Received April 29, 2021, accepted May 7, 2021, date of publication May 11, 2021, date of current version

May 20, 2021.

Digital Object Identifier 10.1109/ACCESS.2021.3079152

A Modified MOEA/D Algorithm for Solving

Bi-Objective Multi-Stage Weapon-Target
Assignment Problem
XIAOCHEN WU 1,2 , CHEN CHEN1,2 , AND SHUXIN DING 3
1 School of Automation, Beijing Institute of Technology, Beijing 100081, China
2 State Key Laboratory of Intelligent Control and Decision of Complex System, Beijing 100081, China
3 Signal and Communication Research Institute, China Academy of Railway Sciences Corporation Ltd., Beijing 100081, China

Corresponding author: Chen Chen ([email protected])

This work was supported by the National Natural Science Foundation of China (NSFC) under Grant 61773066 and Grant 62022015.

ABSTRACT In command of modern intelligent operations, in addition to solving the problem of multi-unit
coordinated task assignment, it is also necessary to obtain a suitable plan according to the needs of decision
makers. Based on these requirements, we established a multi-stage bi-objective weapon-target assignment
model, and designed a new algorithm with niche and region self-adaptive aggregation (named MOEA/
D-NRSA) based on the decomposition-based multi-objective evolutionary algorithm (MOEA/D) to obtain
richer solutions that meet the preferences of different decision makers. Compared with MOEA/D, MOEA/
D-NRSA has advantages in improving the convergence and maintaining the distribution of the solution.
On the one hand, it contains a population evolution method based on niche technology to obtain better
offspring; on the other hand, it has a new neighborhood selection and update strategy. This strategy first
clusters the individuals in the objective space to divide into different regions, in which the subproblems can
independently select the appropriate aggregation mode according to the clustering density of the region and
update its neighborhood. This strategy can improve the uneven distribution of individuals and maintain the
diversity and distribution of the population. Numerical experiments selected state-of-the-art algorithms for
comparison, which proved the superiority of MOEA/D-NRSA.

INDEX TERMS Multi-stage weapon target assignment (MWTA), decomposition-based multi-objective

evolutionary algorithm (MOEA/D), niche, clustering, ideal-nadir Tchebycheff approach.

I. INTRODUCTION WTA is divided into two categories: static WTA (SWTA)

The revolution of intelligent warfare has effectively organized and dynamic WTA (DWTA). Originally modeled by
various weapon platforms, thereby achieving a high level of Manne [4], the SWTA defines a scenario wherein a known
ability matching and increasing the chance of more efficient number of targets are detected and a finite number of
operations [1]. However, how to solve the task allocation weapons (interceptors), with known probabilities of suc-
problem and obtain more diverse solutions is still an urgent cessfully destroying the targets able for a single exchange.
problem to be solved. The weapon-target assignment (WTA) In SWTA, all weapons engage with targets in a single
problem we studied is crucial in the multi-platform coop- stage, and no subsequent actions are considered, since time
erative task allocation process. To obtain suitable solutions, is not considered in the problem. The DWTA problem is
we must start with this problem. Lloyd and Witsenhausen much more complicated than the SWTA, so the current
have shown that the WTA is NP-Complete [2], it involves research results mainly focus on SWTA. The original model
how to obtain a set of weapon-target pairs, which can satisfy was defined by Manne [4], based on this, researchers such
the decision maker’s combat objectives in terms of combat as Ahuja et al. [1], Lee et al. et al. [5], Karasakal [6],
effectiveness and loss [3]. and Kline et al. [7] proposed various new models with
reference to change battlefield requirements. In terms of
The associate editor coordinating the review of this manuscript and algorithms, genetic algorithms (GA) [6], [8], [9], ant colony
approving it for publication was Pavlos I. Lazaridis . algorithms (ACO) [10], [11], and other heuristic intelligent
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
71832 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 9, 2021
X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

