Applied Soft Computing Journal 77 (2019) 567–583

Contents lists available at ScienceDirect

Applied Soft Computing Journal

journal homepage: [Link]/locate/asoc

A Quasi-Oppositional-Chaotic Symbiotic Organisms Search algorithm

for global optimization problems
∗
Khoa H. Truong a , Perumal Nallagownden a , Zuhairi Baharudin a , Dieu N. Vo b ,
a
Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, Malaysia
b
Department of Power Systems, Ho Chi Minh City University of Technology, VNU-HCM, VietNam

highlights graphical abstract

• An improved algorithm called Quasi-

Oppositional Chaotic Symbiotic Or-
ganisms Search (QOCSOS) is
presented for solving global optimiza-
tion problems.
• QOCSOS is tested on twenty-six math-
ematical benchmark functions.
• QOCSOS is applied to solve an opti-
mal DG placement problem in radial
distribution networks.
• QOCSOS is applied to solve five engi-
neering design problems.
• For all application, QOCSOS achieved
better solution quality than other
methods.

article info a b s t r a c t

Article history: This study proposes an improved version of the Symbiotic Organisms Search (SOS) algorithm called
Received 3 July 2018 Quasi-Oppositional Chaotic Symbiotic Organisms Search (QOCSOS). This improved algorithm integrated
Received in revised form 7 November 2018 Quasi-Opposition-Based Learning (QOBL) and Chaotic Local Search (CLS) strategies with SOS for a better
Accepted 28 January 2019
quality solution and faster convergence. To demonstrate and validate the new algorithm’s effectiveness,
Available online 4 February 2019
the authors tested QOCSOS with twenty-six mathematical benchmark functions of different types and
Keywords: dimensions. In addition, QOCSOS optimized placements for distributed generation (DG) units in radial
Chaotic local search distribution networks and solved five structural design optimization problems, as practical optimization
Distributed generation problems challenges. Comparative results showed that QOCSOS provided more accurate solutions than
Global optimization problems SOS and other methods, suggesting viability in dealing with global optimization problems.
Meta-heuristic quasi-opposition-based © 2019 Elsevier B.V. All rights reserved.
learning
Symbiotic organisms search

∗ Corresponding author.
E-mail address: vndieu@[Link] (D.N. Vo).

[Link]
1568-4946/© 2019 Elsevier B.V. All rights reserved.
568 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

1. Introduction the search process. But in SOS, only two controllable parameters
govern the search process. Therefore, its exploration and exploita-
Economic, science and engineering fields constantly need up- tion capacities may not balance well and the solution can become
graded approaches to practical optimization problem solving. Gen- trapped in local optima. To avoid local optima, SOS has been
erally, Optimization Problem Solving approaches (OPS) formulate modified for better performance. The study in [27] proposed three
single or multi-objective functions that are subject to a set of modified versions of SOS by introducing adaptive benefit factors
constraints. OPS may be understood as the use of an optimiza- in the original SOS algorithm. In [28], a modified SOS (mSOS) was
tion method to find the best possible solution that satisfies all refined from the original SOS with five modifications in the three
constraints by minimizing or maximizing formulated objective phases: mutualism, commensalism and parasitism. Also, SOS inte-
functions. Classical optimization methods (analytical and numer- grated the quasi-oppositional based learning (Q-OBL) to avoid the
ical) easily deal with optimization problems that have smooth or suboptimal solution and to accelerate the convergence speed [34].
differentiable objective functions. However, they are not capable A hybrid version of SOS was also introduced by incorporating with
of solving complex problems with non-smooth, non-differentiable simulated annealing to form a new technique (hSOS-SA) [36].
objective functions that represent real-time and real-world prob- In the literature, to avoid local optima issues, the QOBL strat-
lems. Hence, various OPS optimization methods have been devel- egy was integrated with various meta-heuristic optimization al-
oped over the past few decades. gorithms such as DE [39], TLBO [40], SIMBO-Q [41], GWO [42]
Several meta-heuristic optimization methods were developed and HS [43]. It enhances an algorithm’s capacity to explore more
to solve complex optimization problems. Most mime biological or promising search domains, which increases the probability of find-
physical phenomena like the Genetic Algorithm (GA) for example, ing a better solution. Besides the QOBL strategy, the CLS strategy
which is based on evolution [1]. Particle Swarm optimization (PSO) also prevents getting stuck in local optima. This integrated tool
imitates the social behaviour of bird flock or school of fish [2]; locally explores a vicinity of ‘current best solution’ to provide
Simulated Annealing (SA) is based on annealing in metallurgy [3]; better exploitation. Various CLS schemes were integrated with
Differential Evolution (DE) is also inspired by the theory of evolu- PSO [44–46], DE [47–49], ABC [50], HS [51] and TLBO [52]. We
tion [4]; and Ant Colony Optimization (ACO) [5] and Artificial Bee believed the integration of meta-heuristic methods with QOBL
Colony (ABC) [6] methods mimic the foraging behaviours of ants and CLS would balance exploration and exploitation phases and
and honey bees. These meta-heuristic methods have successfully significantly improve performance.
solved complex optimization problems with high-quality results Inspired by the effectiveness of combined QOBL and CLS strate-
in various fields of science and engineering. gies, the present work integrated both with the original SOS al-
According to the ‘‘No-Free-Lunch’’ (NFL) theorem [7], there is no gorithm in an effort to improve performance. We called the im-
meta-heuristic algorithm that can effectively solve all optimization proved algorithm ‘Quasi-Oppositional Chaotic SOS’ (QOCSOS). It
problems, which means if a meta-heuristic algorithm effectively utilizes QOBL for population initialization and generation jumping
solves a specific problem, there is no guarantee it would be effec- while CLS addresses the current best population. To assess perfor-
tive for others. Thus, researchers were/are motivated to improve mance, we tested QOCSOS with twenty-six mathematical bench-
existing algorithms and/or develop innovative schemes. For the mark functions of different types (unimodal and multimodal) and
past few years, many new methods inspired by nature have been dimensions. We also employed the original SOS to solve for these
introduced including Harmony Search (HS) [8]; Teaching Learn- same test functions for comparative purposes. Results demon-
ing Based Optimization (TLBO) [9]; Flower Pollination Algorithm strated superior outcomes for QOCSOS in terms of solution quality
(FPA) [10]; Fruit Fly Optimization Algorithm (FFOA) [11]; Krill Herd and convergence speed.
Algorithm (KHA) [12]; Cuckoo Search Algorithm (CSA) [13]; Swine In addition, we challenged QOCSOS with a real-world opti-
Influenza Model Based Optimization (SIMBO) [14]; Mine Blast Al- mization problem in power system to test its ability. QOCSOS
gorithm (MBA) [15]; Lightning Search Algorithm (LSA) [16]; Grey was applied to find optimal locations and sizes for DG units in
Wolf Optimizer (GWO) [17]; Ant Lion Optimizer (ALO) [18]; Whale radial distribution networks that included 33, 69 and 118 buses,
Optimization Algorithm (WOA) [19]; Multi-Verse Optimization respectively. To further prove the acceptability of the proposed
(MVO) [20]; Crow Search Algorithm (CSA) [21]; Selfish Herd Opti- QOCSOS, we applied QOCSOS to solve five engineering design
mizer (SHO) [22]; and the Squirrel Search Algorithm (SSA) [23]. In problems in structural engineering field. Again, the comparative
addition, various improved and modified versions of existing meta- results confirmed its superiority compared to outcomes from the
heuristic optimization methods and hybrid versions of two or more original SOS and other methods.
meta-heuristic methods have also been developed. Section 2 provides a brief description of the SOS algorithm,
Symbiotic Organisms Search (SOS) [24] mimes relationships which is followed by a presentation of the proposed QOCSOS
such as mutualism, commensalism and parasitism between two in Section 3. Simulation results for benchmark test functions in
organisms in an ecosystem. Its advantage is its simple structure Section 4. The solutions for optimal DG placement problem and
that limits only two controllable parameters (Eco_size (popula- engineering design problems are reported in Sections 5 and 6,
tion) and number of fitness evaluation), which makes it easy to respectively. We conclude with a comprehensive discussion in
implement. For this reason, SOS has recently become a popular Section 6.
application for different optimization problems such as project
scheduling [25], vehicle routing [26], structural design [27,28], 2. The SOS algorithm
multi-objective constrained optimization [29], problems in power
systems [30–34], automatic voltage regulator [35,36], optimiza- SOS is a powerful meta-heuristic algorithm that treats opti-
tion of PI parameters for DSP-based DC motor drives [37], opti- mization problems based on symbiotic relationships (mutualism,
mal scheduling of tasks on cloud, and time-cost-labour utilization commensalism and parasitism) between two different organisms
trade-off problems [38]. However, local optima can jam SOS’s in a natural ecosystem [24]; thus, it mimics symbiosis to generate
performance. The search process of meta-heuristic optimization new solutions. It randomly initializes the search process with a
methods for an optimized outcome includes exploration and ex- population of organisms (ecosystem) where each organism has a
ploitation phases. To balance these, besides basic controllable pa- fitness value and represents a solution for a considered problem.
rameters (population size and number of iterations), more specif- Initial solutions are updated per mutualism, commensalism and
ically defined controllable parameters are required to regulate parasitism phase operations. New solutions from each phase are
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 569

