Applied Soft Computing Journal 82 (2019) 105542

Contents lists available at ScienceDirect

Applied Soft Computing Journal

journal homepage: www.elsevier.com/locate/asoc

Simplified swarm optimization with initialization scheme for dynamic

weapon–target assignment problem
∗
Chyh-Ming Lai , Tsung-Hua Wu
Institute of Resources Management and Decision Science, Management College, National Defense University, Taipei 112, Taiwan

highlights

• The first work to apply SSO to the weapon–target assignment problem.

• Proposed a novel initialization scheme to generate a promising feasible solution.
• Proposed a local search to improve solutions under engagement feasibility constraints.
• Statistical results indicate that the proposed method is better than its competitors.

article info a b s t r a c t

Article history: The dynamic weapon–target assignment (DWTA) problem is a critical issue in the field of military
Received 27 February 2019 operations research. The problem is highly constrained; thus, the use of an evolutionary method to
Received in revised form 7 May 2019 solve the DWTA problem often encounters a population of infeasible solutions resulting in prohibitive
Accepted 29 May 2019
computational burden. Aiming at accelerating the solution process and improving the solution quality,
Available online 10 June 2019
this work proposes an improved simplified swarm optimization called SSODT with two novel schemes:
Keywords: the deterministic initialization scheme, and the target exchange scheme. The deterministic initializa-
Dynamic weapon–target assignment tion scheme is used in population initialization and utilizes problem-specific knowledge of DWTA to
problem speed up the convergence of SSODT by generating a promising feasible solution which has a greater
Simplified swarm optimization potential for evolving globally. The target exchange scheme is a local search updating feasible solutions
Feasible solution in a manner that exchanges their variables without violating the engagement feasibility to enhance
the exploitation capability of SSODT . The proposed method is empirically verified on thirty-six artificial
problems and compared with widely popular evolutionary methods. The results demonstrate that the
proposed SSODT is better than its competitors in terms of both solution quality and efficiency.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction to the time windows involved, which limit the engagement of

weapons. In DWTA, some weapons engage with targets at one
The weapon–target assignment (WTA) originally introduced stage, and the corresponding outcomes are evaluated to deter-
by Manne is an critical problem to be solved in the battlefield mine the engagement strategy for the next stage. Thus, DWTA is a
decision for firepower attack [1]. WTA problem aims to assign a multistage problem, and its goal is to find a global optimal assign-
limited number of weapons to enemy targets with the objective ment which involves the triplet pairing of weapons, targets and
of minimizing the total expected threat value of the targets re- stages for the whole defense process through all stages. The com-
lated to a target-based model, or maximizing the total expected plexity of DWTA is determined by capability constraints, strat-
survival value of the protected assets related to an asset-based egy constraints, resource constraints and engagement feasibility
model [2]. It is a classic NP-complete constrained combinatorial constraints (e.g. the limit of time windows) [5,6].
optimization problem in military operations research [1,3,4], and Existing WTA methods include exact algorithms for some spe-
can be mainly divided into two different versions: static WTA cial cases [7–9], goal programming [10,11], Lagrangian Relaxation
(SWTA) and dynamic WTA (DWTA) [2,5]. method [12], game theoretic framework [13] and heuristic opti-
SWTA assigns all weapons to targets in a single stage to mization algorithms [14–18]. These methods either can only ef-
find the optimal assignment for a temporary defense task. Com- fectively solve smaller problems or has weak global optimization
pared to SWTA, the dynamic version is more complicated due ability. Hence, the main focus has been on developing evolution-
ary computation methods, such as particle swarm optimization
∗ Corresponding author. (PSO) [19–23], genetic algorithm (GA) [24–28], simulated anneal-
E-mail address: [email protected] (C.-M. Lai). ing algorithm [24,29,30], ant colony optimization (ACO) [31–33]

https://doi.org/10.1016/j.asoc.2019.105542
1568-4946/© 2019 Elsevier B.V. All rights reserved.
2 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542

and other algorithms [34–36]. The results show that they have This work mainly focuses on the target-based DWTA problem.
made important contributions. However, there is still room to This problem can be formulated as:
improve the effectiveness and efficiency of the above works con- Ntgt
(Nstg N wpn )
sidering the strict demands for time-to-decision in the battlefield. ∑ ∏∏( )yhw,e
Minimize Ds (Y ) = s
ve h
1 − pw,e (1)
In this work, a novel algorithm based on simplified swarm opti-
e=1 h=s w=1
mization (SSO) is proposed as an alternative method for solving
DWTA. subject to
SSO is a population-based soft computing method which in- Ntgt
∑
tegrates both swarm intelligence and evolutionary computation yhw,e ≤ nw , ∀h ∈ {s, s + 1, s + 2, . . . , Nstg } ,
[37]. After its first application, SSO has been successfully ap- e=1
plied in many fields [38–42]. In addition, computation results ∀w ∈ {1, 2, . . . , N wpn} (2)
indicate that SSO and its variants exhibit better efficiency and N w pn
effectiveness than PSO, GA and ACO [43–46]. Considering the ∑
yhw,e ≤ ms,e , ∀h ∈ {s, s + 1, s + 2, . . . , Nstg } ,
efficient optimization ability of SSO, we adopt SSO to solve DWTA
w=1
problems in this work. The proposed algorithm, called SSODT , is
combined with a deterministic initialization scheme (DIS) and ∀e ∈ {1, 2, . . . , Ntgt } (3)
a target exchange scheme (TES). At the initialization phase of Nstg Ntgt
∑∑
the algorithm, DIS generates a promising feasible solution in a yhw,e ≤ aw , ∀w ∈ {1, 2, . . . , N w pn} (4)
deterministic way which utilizes problem-specific knowledge to h=s e=1
facilitate the proposed algorithm, obtaining better final solutions yhw,e h
fw,
≤ , ∀h ∈ {s, s + 1, s + 2, . . . , Nstg } ,
e
with more efficient convergence speed. TES is a local search
scheme which only serves feasible solutions in order to improve ∀w ∈ {1, 2, . . . , N wpn} , ∀e ∈ {1, 2, . . . , Ntgt } (5)
their quality during the evolution phase without undermining the yhw,e ∈ {0, 1} , ∀h ∈ {s, s + 1, s + 2, . . . , Nstg } ,
engagement feasibility.
The remainder of this paper is arranged as follows. Section 2
∀w ∈ {1, 2, . . . , N wpn} , ∀e ∈ {1, 2, . . . , Ntgt } (6)
formulates the mathematical model for DWTA and briefly intro- From the starting stage s to the final stage, the total expected
duces SSO which is the basis of the proposed method. Section 3 threat value of all targets is minimized as the objective of DWTA,
discusses the procedure of the proposed method for the problem. as shown in Eq. (1). Eq. (2) is a capability constraint of weapons.
The test problems are described, and the numerical results are It limits the number of targets that a weapon can engage in
discussed in Section 4, and Section 5 offers conclusions. each stage. The value of nw is usually set to be 1, because most
weapons can engage only one target at a time. A special weapon
2. Preliminaries which can simultaneously shoot multiple targets can be regarded
as multiple separate weapons. Eq. (3) ensures that at most mh,e
2.1. Problem description weapons can be assigned to target e at stage h. The value of mh,e
is based primarily on weapon performance and firing strategy.
The following notations are used to formulate the DWTA: Eq. (4) ensures that the amount of ammunitions, as used by a
certain weapon w through all stages, cannot exceed its predefined
Nwpns The total number of remaining weapons at allowable amount aw . Eq. (5) is the time window constraint,
starting stage s, and can be briefly represented which considers the influences of time windows on the engage-
as Nwpn. ment feasibility of weapons. It ensures that any shot fired by a
Ntgt s The total number of remaining targets at starting weapon in each stage must satisfy the engagement feasibility.
stage s, and can be briefly represented as Ntgt. Finally, all decision variables yhw,e are required to be binary.
Nstg The total number of stages considered in the
problem.
2.2. Background of SSO
T The set of targets at stage s, T = {1, 2, . . ., Ntgt}.
Y h, Y h, ] , . . ., Y ], with
Y h = [[Y h , Y h+1 Nstg

yhw,e h h
Y = yw,e N wpn×Ntgt , is the decision variables at The proposed SSODT is based on simplified swarm optimiza-
h
tion (SSO), and combined with two proposed novel schemes for
stage h, and yw,e = 1 if weapon w is assigned to DWTA. Before discussing SSODT , the required notations and the
target[ e at stage h; otherwise] yhw,e = 0. basic SSO are introduced below:
V, ve V = v1 , v2 , . . . , ve , . . . , vNtgt in which ve is the
threat[ value Nsol The total number of solutions in the population.
] of target e.
P h , phw,e P h = phw,e N wpn×Ntgt is the probability matrix of Nvar The total number of variables in a solution
Nnfe The total number required to calculate the
stage h, and phw,e is the probability that weapon
fitness function which is the stopping criterion
w destroys target e at stage h. ]
to halt( the algorithm.
N = n1 , n2 , . . . , nw , . . . , nN wpn in which nw is
[
N, nw
Xit = xti,1 , xti,2 , . . . , xti,j , . . . , xti,N v ar is the ith
)
Xit
the fire frequency of weapon w in one stage.
solution in the population at iteration t, where
M, mh,e M = [mh,e ]Nstg ×Ntgt in which mh , e is the weapon
i = 1,( 2, . . ., Nsol.
cost for target e at stage h.
Pi = pi,1 , pi,2 , . . . , pi,j , . . . , pi,N v ar is the pBest,
)
Pi
A, aw A = [a1 , a2 , . . ., aw , . . ., aN wpn ] in which aw is the
which is the best solution with the best fitness
amount [ hof] ammunition equipped for weapon w. value in the ith solution’s history.
F h , fw,
h
F h = fw, e N w pn×Ntgt is the engagement feasibility
e G G = (g1 , g2 , . . ., gj , . . ., gN v ar ) is the gBest, which
h
matrix, and fw, e = 1 if weapon w can shoot is the best solution with the best fitness value
h
target e at stage h; otherwise fw, e = 0. among all solutions.
F (•) The fitness value of •.
 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542 3

