PLoS ONE 10 (5) (2015) 1–36, doi:10.1371/journal. problems, in: International Conference on Machine Learning and Data Science,
pone.0122827. IEEE, 2018. In press.