Fig. 1. SOS algorithm’s Pseudocode.

only updated if their fitness value is better than previous one. selected from the population to communicate with organism Xi .
The process of three phases is repeated to find out the best so- As a result of this interaction, organism Xi receives benefits while
lutions until stopping criteria are satisfied. These phases are now organism Xj is neither harmed nor benefited. A new organism, Xinew ,
described. is thus produced, calculated as follows [24]:

2.1. Mutualism Xinew = Xi + rand(−1, 1) × (Xbest − Xj ) (4)

This phase simulates mutualistic relationships between two If the fitness value of Xinew is better than the previous value, Xinew
different organisms in which both benefit each other. In SOS, is updated.
vectors Xi and Xj denote ith and jth organisms of the ecosystem,
respectively. Xj is randomly chosen from the ecosystem to interact 2.3. Parasitism
with Xi . New organisms for Xi and Xj are generated by following
equations [24]:
This phase describes a parasitic relationship where an organism
Xinew = Xi + rand(0, 1) × (Xbest − MV × bf1 ) (1) benefits while another is harmed. An example is the interaction
Xjnew = Xj + rand(0, 1) × (Xbest − MV × bf2 ) (2) between the Plasmodium parasite from the Anopheles mosquito,
and a human host. The parasite is transferred to the human host
bf1 + bf2
MV = (3) through the mosquito bite. After infection, the parasite grows,
2 reproduces and invades red blood cells causing malaria. If the
where rand(0,1) is a vector of random numbers between 0 and 1; host’s immunity is strong enough it will exterminate the parasite.
Xbest represents the best organism in an ecosystem; MV denotes Otherwise, the human may contract malaria and become seriously
a mutual vector representing a mutual symbiotic relationship be- ill or possibly die.
tween organisms Xi and Xj ; bf 1 and bf 2 are benefit factors that
In this phase, organism Xi , like the Anopheles mosquito, is
describe the level of benefit for each organism — such factors are
duplicated to produce the Parasite_Vector. Some elements of the
stochastically selected as either 1 or 2 (1 for partial benefit, 2 for
Parasite_Vector are randomly selected and modified by using a
full benefit).
random number within limits. Taken from the current ecosystem,
New organisms (Xinew and Xjnew ) are updated only if their fitness
values are better than fitness values for organisms Xi and Xj . another organism, Xj , is randomly chosen to serve as a human
host for the Parasite_Vector. Fitness values are calculated for both
2.2. Commensalism organisms (Xj and Parasite_Vector). If the Parasite_Vector reaps a
better fitness value, it replaces the position of organism Xj in the
This phase portrays a commensal relationship between two current ecosystem. If not, the Parasite_Vector disappears from that
different organisms where one benefits and the other is neither ecosystem and organism Xj remains [24].
harmed nor helped. Similar to mutualism, organism Xj is randomly Algorithm 1 in Fig. 1 describes the SOS algorithm’s pseudocode.
570 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

Fig. 2. QOBL Pseudocode.

3. Proposed QOCSOS algorithm

Two search strategies (QOBL and CLS) were embedded within

the original SOS algorithm to form the proposed QOCSOS algo-
rithm. Both are now described in detail.

3.1. Quasi-oppositional based learning (QOBL)

3.1.1. QOBL: A Concept

Tizhoosh first proposed Oppositional Based Learning (OBL) [53],
which posits that opposite numbers have a higher probability of
obtaining a solution than random numbers. The integration of
meta-heuristic algorithms with OBL improves solution accuracy
Fig. 3. Bifurcation diagram for the logistic map.
and increase convergence speed. Moreover, OBL had already been
extended to QOBL [39–43], which shows that using quasi-opposite
numbers is more effectively than opposite numbers in finding
global optimal results. Some QOBL definitions are mathematically 3.2. Chaotic local search (CLS)
described as follows:
We applied the CLS strategy for purposes of this study to the
Opposite number and quasi-opposite number current best population of the proposed QOCSOS. CLS explores the
Let x be a real number in 1-dimensional space. Let the opposite nearby search space of current best solutions to generate a new
number xo and the quasi-opposite number xqo (of x) be defined by candidate. From the current best population, a new solution is
Eq. (5) and Eq. (6), respectively: generated using the following equation [54]:
new
,k = Xbest ,k + (Zk − 0.5) × Xi,k − Xj,k
( )
xo = a + b − x (5) Xbest (9)
new
where x ∈ R and x ∈ [a, b]. where Xbest ,k and Xbest ,k are the ‘new generated solution’ and the
( ) ‘current best solution’ at the kth iteration, respectively; Zk is the
a+b
xqo = rand , xo (6) chaotic number between 0 and 1 at the kth iteration; Xi,k and
2 Xj,k are two solutions randomly selected from the current best
population.
Opposite point and quasi-opposite point new
If the new solution, Xbest ,k , generated by CLS has a better fitness
Let X (x1 , x2 , . . . , xn ) be a point in n-dimensional space. Let the
value than the current best solution, Xbest ,k , it replaces the current
opposite point, X o (xo1 , xo2 , . . . , xon ), be defined by Eq. (7); and let the
qo qo qo best solution. The CLS stops when local search limit K is reached.
quasi-opposite point, X qo (x1 , x2 , . . . , xn ), be defined by Eq. (8):
The literature reports many types of chaotic maps [44–47].
xoi = ai + bi − xi (7) In this study, the well-known ‘logistic map’ is used to generate
the chaotic number in CLS. Mathematically, the ‘logistic map’ is
where xi ∈ R and xi ∈ [ai , bi ]∀i ∈ 1, 2, . . . , n. expressed as follows [47]:
( )
ai + b i
qo
xi = rand , xoi (8) Zk+1 = µ × Zk × (1 − Zk ) (10)
2
where Z0 = rand(), Zk ∈ (0, 1)∀k ∈ {0, 1, 2, . . .}, and µ ∈ (0, 4].
3.1.2. Implementation of QOBL In Eq. (10), parameter µ controls the behaviour of the logistic
QOBL is employed in QOCSOS for population initialization and map. The value chosen is 4, for which the logistic function shows
generation jumping. QOBL-based initialization simultaneously comprehensive chaotic behaviour at µ = 4 (Fig. 3), while Fig. 4
generates a randomly generated population and a quasi- depicts a visualization of 100 chaotic numbers generated by the
oppositional population to select a set of best solutions for the logistic map.
initial population. QOBL-based generation jumping helps the al-
gorithm jump to a new candidate solution that bears a better 3.3. The proposed QOCSOS algorithm
fitness value. The parameter, jr (jumping rate), decides whether to
keep a current solution or jump to a quasi-oppositional solution. Two strategies, QOBL and CLS, were embedded within the orig-
Algorithm 2 in Fig. 2 describes the QOBL pseudocode for the new inal SOS algorithm to establish the QOCSOS algorithm, which be-
population. gins the search process with a randomly generated population of
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 571

QOCSOS algorithm’s optimization ability [24]. We coded the QOC-

SOS algorithm with the Matlab R2017a platform then executed 100
independent trials on a 3.5 GHz (8CPUs) PC with 16 GB of RAM.

4.1. Description of test functions

Twenty-six benchmark test functions were classified into three

groups according to dimension. The first group (f1 – f11 ) is only
two-dimensional (Table 1). The second group (f12 – f15 ) contains
four, five (f12 – f13 ) and ten-dimensional (f14 – f15 ) functions, respec-
tively (Table 2). The final group (f16 – f26 ) holds thirty-dimensional
functions (Table 3). All functions may be classified into different
types: ‘UN’ denotes unimodal and non-separable; ‘US’ denotes
Fig. 4. Visualization of the logistic map.
unimodal and separable; ‘MS’ denotes multimodal and separable;
‘MN’ denotes multimodal and non-separable.

organisms. After initialization, QOBL is implemented as Algorithm 4.2. Parameter setting

2 to obtain a new initial population, after which the three-phase
process (mutualism, commensalism and parasitism) is executed. We retained two predetermined control parameters from the
Next, QOBL is re-implemented. Finally, CLS is executed to obtain original SOS algorithm (ecosystem size, Eco_size; and maximum
the best organism. The process repeats until stopping criteria are number of fitness function evaluations, maxFE), set at the same
satisfied. Fig. 5 shows a detailed outline of the QOCSOS algorithm values for comparison [24]. For fair appraisal, Eco_size was set at
as Algorithm 3. 50 and maxFE was set at 500,000 for both algorithms and for each
test function. Another two parameters (jumping rate, jr ; and local
4. Experimental studies on benchmark functions search limit, K ) were also predetermined for the proposed QOCSOS
algorithm.
In addition to the original SOS algorithm, we used twenty-six Parameter jr controls the convergence rate of QOCSOS. A higher
well-known benchmark test functions to compare the proposed jr value prompts the algorithm to jump to a new solution at the

Fig. 5. QOCSOS Pseudocode.