SSO is a novel population-based soft computing method, orig-

inally developed in 2009 by Yeh, for overcoming the deficien-
cies of particle swarm optimization (PSO) in solving discrete
problems. It integrates with swarm intelligence and evolutionary
computation, taking advantage of the merits of both [37]. Since
its first appearance, it has been successfully applied in a number
of optimization problems [43,44,47–49].
All candidate solutions in SSO are randomly generated within
the search space at the initialization phase. For each iteration t at
the evolution phase, each variable xti,j in Xit =
xi,1 , xti,2 , . . . , xti,j , . . . , xti,N v ar is updated successively by setting it
( t )

to the value of the same index in Xit −1 , gBest, pBest or a random

feasible value x depending on a uniform random number ρ in [0,
Fig. 1. An example of the encoding scheme.
1] as compared to three predetermined parameters Cg , Cp , and Cr ,
as follows [50]:
⎧
solutions to local rather than global optima. This claim can be
⎪
⎪ gj if ρ ∈ [0, Cg )
⎪
⎪ verified in the results of this study.
ρ ∈ [Cg , Cp )
⎪
pi,j if
⎨
An alternative way to improve on the convergence speed
xti,j = (7)
achieved using RRM [36] is to obtain a good starting solution
⎪
⎪ x if ρ ∈ [Cp , Cr )
to the problem [54,55]. Thus, this work proposes a novel ini-
⎪
⎪
⎪
ρ ∈ [Cr , 1]
⎩ t −1
xi,j if tialization procedure based on problem-specific knowledge, and
integrates it into SSO to ameliorate its solution quality and con-
Eq. (7) is the original SSO update mechanism (UMo ). According
vergence speed. To further improve solutions, a local search
to [44,51], the pBest scheme in UMo can be discarded for more
efficient updating solutions without compromising the solution called the target exchange scheme is also introduced in this
quality, as shown in Eq. (8), called UMf in this work. work.
⎧
⎨ gj
⎪
⎪ if ρ ∈ [0, Cg ) 3.1. Initialization
xti,j = x if ρ ∈ [Cg , Cr ) (8)
⎪
⎪
⎩
xti,−j 1 if ρ ∈ [Cr , 1] Since DWTA is constrained by time windows, the initial so-
lutions are very likely to violate the engagement feasibility. To
For a minimization problem, the primary steps of SSO are address this issue, available target lists (ATLs) for all weapons
summarized as follows: are constructed and maintained using the engagement feasibility
matrix F h . The ATL of weapon w in stage h can be defined as
SSO Procedure in Eq. (9), where 0 represents that weapon w does not engage
Initialization Phase: with any target in stage h. ATLs are conducive for initializing and
Step 0. Randomly generate Xi0 , calculate F (Xi0 ), and let updating solutions in the initialization and evolution phases, re-
pBest Pi = Xi0 for i = 1, 2, . . ., Nsol; find gBest G spectively, to randomly pair a triplet of (h, w, e) without violating
h
among all solutions such that G = arg min F (Xi0 ), the corresponding engagement feasibility, i.e., fw, e.
and let t = 1.
Lhw = 0, ε1 , ε2 , . . . εNtgt
{ }
Evolution Phase:
e, if fw, e = 1,
{ h (9)
Step 1. Let i = 1. εe =
Step 2. Update Xit based on Eq. (7).
∅, otherwise.
Step 3. If F (Xit ) < F (Pi ), let Pi = Xit and go to Step 4. In the initialization phase, each solution i in SSODT adopts
Otherwise, go to Step 5. decimal encoding and is generated with a vector of Nstg × Nwpn
Step 4. If F (Pi ) < F (G), let G = Pi .
variables, encoded as i X = (i X1 ; i X2 ; . . .; i Xh ; . . .; i XNstg ), where i Xh
Step 5. If i < Nsol, let i = i + 1 and go to Step 2.
= (i xh,1 , i xh,2 , . . ., i xh,w , . . ., i xh,N wpn ) corresponding to the decision
Step 6. If t < Nitr, let t = t + 1 and go to Step 1.
matrix Y h in stage h mentioned in Section 2.1, and the variable
Otherwise, halt and gBest is the final solution.
xh,w = e represents that weapon w is assigned to target e in
stage h as a triplet of (h, w, e). With ATL’s aid, xh,w is limited in
3. Proposed method Lhw . Thus, this representation guarantees that all solutions satisfy
the constraint Eqs. (2) and (5) after initialization. Let X = (1, 0,
Due to the complexity of DWTA, solving the problems often 3, 4, 6; 2, 6, 5, 0, 1) is a solution found by SSODT for a DWTA
produce infeasible solutions that violate one or more constraints, problem which Nwpn = 5, Ntgt = 6 and Nstg = 2. The decision
and may take significant time to converge. To address this, Juan matrixes of the problem, say Y 1 for stage 1 and Y 2 for stage 2,
Li et al. proposed a random repair mechanism (RRM) which can be derived from X as shown in Fig. 1. In Fig. 1, The encoded
modifies infeasible solutions to be feasible after the crossover variable x2,1 = 2 represents that weapon 1 is assigned to target
operator [20,36]. However, after evolution through several gen- 2 in stage 2 and thus the corresponding decision variable y21,2
erations, the infeasible solutions may contain useful information is equal to 1 and y21,e = 0, for e ̸ = 2. The encoded variable
to improve the search performance of the algorithm [52,53]. RRM x1,2 = 0 means that weapon 2 has no engagement in stage 1, thus
modifies them instead of keeping them intact, which may lead y12,e = 0, ∀e.
4 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542

Fig. 2. The flowchart of the proposed UMf .

3.2. Deterministic initialization scheme (DIS) Deterministic initialization scheme

Input: Data of a specific DWTA problem.
Output: A feasible solution X .
DIS aims to produce a good starting solution with problem-
Step 0: Let X = 0Nstg ×N wpn , T ∗ = T, υ = V, µ = M, for
specific knowledge to improve the solution quality and conver- h = 1, 2, . . ., Nstg, α = A, Nasn = min(Nvar,
gence rate for SSODT . In the process, the maximum ∑ number of ∑N wpn
N w pn w=1 aw ), and i = 1, where T, V, M and A have
assignments, i.e., Nasn, is determined by min(Nvar, w=1 aw ), defined in Section 2.1.
where Nvar = Nstg × Nwpn. For each assignment, the target τ , ∑Nstg
Step 1: Eliminate the target e with h=1 µh,e = 0 from
which has the highest expected threat value, i.e., υτ , in an valid target set T ∗ .
target set, i.e., T ∗ , has the first priority to be engaged by weapon Step 2: If T ∗ = ∅, go to Step 9.
ω at stage σ , which has the highest kill probability to target Step 3: Select the target τ with maximum threat value
τ , i.e., pσω,τ , among all weapons at all stages, and pσω,τ > 0 for υτ from T ∗ .
satisfying the constraint Eq. (5) (pσω,τ = 0 means that weapon Step 4: Select a pair of (stage, weapon), i.e., (σ , ω),
ω assigned to target τ at stage σ is forbidden because of the which can engage target τ (if xσ ,ω = 0, αω > 0,
and µσ ,τ > 0 are all satisfied) and has the
engagement feasibility). After each effective assignment, pσω,τ , υτ ,
highest kill probability pσω,τ among all available
µσ ,τ , αw and T ∗ are updated for the next assignment, where µσ ,τ weapons at all stages to target τ .
represents the remaining weapon cost for target τ at stage h Step 5: If (σ , ω) = ∅, eliminate the target τ from T ∗ ,
related to Eq. (3), and αw is the remaining amount of ammunition i = i + 1, and go to Step 8. Otherwise, go to
equipped for weapon w related to Eq. (4). Both µσ ,τ and αw are Step 6.
used for the assignment, thus avoiding violation of constraint Eqs. Step 6: Let xσ ,ω = τ , αw = αw − 1, µσ ,τ = µσ ,τ − 1 and
(3) and (4). υτ = υτ ×pσω,τ .
Since DIS is a deterministic process, only one initial solution Step 7: Let pσω,e = 0, ∀e, and i = i + 1.
Step 8: If i > Nasn, go to Step 9; otherwise go to Step 1.
which is guaranteed to be feasible is produced, and the remaining
Step 9: Halt.
Nsol-1 solutions are randomly initialized. Suppose the starting
stage s = 1 and the DIS procedure is given, as follows:
 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542 5