572 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

Table 1
Benchmark test functions: Two-dimensional.
fNo. Name Type D Formulation Range fmin
f1 Beale UN 2 f (x) = (1.5 − x1 + x1 x2 ) + (2.25 − x1 + 2
+ (2.625 − x1 + x1 x22 )2 x1 x32 )2 [−4.5,4.5] 0
f2 Easom UN 2 f (x) = − cos(x1 ) cos(x2 ) exp(−(x1 − π )2 − (x2 − π )2 ) [ −100,100] −1
f3 Matyas UN 2 f (x) = 0.26(x1 + x2 ) − 0.48x1 x2
2 2
[ −10,10] 0
f4 Bohachevsky1 MS 2 f (x) = x21 + 2x22 − 0.3 cos(3π x1 ) − 0.4 cos(4π x2 ) + 0.7 [ −100,100] 0
f5 Booth MS 2 f (x) = (x1 + 2x2 − 7)2 + (2x1 + x2 − 5)2 [ −10,10] 0
D
∑
f6 Michalewicz2 MS 2 f (x) = − sin(xi )(sin(ix2i /π ))20 [0, π ] −1.8013
i=1 √
x21 +x22 )−0.5
sin2 (
f7 Schaffer MN 2 f (x) = 0.5 + [ −100,100] 0
(1+0.0001(x21 +x22 ))2
f8 Six Hump Camel Back MN 2 f (x) = 4x21 − 2.1x41 + 13 x61 + x1 x2 − 4x22 + 4x42 [ −5,5] −1.03163
f9 Bohachevsky2 MN 2 f (x) = x21 + 2x22 − 0.3 cos(3π x1 )(4π x2 ) + 0.3 [ −100,100] 0
f10 Bohachevsky3 MN 2 f (x) = x21 + 2x22 − 0.3 cos(3π x1 + 4π x2 ) + 0.3 [ −100,100] 0
5 5
∑ ∑
f11 Shubert MN 2 f (x) = ( i cos(i + 1)x1 + i)( i cos((i + 1)x2 + i)) [ −10,10] −186.73
i=1 i=1

Table 2
Benchmark test functions: Four, five and ten-dimensional.
fNo. Name Type D Formulation Range fmin
f (x) = 100(x21 − x22 )2 + (x1 − 1)2 + (x3 − 1)2 + 90(x23 − x4 )2
f12 Colville UN 4 [ −10,10] 0
+10.1(x2 − 1)2 + (x4 − 1)2 + 19.8(x2 − 1)(x4 − 1)
D
∑
f13 Michalewicz5 MS 5 f (x) = − sin(xi )(sin(ix2i /π ))20 [0, π ] −4.6877
i=1
D D D
∑ ∑ ∑
f14 Zakharov UN 10 f (x) = x2i + ( 0.5ixi )2 + ( 0.5ixi )4 [ −5,10] 0
i=1 i=1 i=1
D
∑
f15 Michalewicz10 MS 10 f (x) = − sin(xi )(sin(ix2i /π ))20 [0, π ] −9.6602
i=1

Table 3
Benchmark test functions: Thirty-dimensional.
fNo. Name Type D Formulation Range fmin
D
∑
f16 Step US 30 f (x) = (xi + 0.5)2 [−5.12,5.12] 0
i=1
D
∑
f17 Sphere US 30 f (x) = x2i [ −100,100] 0
i=1
D
∑
f18 Sum squares US 30 f (x) = ix2i [ −10,10] 0
i=1
D
∑
f19 Quartic US 30 f (x) = ix4i + Rand [ −1.28,1.28] 0
i=1
D D
∑ ∏
f20 Schwefel 2.22 UN 30 f (x) = |xi | + | xi | [ −10,10] 0
i=1 i=1
D
i
∑∑
f21 Schwefel 1.2 UN 30 f (x) = ( xj )2 [ −100,100] 0
i=1 j=1
D−1
∑
f22 Rosenbrock MN 30 f (x) = 100(xi+1 − x2i )2 + (xi − 1)2 [ −30,30] 0
i=1
D
∑
f23 Dixon-Price UN 30 f (x) = (x1 − 1)2 + i(2x2i − xi − 1) [ −10,10] 0
i=1
D
∑
f24 Rastrigin MS 30 f (x) = (x2i − 10 cos(2π xi ) + 10) [−5.12,5.12] 0
i=1 ( ) ( D
D ( ))
1 ∑ ∏ xi − 100
f25 Griewank MN 30 f (x) = (xi − 100)2
√ +1 − cos [−600,600] 0
4000 i
i=1

 D ⎛ ⎞ i=1 ( D )
1 ∑ 1∑
f26 Ackley MN 30 f (x) = −20 exp ⎝−0.2√ x2i ⎠ − exp cos(2π xi ) + 20 + e [−32,32] 0
n n
i=1 i=1

faster rate; thus leading to premature convergence. To select a suit- trials. Minimum mean fitness was obtained at jr = 0.4. Therefore,
able jr value, we conducted a trial on function f22 . The jr jumping we set the jr value at 0.4.
rate varied from 0.1 to 0.9 for step 0.1 while local search limit K was The next trial used function f22 to set the value for local search
fixed at 1. We tested QOCSOS by using function f22 for each value of limit, K. For this experiment, we varied K ’s value from 2−30 with
jr for 100 independent trials. Fig. 6 shows mean results for all 100 a step of 2 while jr remained fixed at 0.4. We ran QOCSOS for 100
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 573

fitness function evaluations: SOS/QOCSOS per Eq. (11).

AR = FESOS /FEQOCSOS (11)

Tables 4–6 show AR values for all trials. If AR >1, convergence

for QOCSOS was faster than SOS convergence. As observed, most
values of AR were greater than 1. Moreover, the average AR value
was 1.79, indicating QOCSOS convergence speed was 79% faster.
In addition to the AR, we also calculated the convergence time
for SOS and QOCSOS. From Tables 4–6, QOCSOS took less con-
vergence time as compared to SOS for functions: f2−4 , f7 , f9−10 ,
f13−15 , f17−18 , f20−21 , and f 24−25 . The convergence times of SOS and
QOCSOS were almost the same for functions f1 , f5−6 , f8 , f11 , and f16 .
Fig. 6. Effect of the parameter jumping rate (jr ).
For remaining functions (f12 , f19 , f22−23 , and f26"− ), the convergence
times of QOCSOS were slower than those of SOS. It is noted that
both SOS and QOCSOS consumed the same fitness function eval-
uations (500,000) for these functions, however, QOCSOS obtained
better results.
Convergence characteristics for QOCSOS and SOS are seen in
Figs. 8 and 9. QOCSOS converged on optimal solutions sooner than
SOS.