3.3. Fitness function Table 1

ATLs for the Example.
ATLs
In solving DWTA, there is a higher chance that infeasible
L25 = {0, 1, 2, 4}
solutions will be yielded. In order to address this, a penalty
function is a common way to drive the population toward the L11 = {0, 1, 2, 3, 4, 5}

feasible solution region [44]. In this work, the fitness function is L23 = {0, 2, 3, 6}
formulated as follows: L22 = {0, 3, 5, 6}

in Eqs. (3)–(5) are all satisfied, L24 = {0, 1, 3, 5, 6}

⎨D (Y ) if constraints
⎧
⎪
3 L15 = {0, 1, 3, 4, 5, 6}
F (X ) = ∑ (10)
⎩D (Y ) + λ ·
⎪ πi (Y ) , otherwise.
i=1
N w pn Nstg Nstg
∑∑ ∑ SSODT Procedure
π1 (Y ) = max[0, ( yhw,e ) − mh,e ] (11)
Initialization Phase:
w=1 h=s h=s
Step 0. Generate a feasible solution 1 X 0 using DIS,
Ntgt Nstg
∑ ∑ randomly initialize i X 0 for i = 2, 3, . . ., Nsol,
π2 (Y ) = max[0, ( yhw,e ) − aw ] (12) calculate F (i X 0 ) for i = 1, 2, . . ., Nsol, find
e=1 h=s gBest G among all solutions, and let t = 1.
N w pn Ntgt Nstg
[ ]
∑ ∑∑( Evolution Phase:
π3 (Y ) = max 0, h h
)⏐ h
yw,e − fw,e ⏐y
w,e =1 (13)
Step 1. Let i = 1.
w=1 e=1 h=s
Step 2. If the stopping criterion is met, halt and gBest
where D(Y ) is the objective value calculated according to Eq. (1). is the final solution. Otherwise, go to Step 3.
Y is a decision matrix for a DWTA problem and derived from X as Step 3. Update i X t using UMf .
shown in Fig. 1. λ is the degree of penalization set to 102 obtained Step 4. If F (i X t ) < F (G), let G = i X t .
by trial and error. πi (Y ) for i = 1, 2, 3, are the penalty functions Step 5. If ρ < Ntes and i X t is feasible, go to Step 6.
related to Eqs. (3)–(5), respectively. Otherwise, go to Step 8.
Step 6. Update i X t using TES.
3.4. Target exchange scheme (TES) Step 7. If F (i X t ) < F (G), let G = i X t .
Step 8. If i < Nsol, let i = i + 1 and go to Step 2.
Step 9. If t < Nitr, let t = t + 1 and go to Step 1.
During the evolution phase, all solutions are updated by UMf ,
as shown in Eq. (8) and Fig. 2. UMf is an effective update mech-
3.6. Complexity analysis
anism in exploration, and operates by updating all variables of
each solution. However, it focuses on exploring undiscovered
The computational cost of the original SSO algorithm involve
solution space even when the optimum is the neighbor of the cur- the initialization (Tini ), evaluation (Tev a ), and update (Tumo ) for
rent solutions, which means it lacks exploitation ability [45,56]. each solution. Assume D is the dimensionality of the search and
TES is therefore proposed in order to enhance the exploitation Nnfe is the maximum number of fitness function evaluations
capability of SSODT in DWTA. allowed for the algorithm. The time complexity of SSO can be
Since TES is not designed to rectify the infeasibility of solu- estimated as T (D) = Tini + (Tev a + Tumo ) × Nnfe = D + (D + D) ×
tions, feasible solutions are only presented after UMf so as not Nnfe = D(1 + 2·Nnfe). Therefore, O(D·Nnfe) is the time complexity
to waste computation resources, and is activated when ρ ≤ Ntes of the original SSO. Compared with the original SSO, SSODT has
in order to balance the global and local searches, where ρ is additional computation burden on the proposed two schemes:
a random number in [0, 1] and Ntes is predefined parameters DIS (Tdis ) and TES (Ttes ). In the worse cases, we have Tdis = D for
discussed in Section 4. For a feasible solution X in TES, only a sin- the first solution in the population (Tini = D for the others) and
gle variable is randomly selected and repeatedly exchanged with Ttes = D·Ntes for each solution, where Ntes ≤ 1. Therefore, the
the other variables in X by a random sequence. The exchange worst-case time complexity of SSODT can be calculated as T (D) =
is repeated until both the engagement feasibility is satisfied and Tdis (or Tini ) + (Tev a + Tumo + Ttes ) × Nnfe = D + (D + D + D·Ntes)
the two selected variables are different, or the sequence runs × Nnfe. Due to D·Ntes ≤ D, the computational complexity of the
out. That is, only two variables in a solution are exchanged SSODT can be expressed as O(D·Nnfe), which is the same to that
as an adjustment to the trade-off between exploration and ex- of the original SSO algorithm.
ploitation. The TES procedure is demonstrated in the following
example: 4. Experiment results and discussion

Example. Let (Nwpn, Ntgt, Nstg) = (5, 6, 2), X = (1, 5, 2, 0, 4; 5, 6, In order to evaluate the performance of the proposed SSODT
for DWTA, two experiments, EX-1 and EX-2, were carried out.
2, 6, 1) is a feasible solution, the corresponding ATLs constructed
The purpose of EX-1 was to observe the effect of two proposed
according to Eq. (9) are listed in Table 1, and the TES procedure
schemes, DIS and TES, with two parameters, Ndis and Ntes, re-
is given Table 2:
spectively, and then find the best setting for both. In EX-2, SSODT
with the best setting was compared with widely popular evo-
lutionary methods, including genetic algorithm (GA), discrete
3.5. The SSODT procedure particle swarm optimization (PSO) [57,58], population-based sim-
ulated annealing algorithm (PSA), tree seed algorithm (TSA) [59]
Based on the descriptions in Sections 2 and 3, the SSODT and the original SSO with UMo . In all experiments, Cg = 0.6,
flowchart is depicted in Fig. 3, and the overall procedure is Cp = 0.85 and Cr = 0.86 for SSO. For SSODT , Cg = 0.6 and
described below: Cr = 0.61. For GA, the crossover rate and mutation rate were
6 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542

Table 2
The TES procedure
Step 0. If ρ ≤ Ntes is satisfied, then call TES for the solution X . • activate TES
Step 1. Construct the random sequential list for X , i.e., η = {1, 1, 2, 6, 6, 4, • the first index variable in θ is x2,5 meaning that target (η1 = 1) of
5, 5, 0, 2}, and the corresponding index variable list θ = {x2,5 , x1,1 , weapon 5 in stage 2 is selected to be exchanged with the other
x2,3 , x2,2 , x2,4 , x1,5 , x1,2 , x2,1 , x1,4 , x1,3 }. variables in η
Step 2. Let t1 = η1 = 1, with its corresponding assigned index variable
θ1 = x2,5 , i = 2, and start to exchange.
Step 3. Let t2 = ηi=2 = 1, with its corresponding assigned index variable • weapon 1 in stage 1 with its target 1 is selected for exchange
θi=2 = x1,1 .
Step 4. Because t1 = t2 = 1, let i = i + 1 = 3. • targets of two weapons are the same, thus the exchange is not
valid
Step 5. Let t2 = ηi=3 = 2, with its corresponding assigned index variable • weapon 3 in stage 2 with its target 2 is selected for exchange
θi=3 = x2,3 .
Step 6. Because t1 = 1 ∈/ L23 = {0, 2, 3, 6}, let i = i + 1 = 4. • target 1 is forbidden for weapon 3 in stage 2, thus the exchange is
not valid
Step 7. Let t2 = ηi=4 = 6, with its corresponding assigned index variable • weapon 2 in stage 2 with its target 6 is selected for exchange
θi=4 = x2,2 .
Step 8. Because t1 = 1 ∈/ L22 = {0, 3, 5, 6}, let i = i + 1 = 5. • target 1 is forbidden for weapon 2 in stage 2, thus the exchange is
not valid
Step 9. Let t2 = ηi=5 = 6, with its corresponding assigned index variable • weapon 4 in stage 2 with its target 6 is selected for exchange
θi=5 = x2,4 .
Step 10. Because t2 = 6 ∈/ L25 = {0, 1, 2, 4}, let i = i + 1 = 6. • target 6 is forbidden for weapon 5 in stage 2, thus the exchange is
not valid
Step 11. Let t2 = ηi=6 = 4, with its corresponding assigned index variable • weapon 5 in stage 1 with its target 4 is selected for exchange
θi=6 = x1,5 .
Step 12. Because t1 = 1 ∈ L15 = {0, 1, 3, 4, 5, 6}, t2 = 4 ∈ L25 = {0, 1, 2, 4} • The exchange is satisfied
and t1 ̸ = t2, let x2,5 = t2 = 4 and x1,5 = t1 = 1.
Step 13. Original X = (1, 5, 2, 0, 4; 5, 6, 2, 6, 1) is updated as X = (1, 5, 2, • Update solution
0, 1; 5, 6, 2, 6, 4).

set as 0.7 and 0.3, respectively. For PSO, w = 0.2 and c1 = Table 3
c2 = 0.6, which were directly adopted from [57]. For PSA, the Value setting for the problem set.
Parameter Value Type
initial temperature T0 = 0.025, the temperature reduction rate
α = 0.99 and number of neighbors per individual Nmov = 5, Nwpn = (5, 10, 20, 50) Integer
Ntgt = (2 × Nwpn, 3 × Nwpn, 4 × Nwpn)
which were directly adopted from www.yarpiz.com. Nstg = (3, 4, 5)
TSA is originally designed for continuous optimization [60,61]. ve = 0.4 + ρ (0.95 − 0.4)
Float
For better solving DWTA problem, its updated rule is modified phw,e = 0.4 + ρ (0.9 − 0.4)
nw = 1
as Eq. (14). The modified rule includes two operators: crossover Integer
mh,e = [1, 2]
and mutation adopted from GA, where Si is the ith seed that aw = [1, 4]
will be produced by the ith tree Ti , B is the best tree location 1, if ρ ≤ pf ,
{
h
fw, = , where pf = 0.8 − 0.1 × h Float
obtained so far, Tr is the rth tree randomly selected from the e 0, otherwise.
population, ρ is a random number in [0, 1] and i ̸ = r. ST b and
ST r are two parameters of search tendency and set to 0.4 and
0.7, respectively; and the mutation rate is set to 0.01. 4.2. Experimental study of EX-1
crossov er (B, Ti ) if ρ < [0, STb )
{
Si = crossov er (Tr , Ti ) if ρ < [STb , STr ) (14) In order to better observe the effect of the two proposed
mutation (Ti ) if ρ < [STr , 1] schemes, a factorial experiment was conducted with 10 designed
treatments, and two factors were considered, Ndis at two levels:
In order to ensure a fair comparison of SSODT with these
1 and 0 corresponding to DIS being adopted or not in SSODT , and
methods, the number required to calculate the fitness function Ntes at five levels: 0, 0.1, 0.4, 0.7 and 1 corresponding to the
(Nnfe) was adopted as the stopping criterion, and set to Nnfe = trade-off coefficients of TES. Each of the 10 designed treatments
5 × Nstg × Nwpn × Ntgt. All compared methods were coded in was conducted with 10 replications by SSODT with Nsol = 30
MATLAB language and executed on an Intel Core i7 3.7-GHZ PC and tested against four problems selected from the problem set,
with 64 GB of memory. The runtime units were CPU seconds. including Nwpn = 5, 10, 20 and 50 under fixed Ntgt = 3×Nwpn
and Nstg = 5.
4.1. Problem set The main-effects of EX-1 are summarized in Table 4, in which
Fav g and Tav g represent the average of the fitness value and CPU
time over 10 replications, respectively, and the corresponding
In order to test the performance of SSODT , a set of problems standard deviations are represented by Fstd and Tstd . For better
was artificially generated based on three key factors which af- understanding and interpretation of the results, a plot of two
fected problem complexity, such as Nwpn, Ntgt and Nstg. The main effects is provided in Fig. 4. The figure shows that the
values of Nwpn were set to 5, 10, 20 and 50 against its own values adoption of DIS resulted in better Fav g with a slightly better CPU
of Ntgt = (2×Nwpn, 3×Nwpn, 4×Nwpn). For each pair of (Nwpn, time. The Ntes value is inclined to negative with Fav g . The worst
Ntgt), are available for Nstg = 3, 4 and 5. Thus, 36 combinations and best Fav g were attained at Ntes = 0 and 0.4, respectively,
of (Nstg, Nwpn, Ntgt) were constructed in the problem set. For and then the effect receded in tandem with the values of Ntes
each combination, the values of other parameters were generated increase. For Tav g , there was no clear pattern, meaning that Ntes
randomly according to Table 3 [62], where ρ is a random number had no apparent effect on CPU time. This can also be observed
in [0, 1]. by the ANOVA result, summarized in Table 5. The result indicates
 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542 7

Fig. 3. The SSODT flowchart.

that both Ndis and Ntes significantly affect the fitness value, but executed 30 runs with Nsol = 30 for each artificial problem. In
have no significant effect on CPU time. order to ensure a comprehensive comparison, GA, PSO, PSA and
The results of each designed treatment on the four selected TSA were applied with RRM [36], while SSO was retained in its
problems are listed in Table 6, where the best values of all original state. SSO without the assistance of RRM or DIS may
treatments are presented in bold type. The corresponding box converge to infeasible solutions. If its final solution was infeasible,
plot shown in Fig. 5 depicts the full range of results for Fav g , the corresponding fitness value was amended as 1 to narrow the
which summarize the results from EX-1 and provide much of the range of results and maintain comparability.
same information. As can be seen, (Ndis, Ntes) = (1, 0.4) is the The computational results of all problems, including the worst
most effective combination with a moderate computation cost (Fworst ), the average (Fav g ), the best (Fbest ) fitness value, the stan-
compared to the others. Thus, SSODT is implemented with DIS and dard deviation (Fstd ) of the obtained solutions and the average
TES using Ntes = 0.4 for the next experiment. CPU time (Tav g ), are reported in Table 7 for Nwpn = 5, Table 8
for Nwpn = 10, Table 9 for Nwpn = 20 and Table 10 for Nwpn
4.3. Experimental study of EX-2 = 50. The best values of each problem are shown in bold, and
the results are summarized using percentage difference, which is
The aim of EX-2 was to compare the performance of SSODT adopted to find the percent of difference of performance between
with its competitors: GA, PSO, PSA, TSA and SSO. Each method the control method and the others. The percentage difference
8 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542

Table 4
The main-effect of different treatment combinations for Ndis and Ntes.
Statistics Ndis Ntes Average
0 0.1 0.4 0.7 1
0 0.288853 0.286235 0.275954a 0.275974 0.277091 0.279757
Fav g 1 0.246638b 0.244899b 0.242544a b 0.242623b 0.243045b 0.243950
Average 0.267745 0.265567 0.259249 0.259298 0.260068
0 11.88534b 11.87191a 11.92826 11.95564 11.92930 11.91409
Tav g 1 11.95457 11.77746a b 11.83632b 11.93126b 11.91003b 11.88193
Average 11.91996 11.82469 11.88229 11.94345 11.91967
0 0.006641 0.008358 0.004979 0.004871a 0.007302 0.006405
Fstd 1 0.003551b 0.004303b 0.003635b 0.002179a b 0.004790b 0.003692
Average 0.005096 0.006331 0.004307 0.003525 0.006046
0 0.254858 0.225992 0.251831 0.214354 0.204831a b 0.230373
Tstd 1 0.234881b 0.194578b 0.176998a b 0.185658b 0.229449 0.204313
Average 0.244870 0.210285 0.214414 0.200006 0.217140
a
The best value among the values of the same row.
b
The best value among the values of the same column.

Table 5
The ANOVA of EX-1.
Group Source DF SS MS F value P value
Fav g Ndis 1 0.03205 0.03205 1122.94 0.000
Ntes 4 0.00161 0.00040 14.06 0.000
Ndis∗Ntes 4 0.00063 0.00016 5.55 0.000
Error 90 0.00257 0.00003
Total 99 0.03686
S = 0.005343 R2 = 93.03% R2adj = 92.33%

Tav g Ndis 1 0.02587 0.02587 0.54 0.464

Ntes 4 0.17278 0.04320 0.90 0.466
Ndis∗Ntes 4 0.08980 0.02245 0.47 0.758
Error 90 4.30939 0.04788
Total 99 4.59784
S = 0.218820 R2 = 6.27% R2adj = 0.00%

Table 6
The results of each designed treatment on four problems.
Problem Criteria Ntes
Ndis = 0 Ndis = 1
0 0.1 0.4 0.7 1 0 0.1 0.4 0.7 1
Fav g 0.2390 0.2308 0.2305 0.2328 0.2340 0.2120 0.2105 0.2093 0.2079 0.2097
5
Tav g 0.1247 0.1261 0.1295 0.1314 0.1327 0.1292 0.1383 0.1313 0.1319 0.1348
Fav g 0.2948 0.3046 0.2758 0.2769 0.2784 0.2485 0.2434 0.2382 0.2385 0.2384
10
Tav g 0.5532 0.5548 0.5629 0.5634 0.5662 0.5554 0.5558 0.5688 0.5713 0.5772
Fav g 0.3201 0.3149 0.3032 0.3039 0.3041 0.2730 0.2738 0.2720 0.2731 0.2726
20
Tav g 3.4914 3.5033 3.5219 3.5229 3.5508 3.4790 3.5152 3.5455 3.5382 3.5593
Fav g 0.3015 0.2946 0.2942 0.2904 0.2919 0.2531 0.2518 0.2507 0.2510 0.2515
50
Tav g 43.372 43.304 43.499 43.605 43.468 43.655 42.901 43.100 43.484 43.369
Over all Fav g 0.2889 0.2862 0.2760 0.2760 0.2771 0.2466 0.2449 0.2425 0.2426 0.2430
Tav g 11.885 11.872 11.928 11.956 11.929 11.955 11.778 11.836 11.931 11.910

Fig. 4. The main-effect plots of EX-1.

 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542 9

Fig. 5. The box plot of Fav g on different selected problems.

Fig. 6. The plot of percentage difference of methods.

of Tav g and Fav g for a specific method M, denoted by pdT M and larger problems which have a higher chance to encounter infea-
pdF M , respectively, can be calculated by Eq. (15), and SSO is the sible solutions. SSO without any auxiliary maintains its efficiency
control method. The results are listed in Table 11, and presented consistently. Compared with SSO, SSODT is more efficient on the
graphically in Fig. 6. problems for which Nwpn = 5, 10 and 20 and pdT is more than
1.56161%, and slightly less on the problem for which Nwpn = 50
Fav g ,SSO − Fav g ,M Tav g ,SSO − Tav g ,M
pdFM = ( ) , pdTM = ( ) (15) and pdT is −0.43839%. This result is the same as that obtained in
Fav g ,SSO +Fav g ,M Tav g ,SSO +Tav g ,M
EX-1, in that the adoption of the two proposed schemes does not
2 2
impact the efficiency of SSODT .
Tables 7–10 show that the quality of solutions yielded by GA is
4.4. Results and discussion better than those obtained by PSO, PSA and TSA on most problems
except the problem (5, 10, 3), and better than that yielded by
From Table 11, we can observe that among the six algorithms, SSODT on problems (5, 10, 4) and (5, 20, 4). However, the quality
the algorithms which adopt RRM, i.e., GA, PSO PSA and TSA, have deteriorates significantly as the size of the problem increases.
the worst efficiency with pdT, less than −26.14258%, as the repair Also, the solution quality of PSO PSA and TSA, which adopt the
for an infeasible solution is very time consuming, especially for same scheme, i.e., RRM, as GA, has the same pattern. This pattern
10 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542

Fig. 7. The convergence comparison of different methods.