4.5. Robustness and non-parametric test setup

Metaheuristic methods usually employ random variables to

initialize the search for an optimal solution but do not guarantee
the same solution for each run. Hence, performances are best
judged by obtaining results from multiple runs. We measured
Fig. 7. Effects for parameter local search limit K with jr set at 4. robustness with 100 independent trial runs to calculate the mean
and the standard deviation. However, these metrics only evaluate
generally the performance of the algorithms. To verify whether
separate trials for each K value. Fig. 7 shows mean fitness results the results of QOCSOS are significantly superior to the results of
for all 100 trials. The proposed QOCSOS produced the best result at SOS, we performed a statistical test. In this study, a non-parametric
K = 20. Thus, we set K at 20. Wilcoxon rank-sum test was carried out at a significance level
of 5%. A p-value < 0.05 points and h = 1 indicate a significant
difference between two solution sets of the SOS and QOCSOS
4.3. Results for benchmark test functions
algorithms, and the results of QOCSOS is better. h = 0 means there
is no significant difference between two algorithms.
Tables 4–6 show the outcomes for SOS and QOCSOS for bench-
Table 7 shows results of Wilcoxon rank-sum test for functions
mark test functions. These include mean value of fitness function,
f11 , f12 , f15 , f16 , f19 , f22 , and f26
standard deviation (SD), minimum number of fitness evaluations
(FE) and convergence time, respectively. Two-dimensional bench- while Figs. 10 and 11 show QOCSOS vs. SOS distributions for
mark test function results (f1 – f11 ) are listed in Table 4. Both fitness values as a measure of robustness for these functions. All
algorithms obtained the same global minimum but the SD obtained figures show that QOCSOS provided better quality solutions indi-
by QOCSOS proved better than SOS for f11 . For all functions tested, cating a more effective performance in terms of a global search
results from QOCSOS also converged to optimum with a lower capability for optimized solutions. For Wilcoxon rank-sum test, the
number of fitness function evaluations than for SOS. results of QOCSOS dominates the results of SOS for functions f12 , f15 ,
Results for four, five and ten-dimensional test functions (f12 – f22 , and f26 .
f15 ), are seen in Table 5. Mean values obtained by QOCSOS were
better than SOS for functions f12 and f15 . Both algorithms produced 5. solution of ODGP in radial distribution networks using THE
the same global minimum for functions f13 and f14 . In terms of proposed QOCSOS algorithm
convergence, QOCSOS required less fitness function evaluations to
converge to optimum compared to SOS for all functions, excepting We used QOCSOS to solve an Optimal Distributed Generation
f12 . Placement (ODGP) problem for a real-world optimization problem
As for thirty-dimensional test functions (f16 – f26, Table 6), in power system. The problem formulation directed us to identify
optimization outcomes for QOCSOS and SOS were equal for f17 , f18 , optimal locations and sizes for DG unit installations in an existing
f20 , f21 , f24 , and f25 , showing the same global minimums. However, distribution network while minimizing objective functions and
QOCSOS obtained better mean value fitness results for all remain- satisfying all operating constraints within the system.
ing functions. Regarding convergence rate, QOCSOS required the
least number of fitness function evaluations compared to SOS for 5.1. ODGP Problem formulation
the following functions: f16 , f17 , f18 , f20 , f21 , f 24 , and f 25 .
The ODGP problem formulation required the minimization of
4.4. Convergence study
real power loss (PL ) within three distribution networks described
as follows:
We compared convergence rate outcomes by calculating the
acceleration rate (AR) as a ratio between the number of required OF = min(PL ) (12)
574 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

Table 4
Comparison of Results: SOS vs QOCSOS for two-dimensional benchmark functions.
fNo. SOS QOCSOS SOS QOCSOS AR
Mean SD Mean SD FE Time(s) FE Time(s)
f1 0 0 0 0 32,585 0.24 30,197 0.24 1.08
f2 −1 0 −1 0 13,077 0.09 9813 0.08 1.33
f3 0 0 0 0 152,593 1.10 67,197 0.52 2.27
f4 0 0 0 0 5517 0.04 2197 0.02 2.51
f5 0 0 0 0 24,985 0.18 23,997 0.18 1.04
f6 −1.8013 2.90E−15 −1.8013 2.90E−15 10,701 0.08 10,001 0.08 1.07
f7 0 0 0 0 16,629 0.12 2,469 0.02 6.74
f8 −1.03163 1.56E−15 −1.03163 1.56E−15 13,705 0.10 12,309 0.10 1.11
f9 0 0 0 0 5141 0.04 2065 0.02 2.49
f10 0 0 0 0 7137 0.05 2797 0.02 2.55
f11 −186.731 9.97E−08 −186.731 3.86E−13 42,253 0.32 39,193 0.32 1.08

Table 5
Comparison of results: SOS vs QOCSOS for four, five and ten-dimensional functions.
fNo. SOS QOCSOS SOS QOCSOS AR
Mean SD Mean SD FE Time(s) FE Time(s)
f12 9.76E−20 3.71E−19 1.43E−23 5.46E−23 500,000 3.73 500,000 4.00 1
f13 −4.68766 3.40E−15 −4.68766 3.40E−15 36,381 0.31 25,001 0.23 1.46
f14 0 0 0 0 320,081 2.68 294,397 2.59 1.09
f15 −9.65943 0.003933 −9.65981 0.001259 142,929 1.36 78,789 0.82 1.81

Table 6
Results: SOS vs QOCSOS for thirty-dimensional functions.
fNo. SOS QOCSOS SOS QOCSOS AR
Mean SD Mean SD FE Time(s) FE Time(s)
f16 2.10E−33 5.07E−33 1.66E−33 4.92E−33 93,421 0.80 87,057 0.81 1.07
f17 0 0 0 0 233,817 2.23 121,781 1.29 1.92
f18 0 0 0 0 232,469 2.11 119,997 1.21 1.94
f19 9.24E−05 3.39E−05 1.42E−05 9.72E−06 500,000 6.33 500,000 7.10 1
f20 0 0 0 0 231,973 2.59 119,105 1.53 1.95
f21 0 0 0 0 234,217 4.27 121,597 2.94 1.93
f22 0.757135 0.925751 0.000452 0.002322 500,000 4.35 500,000 4.71 1
f23 0.666667 3.09E−14 0.666667 8.67E−16 500,000 4.31 500,000 4.66 1
f24 0 0 0 0 15,393 0.14 5597 0.05 2.75
f25 0 0 0 0 14,217 0.14 5797 0.06 2.45
f26 3.38E−15 1.64E−15 8.88E−16 0 500,000 4.51 500,000 4.90 1

Table 7
Results of Wilcoxon rank-sum test.
f11 f12 f15 f19 f22 f26
p-value 0.416736652 4.81E−34 0.045754599 0.167055521 5.17E−33 2.80E−34
h 0 1 1 0 1 1

where PD,i and QD,i are active and reactive power load demands at the ith
NL
∑ bus, respectively; PL,j and QL,j are active and reactive power losses
PL = Rk Ik2 (13) at the jth branch; PDG,I is the active power output of the ith DG unit.
k=1
(2) Voltage limits:
where NL is the number of branches in a distribution network; Rk is
the resistance of the kth branch; and Ik is current passing through Vimin ≤ Vi ≤ Vimax ; i = 1, . . . , NB (16)
the kth branch.
where Vi is voltage at the ith bus; and Vimin Vimax
are minimum and
Constraints considered in this study were as follows:
maximum acceptable bus voltage levels, respectively.
(1) Power balance constraints:
(3) Power flow constraint:
NDG NB NL
|PFk | < PFkmax
∑ ∑ ∑
Pslack + PDG,i = PD,i + PL,j (14) (17)
i=1 i=1 j=1
where PF k is the power flow in the ith branch; is the maximum
NB NL
∑ ∑ limit of power flow in the ith branch.
Qslack = QD,i + QL,j (15)
i=1 j=1 (4) DG capacity limits:

where N DG is the total number of DG units; N B is the number PDGmin,i ≤ PDG,i ≤ PDGmax,i ; i = 1, . . . , NDG (18)
of total buses in a distribution network; N L is the number of
transmission line in a distribution network; Pslack and Qslack are where PDGmin,i and PDGmax,i are minimum and maximum real power
active and reactive power supplies from a slack bus, respectively; outputs of the ith DG unit, respectively.
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 575

Fig. 8. Convergence profiles: QOCSOS vs SOS for functions f3 , f4 , f7 , f10 , f12 and f14 – f17 .

Table 8
Data for test systems and constraints.
Item 33-busstesttsystem [55] 69-busstest system [56] 118-busstest system [57]
3.72 MW & 3.715 MW & 22.709 MW &
Power load demand
2.3 MVAr 2.300 MVAr 17.041 MVar
210.998 kW & 225.0005 kW & 1298.092 kW &
Power system loss
143.033 kVAr 102.165 kVAr 978.736 kVAr
Base voltage 12.66 kV 12.66 kV 11 kV
Max power flow 3925.99 kW(#L1−2 ) 4027.1 kW (#L1−2 ) 10,677.9 kW (#L1−2 )
Max limit of power flow 4200 kW 4500 kW 11,000 kW
Bus voltage constraint 0.9 ≤ |Vi | ≤ 1.05
Number of DG units to 3 3 7
be installed
Allowable DG range 0–3 MW

5.2. Simulation results (TLBO [40], QOTLBO [40], SIMBO-Q [41], QOSIMBO-Q [41]). Opti-
mal locations for DG installation were found to at buses 13, 24, and
Both QOCSOS and SOS algorithms were used to find optimal 30 with optimal sizes of 801.7, 1091.3 and 1053.6 kW, respectively.
locations and sizes for DG units in 33, 69 and 118-bus radial After installation, power loss was reduced from 210.998 kW to
distribution networks with a view to minimize real power loss. 72.788 kW (65.5%). Both QOCSOS and SOS algorithms obtained
Table 8 shows data for all test systems, including constraints. We equal power loss results and proved slightly better than all other
methods. However, QOCSOS provided faster convergence speed
ran 100 independent trials for both algorithms with a maxFE of
(Fig. 12).
12,000 for each run. Optimized results are reported in the following
sub-sections.
5.2.2. 69-bus test system
Table 10 shows results from QOCSOS, SOS, TLBO [40], QOTLBO
5.2.1. 33-bus test system [40], SIMBO-Q [41] and QOSIMBO-Q [41] the 69-bus radial distri-
Table 9 shows results for the 33-bus radial distribution net- bution network. QOCSOS provided optimal locations at buses 11,
work, including outcomes for QOCSOS, SOS and other methods 18 and 61 with corresponding optimal DG sizes of 526.8, 380.4,
576 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

Fig. 9. Convergence profiles: QOCSOS vs. SOS for functions (f18 – f26 ).