Fig. 8. The best solution on the problem (20, 80, 3).

can be observed clearly in Fig. 6. SSO performs poorly on smaller be that without the aid of RRM or DIS, it needs more time to
problems, i.e., Nwpn = 5 and 10, and the reason for this might converge and has a higher chance to trap in local optimum. This
 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542 11

Fig. 9. The best solution on the problem (20, 80, 4).

Table 7 Table 8
Computational results of all compared methods when Nwpn = 5. Computational results of all compared methods when Nwpn = 10.
(Ntgt, Nstg) Criteria GA PSO PSA TSA SSO SSODT (Nwpn, Ntgt) Criteria GA PSO PSA TSA SSO SSODT
(10, 3) Fworst 0.19902 0.22657 0.21652 0.18358 0.22027 0.20656 (20, 3) Fworst 0.24970 0.26627 0.29886 0.25662 0.30152 0.22957
Fav g 0.16456 0.17207 0.17177 0.15965 0.18197 0.16027 Fav g 0.22039 0.23976 0.24552 0.22880 0.25250 0.19690
Fbest 0.13556 0.12804 0.13687 0.12466 0.14610 0.12956 Fbest 0.19093 0.21052 0.19807 0.20484 0.20410 0.17813
Fstd 0.01312 0.02382 0.01875 0.01578 0.02063 0.02091 Fstd 0.01332 0.01576 0.02483 0.01490 0.02296 0.01152
Tav g 0.06692 0.06756 0.06575 0.06628 0.04811 0.04795 Tav g 0.26184 0.28062 0.26142 0.26572 0.19539 0.18880
(10, 4) Fworst 0.40512 0.44603 0.43882 0.45459 1.00000 0.41794 (20, 4) Fworst 0.21055 0.24141 0.23245 0.24391 1.00000 0.18820
Fav g 0.37558 0.39101 0.38899 0.38816 0.45025 0.38453 Fav g 0.17403 0.19273 0.20171 0.19021 0.27985 0.15738
Fbest 0.34729 0.34738 0.35016 0.35606 0.35366 0.35108 Fbest 0.14389 0.15193 0.17299 0.16553 0.17632 0.13782
Fstd 0.01720 0.02122 0.01853 0.02449 0.11489 0.01715 Fstd 0.01710 0.02008 0.01545 0.01857 0.24456 0.00974
Tav g 0.09660 0.09899 0.09368 0.08979 0.06888 0.06713 Tav g 0.39701 0.41724 0.40362 0.39723 0.29012 0.28182
(10, 5) Fworst 0.32672 0.34443 0.35879 0.35582 1.00000 0.32985 (20, 5) Fworst 0.13439 0.14553 0.15765 0.15362 0.15935 0.12524
Fav g 0.28989 0.30560 0.30373 0.30442 0.40940 0.28547 Fav g 0.11851 0.12539 0.13221 0.12676 0.13353 0.11473
Fbest 0.25859 0.27254 0.26077 0.25245 0.27117 0.25245 Fbest 0.10408 0.10210 0.10870 0.10485 0.11106 0.09951
Fstd 0.01845 0.01621 0.02283 0.02952 0.20224 0.02492 Fstd 0.00682 0.00866 0.01193 0.01070 0.01219 0.00549
Tav g 0.12732 0.13643 0.12931 0.12396 0.09193 0.09045 Tav g 0.55327 0.57515 0.54801 0.53235 0.39843 0.38796
(15, 3) Fworst 0.43240 0.44251 0.43878 0.43173 1.00000 0.43313 (30, 3) Fworst 0.36920 0.38471 0.38228 0.39124 1.00000 0.31875
Fav g 0.40584 0.40814 0.41467 0.40938 0.46691 0.39918 Fav g 0.33815 0.35274 0.35302 0.35046 0.38157 0.30601
Fbest 0.38319 0.37924 0.37921 0.37763 0.40248 0.37681 Fbest 0.30408 0.32339 0.32211 0.31380 0.32685 0.29428
Fstd 0.01294 0.01439 0.01655 0.01460 0.14558 0.01290 Fstd 0.01759 0.01637 0.01578 0.01758 0.11812 0.00587
Tav g 0.09248 0.09957 0.09566 0.09074 0.06990 0.06795 Tav g 0.40342 0.42983 0.40716 0.40627 0.30504 0.29463
(15, 4) Fworst 0.27370 0.29938 0.30748 0.28927 0.33351 0.25477 (30, 4) Fworst 0.28076 0.29526 0.30762 0.30811 0.30525 0.24398
Fav g 0.24542 0.26423 0.26601 0.25288 0.26801 0.22906 Fav g 0.25268 0.27165 0.27554 0.26583 0.27079 0.21475
Fbest 0.21040 0.21940 0.23972 0.21031 0.22874 0.21446 Fbest 0.21623 0.23406 0.24308 0.22941 0.24357 0.19955
Fstd 0.01602 0.01753 0.01900 0.02013 0.02175 0.01020 Fstd 0.01541 0.01447 0.01785 0.01796 0.01572 0.01153
Tav g 0.13872 0.14995 0.14263 0.13542 0.10489 0.10305 Tav g 0.60683 0.64683 0.61090 0.59401 0.45011 0.44542
(15, 5) Fworst 0.24381 0.25930 0.25670 0.26642 0.29429 0.22023 (30, 5) Fworst 0.32393 0.32031 0.33349 0.32834 0.32683 0.25891
Fav g 0.21814 0.23022 0.23335 0.22624 0.24546 0.20418 Fav g 0.28476 0.28855 0.30147 0.29496 0.29956 0.24033
Fbest 0.19568 0.20740 0.18978 0.19686 0.20282 0.19048 Fbest 0.24258 0.25597 0.26980 0.24269 0.26403 0.22804
Fstd 0.01127 0.01327 0.01524 0.01741 0.02406 0.00811 Fstd 0.01746 0.01910 0.01608 0.02211 0.01480 0.00762
Tav g 0.19532 0.20491 0.19451 0.18712 0.13999 0.13795 Tav g 0.86334 0.89899 0.85269 0.82776 0.61769 0.60060
(20, 3) Fworst 0.46523 0.48605 0.50684 0.47570 0.52524 0.45217 (40, 3) Fworst 0.46759 0.49489 0.48339 0.48455 1.00000 0.44693
Fav g 0.44202 0.45613 0.46894 0.44986 0.47136 0.42728 Fav g 0.45368 0.46518 0.46727 0.46139 0.48642 0.43714
Fbest 0.41547 0.42296 0.44109 0.41501 0.42808 0.41280 Fbest 0.43574 0.44271 0.45032 0.43593 0.43785 0.42797
Fstd 0.01190 0.01604 0.01602 0.01243 0.02186 0.00929 Fstd 0.00946 0.01287 0.00816 0.01298 0.09778 0.00474
Tav g 0.12531 0.13905 0.12592 0.12498 0.09479 0.09096 Tav g 0.54601 0.59194 0.55839 0.53755 0.41835 0.40923
(20, 4) Fworst 0.58636 0.60969 0.60675 0.61192 1.00000 0.59838 (40, 4) Fworst 0.48462 0.48456 0.48406 0.47719 1.00000 0.46106
Fav g 0.57007 0.57331 0.58344 0.58006 0.68393 0.57713 Fav g 0.45579 0.46081 0.46598 0.46542 0.48887 0.43934
Fbest 0.54785 0.54832 0.55828 0.55794 0.57682 0.55878 Fbest 0.43535 0.44221 0.44327 0.44917 0.45203 0.42665
Fstd 0.01019 0.01442 0.01394 0.01305 0.16119 0.00999 Fstd 0.01030 0.00892 0.00972 0.00728 0.09705 0.00700
Tav g 0.18807 0.20189 0.18944 0.18270 0.13752 0.13524 Tav g 0.83494 0.88836 0.83494 0.81183 0.62199 0.61777
(20, 5) Fworst 0.39305 0.40397 0.41462 0.41199 0.45214 0.37645 (40, 5) Fworst 0.42299 0.45991 0.45682 0.43976 1.00000 0.40963
Fav g 0.36338 0.37550 0.37936 0.37375 0.39323 0.35755 Fav g 0.41108 0.42361 0.42781 0.42366 0.45556 0.39679
Fbest 0.33790 0.35188 0.34737 0.34870 0.35205 0.34192 Fbest 0.39455 0.40057 0.40661 0.40237 0.41756 0.38688
Fstd 0.01310 0.01500 0.01750 0.01473 0.02326 0.00835 Fstd 0.00779 0.01323 0.01227 0.00907 0.10346 0.00526
Tav g 0.26057 0.27415 0.26131 0.25271 0.18739 0.18728 Tav g 1.16427 1.22341 1.17004 1.13579 0.85420 0.83439
Average Fav g 0.34166 0.35291 0.35670 0.34938 0.39673 0.33607 Average Fav g 0.30101 0.31338 0.31895 0.31194 0.33874 0.27815
Tav g 0.14348 0.15250 0.14425 0.13930 0.10482 0.10311 Tav g 0.62566 0.66138 0.62746 0.61206 0.46126 0.45118
12 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542