Fig. 10. Distribution of fitness values: QOCSOS vs. SOS for functions f11 , f12 , and f15 .
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 577

Fig. 11. Distribution of fitness values: QOCSOS vs. SOS for functions f16 , f19 , f22 , and f26 .

Table 9
Results for the 33-bus system.
Methods Bus DG size PL (kW) % loss
(MW) reduction
TLBO [40] 10 0.8246 75.54 64.20
24 1.0311
31 0.8862
QOTLBO [40] 12 0.8808 74.1 64.88
24 1.0592
29 1.0714
SIMBO-Q [41] 24 0.7613 73.6 65.11
25 0.8657
30 1.1070
QOSIMBO-Q [41] 24 1.0906 72.8 65.49
30 1.0542
13 0.8016
SOS 13 0.8017 72.7869 65.50
24 1.0913
30 1.0536
QOCSOS 13 0.8017 72.7869 65.50
24 1.0913
30 1.0536

and 1719 kW for installed DG units. After DG installation, power optimized real power loss. Fig. 13 shows convergence results for
loss was reduced to 69.4284 kW (69.14%). As in the Table 10, QOC- QOCSOS and SOS and clearly demonstrates speedier convergence
SOS and SOS outperformed other algorithms and obtained equal by QOCSOS.
578 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

Table 10
Results for 69-bus system.
Methods Bus DG size PL (kW) % loss
(MW) reduction
TLBO [40] 15 0.5919 72.406 67.82
61 0.8188
63 0.9003
QOTLBO [40] 18 0.5334 71.625 68.17
61 1.1986
63 0.5672
SIMBO-Q [41] 61 1.5000 71.4 68.27
17 0.4285
67 0.4863
QOSIMBO-Q [41] 9 0.8314 71.0 68.44
17 0.4538
61 1.5000
SOS 11 0.5268 69.4284 69.14
18 0.3804
61 1.7190
QOCSOS 11 0.5268 69.4284 69.14
18 0.3804
61 1.7190

Fig. 12. Convergence profiles: QOCSOS vs. SOS for 33-bus system. Fig. 14. Convergence profiles: QOCSOS vs. SOS for 119-bus Test System.

5.2.4. Effect of Eco-size

We varied Eco-size from 10 to 50 to study its impact on out-
comes. Table 12 shows results obtained for QOCSOS vs. SOS where
an Eco-size of 20 proved sufficient to deal with ODGP problem
solving. Moreover, average power losses obtain by QOCSOS were
better than those obtained by SOS.

6. Engineering design problems

To further investigate the performance of the proposed QOC-

SOS algorithm, we used QOCSOS to solve five engineering design
problems which are real-world optimization problems in struc-
tural engineering field. For parameter setting, QOCSOS used 20
Fig. 13. Convergence profiles: QOCSOS vs. SOS for 69-bus Test System. organisms for all cases. The parameters jr and K are set to 0.4
and 20 respectively. The parameter maxFE is chosen depend on
the complexity of each problem. The results obtained by QOCSOS
5.2.3. 118 bus test system were compared to the results from the original SOS algorithm and
Table 11 shows results for a large scale 118-bus radial distribu- other optimization methods in the literature. Both QOCSOS and
tion system, including outcomes for QOCSOS, SOS, TLBO [40] and SOS were run 100 independent trials. The results of QOCSOS and
QOTLBO [40]. Optimal locations obtained by QOCSOS were at buses SOS for these engineering design problems are shown in Appendix
20, 42, 50, 72, 80, 96, and 110. DG units installed at those buses (Table 18). Optimized results of each problems are reported in the
had optimal sizes of 1817.6, 1276.4, 2767, 2533.3, 2095, 1663.1 following sub-sections.
and 2869.3 kW, respectively. After DG installation, power loss was
reduced to 516.4187 kW (60.22%) by QOCSOS and SOS equally, 6.1. Tension/compression spring design problem
showing a significantly superior performance. Fig. 14 shows con-
vergence curves for QOCSOS vs SOS, again demonstrating a quicker In the tension/compression spring design problem, QOCSOS
convergence capacity for QOCSOS. optimized three design variables: wire diameter (d), mean coil
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 579

Table 11
Result: 118-bus system tests.
Methods Bus DG size PL (kW) % loss
(MW) reduction
TLBO [40] 8 1.7553 590.6974 54.49
10 0.5910
36 1.5368
49 2.6865
71 2.5014
79 2.4941
110 2.6628
QOTLBO [40] 24 1.2463 576.1823 55.61
42 0.7322
47 3.5392
74 2.6792
78 1.2483
94 1.0865
108 3.2432
SOS 20 2.0956 516.4188 60.22
42 1.8185
50 1.6613
72 1.2757
80 2.7679
96 2.5338
110 2.8691
QOCSOS 20 1.8176 516.4187 60.22
42 1.2764
50 2.7670
72 2.5333
80 2.0950
96 1.6631
110 2.8693

Table 12
Results from Eco_size variance: QOCSOS vs. SOS.
Test system Eco-size SOS QOCSOS
Min. PL (kW) Avg. PL (kW) Max. PL (kW) Std. PL (kW) Min. PL (kW) Avg. PL (kW) Max. PL (kW) Std. PL (kW)
33-bus 10 72.7869 72.8869 78.4538 0.7035 72.7869 72.7879 72.7910 0.0018
20 72.7869 72.7871 72.7910 0.0011 72.7869 72.7869 72.7910 0.0007
30 72.7869 72.7876 72.8086 0.0025 72.7869 72.7871 72.7910 0.0010
40 72.7869 72.7978 73.2087 0.0533 72.7869 72.7870 72.7910 0.0009
50 72.7869 72.8190 73.8910 0.1388 72.7869 72.7875 72.8196 0.0035
69-bus 10 69.4284 69.4633 70.1662 0.1482 69.4284 69.5023 70.1605 0.2129
20 69.4284 69.4414 70.1605 0.0817 69.4284 69.4311 69.6936 0.0265
30 69.4284 69.4428 69.7023 0.0565 69.4284 69.4370 69.6945 0.0456
40 69.4284 69.4556 69.9846 0.0938 69.4284 69.4339 69.6954 0.0363
50 69.4284 69.4467 70.1605 0.0812 69.4284 69.4444 69.7254 0.0593
118-bus 10 516.4187 523.0863 544.9740 6.6816 516.4187 522.2628 551.0279 6.0681
20 516.4188 522.3910 552.8007 6.0671 516.4187 519.6234 536.7695 4.4125
30 516.4415 529.5661 564.2080 11.027 516.4191 521.6004 540.2678 5.9672
40 516.6930 542.5124 572.8994 13.754 516.5337 529.6025 560.3447 8.7792
50 521.7607 547.3204 571.5653 12.199 517.4860 536.7665 560.3126 9.4033

diameter (D), and the number of active coils (N). The objective of
this problem is to minimize the weight of a spring as in Fig. 15. The
mathematical formulation of this problem is expressed as follows:

Consider
−
→ x = [x1 x2 x3 ] = [dDN ] (19)
Minimize
−
→
f ( x ) = (2 + x )x x2 (20)
3 2 1

−
→ x32 x3 Fig. 15. Tension/compression spring design.
Subject to g1 ( x ) = 1 − ≤0 (21)
71,785x41
−
→ 4x22 − x1 x2 1
g2 ( x ) = + ≤ 0 (22) 0.25 ≤ x2 ≤ 1.3 (26)
12,566(x2 x31 − x41 ) 5108x21
−
→ 140.45x1 2 ≤ x3 ≤ 15 (27)
g3 ( x ) = 1 − ≤0 (23)
x22 x3 The result of QOCSOS was reported and compared to results
−
→ x1 + x2 of SOS, HS [58], CDE [59], CPSO [60], GWO [17] and MFO [61]
g4 ( x ) = −1≤0 (24)
1.5 in Table 13. The comparison results show that QOCSOS provided
Variable range 0.05 ≤ x1 ≤ 2 (25) minimum weight of the spring as compared to other methods.
580 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

Table 13
Comparison results for the tension/compression spring design problem.
HS [58] CDE [59] CPSO [60] GWO [17] MFO [61] SOS QOCSOS
x1 0.051154 0.051609 0.051728 0.05169 0.051994457 0.051760679 0.055130816
x2 0.349871 0.354714 0.357644 0.356737 0.36410932 0.358426195 0.444775663
x3 12.076432 11.410831 11.244543 11.28885 10.868421862 11.19023115 7.544716912
Min. weight 0.0126706 0.0126702 0.0126747 0.012666 0.0126669 0.012666368 0.012665282

6.2. Welded beam design problem

In this problem, QOCSOS optimized four design parameters:

thickness of weld (h), length of attached part of bar (l), the height
of the bar (t), and thickness of the bar (b) with the objective of
minimizing the fabrication cost of a welded beam shown in Fig. 16.
The mathematical formulation of welded beam design problem is
described as follows:

Consider:
−
→ x = [x1 x2 x3 x4 ] = [hltb] (28)
Minimize:
−
→
f ( x ) = 1.10471x21 x2 + 0.04811x3 x4 (14 + x2 ) (29)
Subject to:
−
→ −
→
g ( x ) = τ( x ) − τ ≤0 (30)
1 max
−
→ −
→
g2 ( x ) = σ ( x ) − σmax ≤ 0 (31)
−
→
g3 ( x ) = x1 − x4 ≤ 0 (32) Fig. 16. Welded beam design.
−
→
g4 ( x ) = 0.10471x21 + 0.04811x3 x4 (14 + x2 )
−5≤0 (33)
−
→
g5 ( x ) = 0.125 − x1 ≤ 0 (34)
−
→ −
→
g ( x ) = δ( x ) − δ ≤0 (35)
6 max
−
→ −→
g7 ( x ) = P − Pc ( x ) ≤ 0 (36)
√
x2
where τ (−
→x ) = (τ ′ )2 + 2τ ′ τ ′′ + (τ ′′ )2 (37)
2R
P MR ( x2 )
τ′ = √ , τ ′′ = ,M = P L + (38)
2x x J 2
√ 1 2
2
( )2 Fig. 17. Pressure vessel design.
x2 x1 + x3
R= + (39)
4 2
{ [ )2 ]}
√ x22
(
x1 + x3 thickness of the head (Th ), the inner radius (R), the length of the
J =2 2x1 x2 + (40) cylindrical section without considering the head (L). The formula-
4 2
tion of this optimization problem can be expressed as follows:
6PL 6PL3
σ −
( →) (−
→ −
→
, δ
)
x = x = (41) Consider: x = [x1 x2 x3 x4 ] = [Ts Th RL] (45)
x4 x23 Ex23 x4 −
→
√
2 6
Minimize: f ( x ) = 0.6224x1 x3 x4 + 1.7781x2 x23
x3 x4 ( √ )
−
→ 4.013E 36 x3 E + 3.1661x21 x4 + 19.84x21 x3 (46)
Pc ( x ) = 2
1− (42)
L 2L 4G
Subject to:
−
→
g1 ( x ) = −x1 + 0.0193x3 ≤ 0 (47)
where P = 6000 lb, L = 14 in, E = 30 × 10 psi, G = 12 × 10 psi
6 6 −
→
g ( x ) = −x + 0.00954x ≤ 0 (48)
2 2 3

τmax = 13,600 psi, σmax = 30,000 psi, σmax = 0.25 in −

→ 4
g3 ( x ) = −π x23 x4 − π x33 + 1, 296, 000 ≤ 0
3
Variable range: 0.1 ≤ x1 , x4 ≤ 2 (43) (49)
−
→
g4 ( x ) = x4 − 240 ≤ 0 (50)
0.1 ≤ x2 , x3 ≤ 10 (44)
Variable range: 0 ≤ x1 , x2 ≤ 99 (51)
Table 14 shows results of QOCSOS, SOS, CPSO [60], GA [62], GSA
[61] and GWO [17] for welded beam design problem. As in the 10 ≤ x3 , x4 ≤ 200 (52)
Table 14, QOCSOS obtained minimum fabrication cost as compared
The results of QOCSOS, SOS, MFO [61], CSA [63], CPSO [60] and
to other reported optimization methods. ACO [64] for the pressure vessel design problem are reported in
Table 15. It can be seen that QOCSOS obtained a result superior to
6.3. Pressure vessel design problem all other methods.

The objective of the pressure vessel design problem is to min- 6.4. Cantilever beam design problem
imize the total cost including welding cost, material and forming
cost of a cylindrical pressure vessel illustrated in Fig. 17. There are This optimization problem is to minimize the weight of a can-
four variables to be optimized: the thickness of the shell (Ts ), the tilever beam with five square blocks in Fig. 18. The beam is fixed at
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 581

Table 14
Comparison results for the welded beam design problem.
CPSO [60] GA [62] GSA [61] GWO [17] SOS QOCSOS
x1 0.202369 0.1829 0.182129 0.205676 0.20572965 0.20572943
x2 3.544214 4.0483 3.856979 3.478377 3.47048871 3.47050509
x3 9.048210 9.3666 10.0000 9.03681 9.03662371 9.03662141
x4 0.205723 0.2059 0.202376 0.205778 0.20572965 0.20573002
Min. cost 1.73148 1.82420 1.87995 1.72624 1.7248523963 1.7248523440

Table 15
Comparison results for the pressure vessel design problem.
MFO [61] CSA [63] CPSO [60] ACO [64] SOS QOCSOS
x1 0.8125 0.8125 0.8125 0.8125 0.778169 0.778238
x2 0.4375 0.4375 0.4375 0.4375 0.384649 0.384893
x3 42.098445 42.0984456 42.091266 42.103624 40.319619 40.322081
x4 176.636596 176.6365958 176.746500 176.572656 199.999999 199.966711
Min. cost 6059.7143 6059.7143348 6061.0777 6059.0888 5885.332777 5885.332774

−
→ 397.5
g2 ( x ) = −1≤0 (60)
x1 x22 x23
−
→ 1.93x34
g3 ( x ) = −1≤0 (61)
x2 x3 x46
−
→ 1.93x35
g4 ( x ) = −1≤0 (62)
x2 x3 x47
√
( )2
−
→ 1 745x4
g5 ( x ) = + 16.9 × 106
110x36 x2 x3

Fig. 18. Cantilever beam design.

−1≤0 (63)
√
( )2
−
→ 1 745x5
g6 ( x ) = + 157.5 × 106
85x37 x2 x3
the first block and carries a vertical load at the fifth one. The design
variables are the heights (or widths) of five different square blocks. −1≤0 (64)
The problem formulation of cantilever beam design is described as −
→ x2 x3
g7 ( x ) = −1≤0 (65)
follows: 40
Consider
−
→ x = [x1 x2 x3 x4 x5 ] (53) −
→ 5x2
g8 ( x ) = −1≤0 (66)
−
→ x1
Minimize f ( x ) = 0.0624(x1 + x2 + x3 + x4 + x5 ) (54) x1
−
→
g9 ( x ) = −1≤0 (67)
−
→ 61 27 19 7 1 12x2
Subject to: g( x ) = 3 + 3 + 3 + 3 + 3 − 1 (55)
x1 x2 x3 x4 x5 −
→ 1.5x6 + 1.9
g10 ( x ) = −1≤0 (68)
Variable range 0.01 ≤ x1 , x2 , x3 , x4 , x5 ≤ 100 (56) x4
−
→ 1.1x7 + 1.9
Table 16 shows the comparison result between QOCSOS and g11 ( x ) = −1≤0 (69)
other methods such as: SOS, MMA(II), GCA(II), CSA, and MFO. As x5
in Table 16, QOCSOS achieved a best solution as compared to other Variable range 2.6 ≤ x1 ≤ 3.6 (70)
methods.
0.7 ≤ x2 ≤ 0.8 (71)
6.5. Speed reducer design problem 17 ≤ x3 ≤ 28 (72)
7.3 ≤ x4 , x5 ≤ 8.3 (73)
QOCSOS was applied to solve the design of a speed reducer
design optimization problem in Fig. 19. There are seven parameters 2.9 ≤ x6 ≤ 3.9 (74)
to be optimized: the face width (b), module of teeth (m), number of 5 ≤ x7 ≤ 5.5 (75)
teeth on pinion (z), length of shaft 1 between bearings (l1 ), length of
shaft 2 between bearings (l2 ), diameter of shaft 1 (d1 ), and diameter Table 17 presents a comparison of the results of QOCSOS, SOS,
of shaft 2 (d2 ). The objective is to minimize the total weight of the MDHGA [66], CSA [63], IPHS [67]. QOCSOS and SOS obtained the
speed reducer. The mathematical formulation of this problem is as same minimum weight of the speed reducer design problem,
follows: which is better solution than those from other reported methods
−
→ in Table 17.
Consider x = [x1 x2 x3 x4 x5 x6 x7 ] = [bmzl1 l2 d1 d2 ] (57)
−
→
f ( x ) = 0.7854x x2 (3.3333x2
1 2 3 7. Conclusion
+14.9334x3 − 43.0934)
Minimize (58)
−1.508x1 (x26 + x27 ) + 7.4777(x36 + x37 ) This paper presented an improved version of the SOS algorithm
+0.7854(x4 x26 + x5 x27 ) called Quasi-Oppositional Chaotic Symbiotic Organisms Search
−
→ 27 (QOCSOS). In QOCSOS, the QOBL strategy provided for better explo-
Subject to g1 ( x ) = −1≤0 (59)
x1 x22 x3 ration capacity because it enhanced the quality of initial solution
582 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583