Table 9 4.5. Nonparametric statistical test

Computational results of all compared methods when Nwpn = 20.
(Ntgt, Nstg) Criteria GA PSO PSA TSA SSO SSODT
In this section, a statistical analysis suggested in [63] is em-
Fworst 0.19042 0.20532 0.23253 0.20330 0.20780 0.16045
(40, 3) ployed to further confirm whether SSODT provides a statistically
Fav g 0.16850 0.18315 0.19409 0.18288 0.18154 0.14225
Fbest 0.14405 0.16096 0.17774 0.15279 0.16273 0.13025 significant improvement over its competitors for DWTA prob-
Fstd 0.01048 0.01239 0.01294 0.01248 0.01208 0.00618 lems. The analysis first applies Friedman’s test to rank the perfor-
Tav g 1.34269 1.42786 1.35349 1.37987 1.02644 0.99501 mance of all compared methods and detect significant differences
Fworst 0.17408 0.18415 0.20275 0.18779 0.19833 0.14712
(40, 4) over multiple comparisons. If the null hypothesis, which states
Fav g 0.15205 0.16969 0.17825 0.16455 0.16756 0.13484
Fbest 0.13596 0.14853 0.15477 0.14243 0.14926 0.12370 that there is no significant difference between the performance
Fstd 0.00874 0.01064 0.01055 0.01137 0.01150 0.00500 of any two methods, is rejected, then Holm’s method is adopted
Tav g 2.04956 2.16789 2.09599 2.13737 1.58808 1.53610 as a post hoc test to perform pairwise comparisons considering
Fworst 0.12433 0.13296 0.13686 0.12922 0.14661 0.10859
(40, 5)
Fav g 0.11162 0.11806 0.12403 0.11643 0.12155 0.10118
SSODT and the others. For all comparisons, the significance level
Fbest 0.09473 0.10615 0.11008 0.10527 0.10439 0.09257 is set to α = 0.05 to determine whether the hypothesis is
Fstd 0.00702 0.00750 0.00609 0.00615 0.00825 0.00409 rejected. Since Fav g represents the quality of solutions obtained
Tav g 2.89348 3.02343 2.92028 3.02378 2.20613 2.16843 on all tested problems, it is adopted as the target value for the
Fworst 0.40711 0.41833 0.43036 0.43193 0.43074 0.36923
(60, 3)
Fav g 0.38952 0.39643 0.40223 0.39674 0.39800 0.35873
statistical analysis.
Fbest 0.37335 0.37927 0.38391 0.37343 0.38120 0.34722 Table 12 shows the average rankings computed using the
Fstd 0.00764 0.01204 0.01018 0.01165 0.01213 0.00573 Friedman test. The results indicate that SSODT is the best per-
Tav g 2.15245 2.30421 2.18907 2.22226 1.68376 1.66023 forming method of the comparisons, followed by GA, PSO, TSA,
Fworst 0.32836 0.34840 0.34009 0.33428 0.33821 0.28590
(60, 4)
Fav g 0.31107 0.31905 0.32336 0.32199 0.31755 0.27628 SSO and PSA, consecutively. The p-value of the Friedman test
Fbest 0.29012 0.30032 0.29490 0.30590 0.30517 0.26798 strongly suggests the existence of significant differences among
Fstd 0.00932 0.01006 0.00986 0.00722 0.00787 0.00510 the considered methods, and then the statistical analysis contin-
Tav g 3.37428 3.56919 3.42828 3.53032 2.64848 2.57646
ues with Holm’s test. Table 12 also shows the results of the post
Fworst 0.34601 0.34048 0.34853 0.34647 0.34193 0.28160
(60, 5) hoc test, in which the adjusted p-vale (APV) is more rigorous than
Fav g 0.31495 0.31682 0.32094 0.32301 0.31589 0.27141
Fbest 0.28954 0.30232 0.29902 0.30248 0.29341 0.25928 the p-value, as it considers the accumulated family errors [64].
Fstd 0.01110 0.01102 0.01152 0.01170 0.01185 0.00509 According to the p-values and APVs of Holm’s test, it is safe to
Tav g 4.88054 5.07970 4.91392 5.45526 3.72404 3.68865
conclude that the proposed method, SSODT , is significantly more
Fworst 0.43290 0.44754 0.45290 0.43834 0.44408 0.40318
(80, 3) effective than the other considered methods.
Fav g 0.41795 0.43200 0.42978 0.42719 0.42923 0.39742
Fbest 0.40107 0.41705 0.41114 0.41576 0.41404 0.39121
Fstd 0.00697 0.00738 0.00949 0.00567 0.00633 0.00305
Tav g 3.06053 3.19248 3.08735 3.17214 2.33719 2.34440
5. Conclusion
Fworst 0.40960 0.42971 0.42167 0.41743 0.41441 0.36866
(80, 4)
Fav g 0.39312 0.39876 0.40145 0.40264 0.40082 0.36232 DWTA is a highly complex constrained combinatorial opti-
Fbest 0.37221 0.38595 0.38894 0.38503 0.38648 0.35628
Fstd 0.00908 0.00961 0.00816 0.00930 0.00678 0.00325
mization problem due to its capability constraints, strategy con-
Tav g 4.90895 5.09685 4.85676 5.07423 3.73752 3.68354 straints, resource constraints, and engagement feasibility con-
(80, 5)
Fworst 0.41849 0.42797 0.43914 0.42951 0.43963 0.40482 straints. The method without any assistance, e.g. RRM, may of-
Fav g 0.40887 0.41681 0.41963 0.41619 0.42369 0.39490 ten encounter infeasible solutions that violate one or more con-
Fbest 0.39874 0.40649 0.40699 0.40627 0.41221 0.38895
Fstd 0.00548 0.00587 0.00811 0.00598 0.00704 0.00349 straints and lead to long computational times. In order to address
Tav g 7.06655 7.43644 7.16125 7.17303 5.44152 5.36232 this problem, this study proposes an SSO-based method called
Fav g 0.29641 0.30564 0.31042 0.30574 0.30620 0.27104 SSODT to solve DWTA problems.
Average
Tav g 3.52545 3.69978 3.55627 3.68536 2.71035 2.66835 SSODT integrates two novel schemes, DIS and TES, into the
existing SSO to efficiently optimize DWTA. DIS is a deterministic
process in which the target with the highest expected threat
value always takes priority for engagement by the weapon with
situation can be seen in Fig. 7 which depicts the progress of the the highest kill probability for that target. This method generates
average gBest over 30 runs with respect to the number of fitness a promising feasible initial solution for SSODT at the initialization
function evaluated (nfe) on four selected problems. Fig. 7 provides phase. During the evolution phase, TES, which is designed as a
insights into the convergence behavior of the compared methods. local search, exchanges targets between two triplets selected ran-
As can be seen, SSO without RRM or DIS exhibits the slowest domly from a feasible solution under the engagement feasibility
convergence among the compared algorithm. Compared with the constraints, thus improving the exploitation capability of SSODT
peer algorithms, SSODT has the best starting points generated in DWTA.
by DIS to speed up its convergence rate, and meanwhile, it is An extensive experimental study on 36 DWTA problems was
not easy to be trapped in local optima. The proposed method conducted. The results indicate that the two proposed schemes
obtains the highest scores for Fworst , Fav g , Fbest and Fstd on most enhance the performance of SSODT to achieve better quality solu-
problems with overall pdf = 15.24366%. This empirically demon- tions in a more efficient way than its competitors. Furthermore,
strates that the two proposed schemes, DIS and TES, enhance the statistical analysis significantly confirms these findings.
performance of SSODT to more effectively and efficiently deal with
DWTA problems. Acknowledgments
The best solutions obtained by SSODT on the problems (20,
80, 3), (20, 80, 4) and (20, 80, 5) are presented graphically in We are thankful to lieutenants Jo-Yun Ma, Pei-Ling Wang,
Figs. 8–10. As can be seen, solutions have more engagements Wei-Chien Chang and Yi-Han Fang for reorganizing literature,
at the former stages, since the fact that fewer weapons can be and this research was supported by the Ministry of Science and
used at latter stages. This situation results from the engagement Technology of Taiwan, R. O. C, under grant MOST 107-2221-E-
feasibility and depletion of ammunition. 606-009.
 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542 13

Fig. 10. The best solution on the problem (20, 80, 5).