Table 16
Comparison results for the cantilever beam design problem.
MMA(II) [65] GCA(II) [65] CSA [63] MFO [61] SOS QOCSOS
x1 6.0100 6.0100 6.0089 5.9848717732166 6.02299643 6.013069155
x2 5.3000 5.3000 5.3049 5.31672692429783 5.306679042 5.294432112
x3 4.4900 4.4900 4.5023 4.49733258583062 4.494979932 4.513527407
x4 3.4900 3.4900 3.5077 3.51361646768954 3.499628184 3.493453969
x5 2.1500 2.1500 2.1504 2.16162029338550 2.14941363 2.159530941
Min. weight 1.3400 1.3400 1.33999 1.33998808597181 1.339973282 1.339956387

Table 17
Comparison results of the pressure vessel design problem.
MDHGA [66] CSA [63] IPHS [67] SOS QOCSOS
x1 3.5 3.5015 3.5 3.5 3.5
x2 0.7 0.7000 0.7 0.7 0.7
x3 17 17.0000 17 17 17
x4 7.3 7.6050 7.3 7.3 7.3
x5 7.8 7.8181 7.71533233833903 7.7153199115 7.7153199115
x6 3.36 3.3520 3.35021510925684 3.3502146661 3.3502146661
x7 5.29 5.2875 5.28666403545462 5.2866544650 5.2866544650
Min. weight 3000.83 3000.9810 2994.477531 2994.47106615 2994.47106615

Table 18
Results of QOCSOS vs. SOS for engineering design problem.
Sections Methods Minimum Average Maximum Std. maxFE
SOS 0.0126663675 0.0126960824 0.0130047981 4.12E−05 16,000
6.1
QOCSOS 0.0126652823 0.0126826827 0.0128109294 2.65E−05 16,000
SOS 1.7248523963 1.7248707938 1.7258280755 1.04E−04 12,000
6.2
QOCSOS 1.7248523440 1.7248540792 1.7248694364 2.33E−06 12,000
SOS 5885.3327773366 5895.0450245940 6075.3073530963 26.6817 30,000
6.3
QOCSOS 5885.3327736012 5885.8333884916 5893.1729085511 1.2288 30,000
SOS 1.3399587065 1.3399732823 1.3400171021 1.07E−05 16,000
6.4
QOCSOS 1.3399563873 1.3399568342 1.3399589159 4.04E−07 16,000
SOS 2994.471066146 2994.471066146 2994.471066146 0 16,000
6.5
QOCSOS 2994.471066146 2994.471066146 2994.471066146 0 16,000

applicability of QOCSOS for complex real-world optimization prob-

lems.

Acknowledgement

This research work is supported by

Universiti Teknologi PETRONAS, Malaysia’ Graduate Assistant
Scheme.

Appendix

Fig. 19. Speed reducer design. See Table 18.

References
sets and diversified search direction when generating new solu-
[1] J.H. Holland, Adaptation in Natural and Artificial Systems, MIT Press, 1992.
tions. In addition, the CLS strategy pushed the search process from
[2] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Neural Networks,
within a current best population set to offer better exploitation 1995. Proceedings. IEEE International Conference on, vol. 4, 1995, pp. 1942–
capacity. Thanks to the integration of both strategies, QOCSOS per- 1948.
[3] S. Kirkpatrick, C.D. Gelatt, M.P. Vecchi, Optimization by simulated annealing,
formance greatly improved the original SOS algorithm in terms of
Science 220 (1983) 671–680.
solution quality and convergence speed. We validated the claim by [4] R. Storn, K. Price, Differential evolution – A simple and efficient heuristic for
testing QOCSOS with different types and dimensions for twenty-six global optimization over continuous spaces, J. Global Optim. 11 (1997) 341–
mathematical benchmark functions. QOCSOS obtained the theoret- 359.
[5] M. Dorigo, M. Birattari, T. Stutzle, Ant colony optimization, IEEE Comput.
ical global minima for fifteen test functions and reached very closer Intell. Mag. 1 (2006) 28–39.
to the theoretical global minima for remaining test functions. It is [6] D. Karaboga, B. Basturk, A powerful and efficient algorithm for numerical
observed from the convergence study that QOCSOS achieved the function optimization: artificial bee colony (ABC) algorithm, J. Global Optim.
39 (2007) 459–471.
global optimum solutions with remarkable convergence behaviour
[7] D.H. Wolpert, W.G. Macready, No free lunch theorems for optimization, IEEE
compared to the original SOS. Moreover, we applied the proposed Trans. Evol. Comput. 1 (1997) 67–82.
algorithm successfully to solve for optimal DG placements in ra- [8] G. Zong Woo, K. Joong Hoon, G.V. Loganathan, A new heuristic optimization
algorithm: Harmony search, Simulation 76 (2001) 60–68.
dial distribution networks and five structural design optimization
[9] R.V. Rao, V.J. Savsani, D.P. Vakharia, Teaching–learning-based optimization:
problems. For all application, QOCSOS achieved better solution A novel method for constrained mechanical design optimization problems,
quality than SOS and other methods. This further confirms the Comput. Aided Des. 43 (2011) 303–315.
 K.H. Truong, P. Nallagownden, Z. Baharudin et al. / Applied Soft Computing Journal 77 (2019) 567–583 583

[10] X.-S. Yang, Flower Pollination Algorithm for Global Optimization, Berlin, [40] S. Sultana, P.K. Roy, Multi-objective quasi-oppositional teaching learning
Heidelberg, 2012, pp. 240–249. based optimization for optimal location of distributed generator in radial
[11] W.-T. Pan, A new fruit fly optimization algorithm: Taking the financial distress distribution systems, Int. J. Electr. Power Energy Syst. 63 (2014) 534–545.
model as an example, Knowl.-Based Syst. 26 (2012) 69–74. [41] S. Sharma, S. Bhattacharjee, A. Bhattacharya, Quasi-oppositional swine in-
[12] A.H. Gandomi, A.H. Alavi, Krill herd: A new bio-inspired optimization algo- fluenza model based optimization with quarantine for optimal allocation of
rithm, Commun. Nonlinear Sci. Numer. Simul. 17 (2012) 4831–4845. DG in radial distribution network, Int. J. Electr. Power Energy Syst. 74 (2016)
[13] X.S. Yang, D. Suash, Cuckoo search via Lévy flights, in: 2009 World Congress 348–373.
on Nature & Biologically Inspired Computing, NaBIC, 2009, pp. 210–214. [42] D. Guha, P.K. Roy, S. Banerjee, Load frequency control of large scale power
[14] S.S. Pattnaik, K.M. Bakwad, B.S. Sohi, R.K. Ratho, S. Devi, Swine Influenza system using quasi-oppositional grey wolf optimization algorithm, Eng. Sci.
Models Based Optimization (SIMBO), Appl. Soft Comput. 13 (2013) 628–653. Technol. Int. J. 19 (2016) 1693–1713.
[15] A. Sadollah, A. Bahreininejad, H. Eskandar, M. Hamdi, Mine blast algorithm: [43] C.K. Shiva, V. Mukherjee, A novel quasi-oppositional harmony search algo-
A new population based algorithm for solving constrained engineering opti- rithm for automatic generation control of power system, Appl. Soft Comput.
mization problems, Appl. Soft Comput. 13 (2013) 2592–2612. 35 (2015) 749–765.
[16] H. Shareef, A.A. Ibrahim, A.H. Mutlag, Lightning search algorithm, Appl. Soft [44] L.-Y. Chuang, C.-J. Hsiao, C.-H. Yang, Chaotic particle swarm optimization for
Comput. 36 (2015) 315–333. data clustering, Expert Syst. Appl. 38 (2011) 14555–14563.
[17] S. Mirjalili, S.M. Mirjalili, A. Lewis, Grey wolf optimizer, Adv. Eng. Softw. 69 [45] X. Xia, Particle swarm optimization method based on chaotic local search and
(2014) 46–61. Roulette wheel mechanism, Physics Procedia 24 (2012) 269–275.
[18] S. Mirjalili, The ant lion optimizer, Adv. Eng. Softw. 83 (2015) 80–98. [46] O.E. Turgut, Hybrid chaotic quantum behaved particle swarm optimization
[19] S. Mirjalili, A. Lewis, The whale optimization algorithm, Adv. Eng. Softw. 95 algorithm for thermal design of plate fin heat exchangers, Appl. Math. Model.
(2016) 51–67. 40 (2016) 50–69.
[20] S. Mirjalili, S.M. Mirjalili, A. Hatamlou, Multi-verse optimizer: a nature- [47] D. Jia, G. Zheng, M. Khurram Khan, An effective memetic differential evolution
inspired algorithm for global optimization, Neural Comput. Appl. 27 (2016) algorithm based on chaotic local search, Inform. Sci. 181 (2011) 3175–3187.
495–513. [48] Y. He, Q. Xu, S. Yang, A. Han, L. Yang, A novel chaotic differential evolution
[21] A. Askarzadeh, A novel metaheuristic method for solving constrained engi- algorithm for short-term cascaded hydroelectric system scheduling, Int. J.
neering optimization problems: Crow search algorithm, Comput. Struct. 169 Electr. Power Energy Syst. 61 (2014) 455–462.
(2016) 1–12. [49] J. Zhang, S. Lin, W. Qiu, A modified chaotic differential evolution algorithm
[22] F. Fausto, E. Cuevas, A. Valdivia, A. González, A global optimization algorithm for short-term optimal hydrothermal scheduling, Int. J. Electr. Power Energy
inspired in the behavior of selfish herds, Biosystems 160 (2017) 39–55. Syst. 65 (2015) 159–168.
[23] M. Jain, V. Singh, A. Rani, A novel nature-inspired algorithm for optimization: [50] K.K. Bharti, P.K. Singh, Chaotic gradient artificial bee colony for text clustering,
Squirrel search algorithm, Swarm Evol. Comput. (2018). Soft Comput. 20 (2016) 1113–1126.
[24] M.-Y. Cheng, D. Prayogo, Symbiotic organisms search: A new metaheuristic [51] Q.-K. Pan, L. Wang, L. Gao, A chaotic harmony search algorithm for the flow
optimization algorithm, Comput. Struct. 139 (2014) 98–112. shop scheduling problem with limited buffers, Appl. Soft Comput. 11 (2011)
[25] M.-Y. Cheng, D. Prayogo, D.-H. Tran, Optimizing multiple-resources leveling 5270–5280.
in multiple projects using discrete symbiotic organisms search, J. Comput. Civ. [52] X. He, Y. Rao, J. Huang, A novel algorithm for economic load dispatch of power
Eng. 30 (2016) 04015036. systems, Neurocomputing 171 (2016) 1454–1461.
[26] V.F. Yu, A.A.N.P. Redi, C.-L. Yang, E. Ruskartina, B. Santosa, Symbiotic organ- [53] H.R. Tizhoosh, Opposition-based learning: A new scheme for machine intel-
isms search and two solution representations for solving the capacitated ligence, in: International Conference on Computational Intelligence for Mod-
vehicle routing problem, Appl. Soft Comput. 52 (2017) 657–672. elling, Control and Automation and International Conference on Intelligent
[27] G.G. Tejani, V.J. Savsani, V.K. Patel, Adaptive symbiotic organisms search (SOS) Agents, Web Technologies and Internet Commerce, CIMCA-IAWTIC’06, 2005,
algorithm for structural design optimization, J. Comput. Des. Eng. 3 (2016) pp. 695–701.
226–249. [54] J. Ji, S. Gao, S. Wang, Y. Tang, H. Yu, Y. Todo, Self-adaptive gravitational search
[28] D.T.T. Do, J. Lee, A modified symbiotic organisms search (mSOS) algorithm for algorithm with a modified chaotic local search, IEEE Access 5 (2017) 17881–
optimization of pin-jointed structures, Appl. Soft Comput. 61 (2017) 683–699. 17895.
[29] A. Panda, S. Pani, A symbiotic organisms search algorithm with adaptive [55] M.E. Baran, F.F. Wu, Network reconfiguration in distribution systems for loss
penalty function to solve multi-objective constrained optimization problems, reduction and load balancing, IEEE Trans. Power Deliv. 4 (1989) 1401–1407.
Appl. Soft Comput. 46 (2016) 344–360. [56] M.E. Baran, F.F. Wu, Optimal capacitor placement on radial distribution sys-
[30] S. Duman, Symbiotic organisms search algorithm for optimal power flow tems, IEEE Trans. Power Deliv. 4 (1989) 725–734.
problem based on valve-point effect and prohibited zones, Neural Comput. [57] D. Zhang, Z. Fu, L. Zhang, An improved TS algorithm for loss-minimum re-
Appl. 28 (2017) 3571–3585. configuration in large-scale distribution systems, Electr. Power Syst. Res. 77
[31] H. Kamankesh, V.G. Agelidis, A. Kavousi-Fard, Optimal scheduling of renew- (2007) 685–694.
able micro-grids considering plug-in hybrid electric vehicle charging demand, [58] M. Mahdavi, M. Fesanghary, E. Damangir, An improved harmony search
Energy 100 (2016) 285–297. algorithm for solving optimization problems, Appl. Math. Comput. 188 (2007)
[32] S. Verma, S. Saha, V. Mukherjee, A novel symbiotic organisms search al- 1567–1579.
gorithm for congestion management in deregulated environment, J. Exp. [59] F.-z. Huang, L. Wang, Q. He, An effective co-evolutionary differential evolution
Theoret. Artif. Intell. 29 (2017) 59–79. for constrained optimization, Appl. Math. Comput. 186 (2007) 340–356.
[33] S. Saha, V. Mukherjee, Optimal placement and sizing of DGs in RDS using chaos [60] Q. He, L. Wang, An effective co-evolutionary particle swarm optimization for
embedded SOS algorithm, IET Gen. Transm. Distrib. 10 (2016) 3671–3680. constrained engineering design problems, Eng. Appl. Artif. Intell. 20 (2007)
[34] D. Guha, P. Roy, S. Banerjee, Quasi-oppositional symbiotic organism search 89–99.
algorithm applied to load frequency control, Swarm Evol. Comput. 33 (2017) [61] S. Mirjalili, Moth-flame optimization algorithm: A novel nature-inspired
46–67. heuristic paradigm, Knowl.-Based Syst. 89 (2015) 228–249.
[35] E. Çelik, R. Durgut, Performance enhancement of automatic voltage regulator [62] C.A. Coello Coello, Constraint-handling using an evolutionary multiobjective
by modified cost function and symbiotic organisms search algorithm, Eng. Sci. optimization technique, Civ. Eng. Environ. Syst. 17 (2000) 319–346.
Technol. Int. J. 21 (2018) 1104–1111. [63] A.H. Gandomi, X.-S. Yang, A.H. Alavi, Cuckoo search algorithm: a metaheuris-
[36] E. Çelik, N. Öztürk, A hybrid symbiotic organisms search and simulated tic approach to solve structural optimization problems, Eng. Comput. 29
annealing technique applied to efficient design of PID controller for automatic (2013) 17–35.
voltage regulator, Soft Comput. (2018). [64] A. Kaveh, S. Talatahari, An improved ant colony optimization for constrained
[37] E. Çelik, N. Öztürk, First application of symbiotic organisms search algorithm engineering design problems, Eng. Comput. 27 (2010) 155–182.
to off-line optimization of PI parameters for DSP-based DC motor drives, [65] H. Chickermane, H.C. Gea, Structural optimization using a new local approxi-
Neural Comput. Appl. 30 (2018) 1689–1699. mation method, Internat. J. Numer. Methods Engrg. 39 (1996) 829–846.
[38] D.-H. Tran, M.-Y. Cheng, D. Prayogo, A novel Multiple Objective Symbiotic [66] S.S. Rao, Y. Xiong, A hybrid genetic algorithm for mixed-discrete design
Organisms Search (MOSOS) for time–cost–labor utilization tradeoff problem, optimization, J. Mech. Des. 127 (2004) 1100–1112.
Knowl.-Based Syst. 94 (2016) 132–145. [67] M. Jaberipour, E. Khorram, Two improved harmony search algorithms for
[39] S. Rahnamayan, H.R. Tizhoosh, M.M.A. Salama, Quasi-oppositional differential solving engineering optimization problems, Commun. Nonlinear Sci. Numer.
evolution, in: 2007 IEEE Congress on Evolutionary Computation, 2007, pp. Simul. 15 (2010) 3316–3331.
2229–2236.