Table 10
Computational results of all compared methods when Nwpn = 50.
(Nwpn, Ntgt) Criteria GA PSO PSA TSA SSO SSODT
(100, 3) Fworst 0.17668 0.18408 0.21278 0.19687 0.16557 0.13517
Fav g 0.16113 0.17161 0.19168 0.18481 0.15093 0.12911
Fbest 0.14699 0.15892 0.17967 0.16635 0.14172 0.12182
Fstd 0.00634 0.00603 0.00881 0.00740 0.00597 0.00311
Tav g 15.40504 16.09844 15.28134 15.07158 11.07099 11.10133
(100, 4) Fworst 0.13970 0.14508 0.17105 0.15623 0.13767 0.12149
Fav g 0.12999 0.13689 0.15213 0.14298 0.12895 0.11548
Fbest 0.12110 0.12731 0.13570 0.13445 0.12030 0.10986
Fstd 0.00472 0.00498 0.00710 0.00566 0.00417 0.00325
Tav g 25.13350 26.54844 25.65071 24.71101 18.32753 18.17943
(100, 5) Fworst 0.13414 0.11812 0.15081 0.14564 0.10880 0.09438
Fav g 0.11682 0.11044 0.14009 0.13646 0.10279 0.08857
Fbest 0.10601 0.09917 0.12551 0.12503 0.09630 0.08093
Fstd 0.00660 0.00454 0.00594 0.00549 0.00302 0.00249
Tav g 37.36647 37.75061 35.63524 37.15663 24.43429 23.87315
(150, 3) Fworst 0.31760 0.31544 0.34218 0.33556 0.29820 0.24612
Fav g 0.30561 0.30161 0.32764 0.31801 0.28543 0.23908
Fbest 0.29403 0.28762 0.31100 0.30427 0.27545 0.23149
Fstd 0.00721 0.00631 0.00829 0.00754 0.00621 0.00293
Tav g 24.88161 26.45959 25.85630 27.53860 18.63722 18.34009
(150, 4) Fworst 0.28370 0.29427 0.29994 0.29487 0.27414 0.21618
Fav g 0.26766 0.26960 0.28344 0.28109 0.25619 0.20991
Fbest 0.25033 0.24956 0.26743 0.26604 0.24483 0.20111
Fstd 0.00805 0.00909 0.00905 0.00784 0.00749 0.00394
Tav g 42.14268 43.66848 42.04012 45.33457 29.84173 29.68406
(150, 5) Fworst 0.32213 0.32661 0.34741 0.34802 0.31055 0.25887
Fav g 0.30687 0.30867 0.33476 0.32931 0.29078 0.25077
Fbest 0.29652 0.29394 0.31583 0.31387 0.27471 0.24285
Fstd 0.00696 0.00798 0.00812 0.00867 0.00766 0.00393
Tav g 60.90008 63.00610 61.99925 67.11380 44.33655 43.91403
(200, 3) Fworst 0.47006 0.47045 0.48799 0.48122 0.46117 0.42731
Fav g 0.45926 0.46196 0.47602 0.47191 0.45379 0.42192
Fbest 0.44991 0.45256 0.46549 0.46334 0.44687 0.41834
Fstd 0.00475 0.00418 0.00513 0.00491 0.00389 0.00197
Tav g 40.41742 42.19455 41.90846 45.31574 31.26270 32.01318
(200, 4) Fworst 0.43186 0.43609 0.44539 0.44190 0.43160 0.39206
Fav g 0.42164 0.42694 0.43486 0.42892 0.41965 0.38663
Fbest 0.41185 0.41485 0.42016 0.42155 0.41031 0.38208
Fstd 0.00446 0.00519 0.00579 0.00524 0.00526 0.00283
Tav g 73.37474 78.01042 72.35262 75.91309 54.32905 54.83622
(200, 5) Fworst 0.43384 0.42872 0.45482 0.46264 0.42238 0.39131
Fav g 0.42167 0.41982 0.44607 0.44483 0.41390 0.38625
Fbest 0.40527 0.41348 0.42717 0.43204 0.40618 0.37987
Fstd 0.00572 0.00358 0.00607 0.00662 0.00408 0.00310
Tav g 106.62566 110.90182 110.82188 113.30216 79.52822 81.19654
Average Fav g 0.28785 0.28973 0.30963 0.30426 0.27804 0.24752
Tav g 47.36080 49.40427 47.94955 50.16191 34.64092 34.79312
14 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542

Table 11
The percentage difference of methods compared with SSO.
Nwpn Criteria GA PSO PSA TSA SSODT
5 Fav g 14.91624 11.68961 10.62618 12.69196 16.55349
Tav g −31.13733 −37.05718 −31.65724 −28.24757 1.64835
10 Fav g 11.79571 7.77768 6.01835 8.23616 19.64286
Tav g −30.25105 −35.65163 −30.53258 −28.09956 2.20905
20 Fav g 3.25186 0.18343 −1.36707 0.15289 12.18388
Tav g −26.14258 −30.87100 −26.99762 −30.48963 1.56161
50 Fav g −3.46583 −4.11521 −10.74949 −9.00313 11.61437
Tav g −31.02345 −35.13194 −32.22800 −36.60489 −0.43839
Fav g 7.28745 4.49786 1.83681 3.73584 15.24366
Over all
Tav g −30.67460 −34.84627 −31.84256 −36.06141 −0.25920

Table 12 [17] B. Xin, J. Chen, J. Zhang, L. Dou, Z. Peng, Efficient decision makings
The results of the statistical test. for dynamic weapon-target assignment by virtual permutation and tabu
Method Friedman’s test Holm’s test search heuristics, IEEE Trans. Syst. Man Cybern. C 40 (2010) 649–662.
[18] A.G. Kline, D.K. Ahner, B.J. Lunday, Real-time heuristic algorithms for the
Rank Statistic p-value Statistic p-value APV
static weapon target assignment problem, J. Heuristics (2017) 1–21.
GA 2.30556 2.70875 0.00675 0.00675 [19] Y. Zhou, X. Li, Y. Zhu, W. Wang, A discrete particle swarm optimization
PSO 3.88889 6.29941 0.00000 0.00000 algorithm applied in constrained static weapon-target assignment problem,
PSA 5.33333 120.46032 0.00000 9.57510 0.00000 0.00000 in: Proceedings of the 12th World Congress on Intelligent Control and
TSA 3.91667 6.36240 0.00000 0.00000 Automation (WCICA), IEEE, 2016, pp. 3118–3123.
SSO 4.44444 7.55929 0.00000 0.00000 [20] G. Peng, Y. Fang, S. Chen, W. Peng, D. Yang, A hybrid multi-objective
SSODT 1.11111 discrete particle swarm optimization algorithm for cooperative air com-
bat DWTA, in: Proceedings of Internatonal Conference on Bio-Inspired
Computing-Theories and Applications, Springer, Singapore, 2016, pp.
114–119.
Declaration of competing interest [21] X. Zeng, Y. Zhu, L. Nan, K. Hu, B. Niu, X. He, Solving weapon-target assign-
ment problem using discrete particle swarm optimization, in: Proceedings
of the 6th World Congress on Intelligent Control and Automation (WCICA),
No author associated with this paper has disclosed any po- IEEE, 2006, pp. 3562–3565.
tential or pertinent conflicts which may be perceived to have [22] C. Leboucher, H.S. Shin, P. Siarry, R. Chelouah, S. Le Ménec, A. Tsourdos,
impending conflict with this work. For full disclosure statements A two-step optimisation method for dynamic weapon target assignment
refer to https://doi.org/10.1016/j.asoc.2019.105542. problem, in: Recent Advances on Meta-Heuristics and their Application To
Real Scenarios, InTech, 2013.
[23] L. Yang, Z. Zhai, Y. Li, Y. Huang, A multi-information particle swarm
References optimization algorithm for weapon target assignment of multiple kill ve-
hicle, in: Proceedings of International Conference on Advanced Intelligent
[1] A.S. Manne, A target-assignment problem, Oper. Res. 6 (1958) 346–351. Mechatronics (AIM), IEEE, Auckland, New Zealand, 2018, pp. 1160–1165.
[2] P.A. Hosein, J.T. Walton, M. Athans, Dynamic Weapon-Target Assignment [24] S. Bisht, Hybrid genetic-simulated annealing algorithm for optimal weapon
Problems with Vulnerable C2Ì3 Nodes, 1988. allocation in multilayer defence scenario, Def. Sci. J. 54 (2004) 395.
[3] S.P. Lloyd, H.S. Witsenhausen, Weapons allocation is NP-complete, in: [25] H. Lu, H. Zhang, X. Zhang, R. Han, An improved genetic algorithm for target
Proceedings of IEEE Summer Computer Simulation Conference, Reno, NV, assignment, optimization of naval fleet air defense, in: Proceedings of the
1986, pp. 1054–1058. 6th World Congress on Intelligent Control and Automation (WCICA), IEEE,
Dalian, China, 2006, pp. 3401–3405.
[4] R.H. Day, Allocating weapons to target complexes by means of nonlinear
programming, Oper. Res. 14 (1966) 992–1013. [26] Z.J. Lee, S.F. Su, C.Y. Lee, Efficiently solving general weapon-target assign-
ment problem by genetic algorithms with greedy eugenics, IEEE Trans.
[5] A. Kline, D. Ahner, R. Hill, The weapon-target assignment problem, Comput.
Syst. Man Cybern. B 33 (2003) 113–121.
Oper. Res. (2018) 226–236.
[27] T.p. Fu, Y.s. Liu, J.h. Chen, Improved genetic and ant colony optimization
[6] C. Huaiping, L. Jingxu, C. Yingwu, W. Hao, Survey of the research on
algorithm for regional air defense wta problem, in: Proceedings of the
dynamic weapon-target assignment problem, J. Syst. Eng. Electron. 17
First International Conference on Innovative Computing, Information and
(2006) 559–565.
Control (ICICIC), IEEE, Beijing, China, 2006, pp. 226–229.
[7] G. denBroeder Jr, R. Ellison, L. Emerling, On optimum target assignments,
[28] J. Chen, B. Xin, Z. Peng, L. Dou, J. Zhang, Evolutionary decision-makings for
Oper. Res. 7 (1959) 322–326.
the dynamic weapon-target assignment problem, Sci. China Ser. F: Inf. Sci.
[8] S.C. Chang, R.M. James, J.J. Shaw, Assignment algorithm for kinetic en- 52 (2009) 2006.
ergy weapons in boost phase defence, in: Proceedings of the 26th IEEE
[29] E. Sonuc, B. Sen, S. Bayir, A parallel simulated annealing algorithm for
Conference on Decision and Control, IEEE, Los Angeles, CA, 1987, pp.
weapon-target assignment problem, Int. J. Adv. Comput. Sci. Appl. 8 (2017)
1678–1683.
87–92.
[9] D. Orlin, Optimal weapons allocation against layered defenses, Nav. Res. [30] L. Yan, D. Yu’na, Weapon-target assignment based on simulated annealing
Logist. 34 (1987) 605–617. and discrete particle swarm optimization in cooperative air combat, Acta
[10] D.J. Green, J.T. Moore, J.J. Borsi, An integer solution heuristic for the arsenal Aeronaut. Astronaut. Sin. 3 (2010).
exchange model (AEM), Mil. Oper. Res. (1997) 5–15. [31] Z.J. Lee, C.Y. Lee, S.F. Su, An immunity-based ant colony optimization algo-
[11] M.F. Hocaoğlu, Weapon target assignment optimization for land based rithm for solving weapon-target assignment problem, Appl. Soft Comput.
multi-air defense systems: A goal programming approach, Comput. Ind. 2 (2002) 39–47.
Eng. 128 (2019) 681–689. [32] W. Yanxia, Q. Longjun, G. Zhi, M. Lifeng, Weapon target assignment
[12] M. Ni, Z. Yu, F. Ma, X. Wu, A lagrange relaxation method for solving problem satisfying expected damage probabilities based on ant colony
weapon-target assignment problem, Math. Probl. Eng. 2011 (2011). algorithm, J. Syst. Eng. Electron. 19 (2008) 939–944.
[13] Z. Yao, M. Li, Z. Chen, R. Zhou, Mission decision-making method of [33] Y. Li, Y. Kou, Z. Li, A. Xu, Y. Chang, A modified pareto ant colony optimiza-
multi-aircraft cooperatively attacking multi-target based on game theoretic tion approach to solve biobjective weapon-target assignment problem, Int.
framework, Chin. J. Aeronaut. 29 (2016) 1685–1694. J. Aerosp. Eng. 2017 (2017).
[14] M.Z. Lee, Constrained weapon-target assignment: enhanced very large [34] J. Li, J. Chen, B. Xin, L. Dou, Z. Peng, Solving the uncertain multi-objective
scale neighborhood search algorithm, IEEE Trans. Syst. Man Cybern. A 40 multi-stage weapon target assignment problem via MOEA/d-AWA, in:
(2010) 198–204. Proceedings of IEEE Congress on Evolutionary Computation (CEC), IEEE,
[15] R.K. Ahuja, A. Kumar, K.C. Jha, J.B. Orlin, Exact and heuristic algorithms for Vancouver, BC, Canada, 2016, pp. 4934–4941.
the weapon-target assignment problem, Oper. Res. 55 (2007) 1136–1146. [35] Q. Pan, D. Zhou, Y. Tang, X. Li, A novel antagonistic weapon-target assign-
[16] A.M. Madni, M. Andrecut, Efficient heuristic approach to the weapon-target ment model considering uncertainty and its solution using decomposition
assignment problem, J. Aerosp. Comput. Inf. Commun. 6 (2009) 405–414. co-evolution algorithm, IEEE Access 7 (2019) 37498–37517.
 C.-M. Lai and T.-H. Wu / Applied Soft Computing Journal 82 (2019) 105542 15

[36] J. Li, J. Chen, B. Xin, L. Dou, Solving multi-objective multi-stage weapon [50] W.C. Yeh, Optimization of the disassembly sequencing problem on the
target assignment problem via adaptive NSGA-II and adaptive MOEA/d: basis of self-adaptive simplified swarm optimization, IEEE Trans. Syst. Man
A comparison study, in: Proceedings IEEE Congress on Evolutionary Cybern. A 42 (2012) 250–261.
Computation (CEC), IEEE, Sendai, Japan, 2015. [51] W.C. Yeh, C.M. Lai, K.H. Chang, A novel hybrid clustering approach based
[37] W.C. Yeh, A two-stage discrete particle swarm optimization for the on k-harmonic means using robust design, Neurocomputing 173 (2016)
problem of multiple multi-level redundancy allocation in series systems, 1720–1732.
Expert Syst. Appl. 36 (2009) 9192–9200. [52] Z. Michalewicz, Do not kill unfeasible individuals, in: Proceedings of
[38] W.C. Yeh, P.J. Lai, W.C. Lee, M.C. Chuang, Parallel-machine scheduling the Fourth Intelligent Information Systems Workshop, Citeseer, 1995, pp.
to minimize makespan with fuzzy processing times and learning effects, 110–123.
Inform. Sci. 269 (2014) 142–158. [53] M. Xie, T. Yamaguchi, T. Odaka, H. Ogura, An analysis of evolutionary states
[39] R. Azizipanah-Abarghooee, A new hybrid bacterial foraging and simplified in the GA with lethal genes, Trans. Inst. Electron. Inf. Commun. Eng. D-II
swarm optimization algorithm for practical optimal dynamic load dispatch, J79D-II (5) (1996) 870–878.
Int. J. Electr. Power Energy Syst. 49 (2013) 414–429. [54] W.C. Yeh, C.M. Lai, Accelerated simplified swarm optimization with
[40] R. Azizipanah-Abarghooee, T. Niknam, M. Gharibzadeh, F. Golestaneh, exploitation search scheme for data clustering, PLoS One 10 (2015)
Robust, Fast and optimal solution of practical economic dispatch by a e0137246.
new enhanced gradient-based simplified swarm optimisation algorithm, [55] Z. Juman, M. Hoque, An efficient heuristic to obtain a better initial feasible
IET Gener. Transm. Distrib. 7 (2013) 620–635. solution to the transportation problem, Appl. Soft Comput. 34 (2015)
[41] Y.Y. Chung, N. Wahid, A hybrid network intrusion detection system 813–826.
using simplified swarm optimization (SSO), Appl. Soft Comput. 12 (2012) [56] W.C. Yeh, A novel boundary swarm optimization method for reliability
3014–3022. redundancy allocation problems, Reliab. Eng. Syst. Saf. (2018).
[42] C.L. Huang, A particle-based simplified swarm optimization algorithm [57] Q. Kang, H. He, A novel discrete particle swarm optimization algo-
for reliability redundancy allocation problems, Reliab. Eng. Syst. Saf. 142 rithm for meta-task assignment in heterogeneous computing systems,
(2015) 221–230. Microprocessors Microsyst. 35 (2011) 10–17.
[43] C.M. Lai, Multi-objective simplified swarm optimization with weighting [58] Q.K. Pan, M.F. Tasgetiren, Y.C. Liang, A discrete particle swarm optimization
scheme for gene selection, Appl. Soft Comput. 65 (2018) 58–68. algorithm for the no-wait flowshop scheduling problem, Comput. Oper.
[44] C.M. Lai, W.C. Yeh, Y.C. Huang, Entropic simplified swarm optimization for Res. 35 (2008) 2807–2839.
the task assignment problem, Appl. Soft Comput. 58 (2017) 115–127. [59] M.S. Kiran, TSA: Tree-seed algorithm for continuous optimization, Expert
[45] C.M. Lai, W.C. Yeh, C.Y. Chang, Gene selection using information gain Syst. Appl. 42 (2015) 6686–6698.
and improved simplified swarm optimization, Neurocomputing 218 (2016) [60] M.S. Kıran, An implementation of tree-seed algorithm (TSA) for constrained
331–338. optimization, in: Intelligent and Evolutionary Systems, Springer, 2016, pp.
[46] W.C. Yeh, Orthogonal simplified swarm optimization for the series–parallel 189–197.
redundancy allocation problem with a mix of components, Knowl.-Based [61] A. Babalik, A.C. Cinar, M.S. Kiran, A modification of tree-seed algorithm
Syst. 64 (2014) 1–12. using Deb’s rules for constrained optimization, Appl. Soft Comput. 63
[47] P. Lin, S. Cheng, W. Yeh, Z. Chen, L. Wu, Parameters extraction of solar (2018) 289–305.
cell models using a modified simplified swarm optimization algorithm, [62] B. Xin, J. Chen, Z. Peng, L. Dou, J. Zhang, An efficient rule-based constructive
Sol. Energy 144 (2017) 594–603. heuristic to solve dynamic weapon-target assignment problem, IEEE Trans.
[48] C.L. Huang, A particle-based simplified swarm optimization algorithm Syst. Man Cybern. A 41 (2011) 598–606.
for reliability redundancy allocation problems, Reliab. Eng. Syst. Saf. 142 [63] J. Derrac, S. García, D. Molina, F. Herrera, A practical tutorial on the
(2015) 221–230. use of nonparametric statistical tests as a methodology for comparing
[49] W.C. Yeh, Y.M. Yeh, P.C. Chang, Y.C. Ke, V. Chung, Forecasting wind power evolutionary and swarm intelligence algorithms, Swarm Evol. Comput. 1
in the mai liao wind farm based on the multi-layer perceptron artificial (2011) 3–18.
neural network model with improved simplified swarm optimization, Int. [64] A. Benavoli, G. Corani, F. Mangili, Should we really use post-hoc tests based
J. Electr. Power Energy Syst. 55 (2014) 741–748. on mean-ranks?, J. Mach. Learn. Res. 17 (2016) 152–161.