optimization algorithms [12]–[14] are widely used in SWTA optimization problem. We would like to maximize combat
problems. results while ensuring fewer combat losses and develop an
However, due to the continuous development of war optimization model with the above two conflicting optimiza-
modes, DWTA problems and related models are more suit- tion goals.
able for describing the current war situation with a larger area, In terms of algorithms, after some researchers used
multiple combat platforms, and a larger time scale. DWTA MOEAs to solve the task assignment problem, we also used
contains a time dimension, which is a multi-stage global the MOEA/D framework and improved it to form a new
decision-making process. In solving the DWTA problem, all algorithm to solve the problem, aiming to generate more
stages of the offense or defense must be considered, and its differentiated and richer solutions.
distribution results must be solved. Less attention has been The main contributions of this paper can be summa-
given to the DWTA as compared to the SWTA. Thus, there rized as follows. Firstly, a bi-objective dynamic collaborative
are fewer heuristic algorithms shared among researchers. weapon-target assignment model is formulated. The objective
Currently, hybrid heuristics are often used to solve DWTA used to describe the cost in the model includes ammunition
problems. Xin et al. [15], [16], solved this problem by using consumption and considers the loss of the combat platform
virtual permutation (VP) and tabu search heuristics (TSH); itself. Secondly, a population evolution strategy based on
and Leboucher et al. [17] used Hungarian algorithm and niche technology suitable for the MOEA/D framework is
GA-PSO hybrid algorithm in solving DWTA problems. proposed. It selects parents based on the sharing degree
In the context of DWTA, this paper studies a kind of of individuals and produces excellent offspring. Thirdly,
derivative problem of DWTA, that is, the multi-stage weapon a neighborhood update strategy based on ideal point and nadir
target allocation problem (MWTA). This problem divides the point is proposed to maintain the diversity of evolutionary
entire combat process into different stages. Each stage needs groups and improve the distribution of solutions.
to make decisions based on the previous stage results and The remainder of the paper is organized as follows.
the changes in the battlefield in the new stage and allocate In Section II, we formulate the problem. The improved algo-
weapons to the target. The decision-making process is the rithm we designed is explained in Section III Some numerical
same at each stage, due to different problem scale (number of experiment is carried in Section IV. The conclusion and
targets, number of weapons, etc.), there are also differences future work are presented in Section V.
in computational complexity. The multi-stage weapon target
assignment (MWTA) problem is intermediate between the II. PROBLEM FORMULATION
SWTA one and the DWTA one. It also takes time windows Models of WTA problems depend on many factors, e.g.,
into account, but does not have a fully dynamic process as offense or defense strategies, features of targets and weapons,
the DWTA does. etc. The combat scenarios considered in this paper are as
At the level of problem models and algorithms, the objec- follows. A total of W weapon platforms with vet types attack
tive of traditional WTA problems is focused on operational an enemy area together. Our weapons need to explore, dis-
effects. It can also be called the target-based goal, which cover, and strike against enemy targets continuously. The
aims to maximize the expected damage of the targets, while entire offensive process of weapon-target assignment can be
the goal of the asset-based one is to minimize the expected divided into multiple stages, where one stage is the minimum
loss. Therefore, many researchers transform this problem combat time unit. Suppose there are S stages in total.
into a series of multi-objective optimization problems with The offense overview and timeline of each stage are shown
constraints. In the process of solving such problems, various in Figs. 1 and 2. In Fig. 1, the units of the enemy and
multi-objective evolutionary algorithms (MOEA), which are ours are represented by three colors of red, blue, and gray.
widely used as the primary method to approximate the true The red units are our offensive weapons, the blue and gray
Pareto front (PF) of MOPs, have also become the first choice units together constitute the enemy forces in the area of
of experts [18]–[20]. engagement. Among them, the blue ones are targets that are
As for the study of MOEAs, Schaffer [21] designed the known and are within our attack range, and the gray ones
first MOEA. Recently, proposed by Zhang and Li [22], are unknown targets or targets that cannot be attacked yet.
a branch of MOEA based on decomposition has become Throughout the offensive process, we stipulate that the length
increasingly popular. It decomposes a multi-objective prob- of each phase is equal. In stage t, our entities launch attacks
lem into a set of scale subproblems and optimizes them simul- on enemy forces within the range according to the current
taneously. Because of its advantages in solution efficiency, allocation plan.
many researchers have conducted in-depth research on it and In Fig. 2, St and St+1 are the starting times of stage t and
achieved certain results [23]–[27]. t +1, respectively. The start of the attack marks the beginning
Unlike the general WTA model, the model studied in this of stage t. In stage t, the tasks that need to be performed
paper considers the possible losses of the weapon platform include offense, advancement, and action adjustment. Each
itself. The task assignment process is to assign the appropriate stage contains dynamic events, which are the destruction of
weapons to the required targets. In the problem model design old targets (blue dot) and the appearance of new targets (red
process, it transforms the WTA problem into a bi-objective dot).

VOLUME 9, 2021 71833

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

FIGURE 1. Multi-stage weapon-target assignment scenario.

oncoming targets through all stages. The formulation of the

expected damage at stage t is expressed as:
T S YW
!
X Y
xij (s)
F1 (t) = vj 1 − 1 − pij (s) (1)
j=1 s=t i=1

where t and s are the indexes of offense stages, X t =

[Xt , Xt+1 , . . . , XS ] with Xt = xij (t) W ×T is the decision

FIGURE 2. Timeline of task assignment. matrix at stage t, and xij (t) is a binary decision variable taking
a value of one (i.e., xij (t) = 1 ) if weapon i is assigned to
Returning to Fig. 1, in stage (t + 1), the processing of target i at stage t, or zero (i.e., xij (t) = 0) otherwise. W (t) and
dynamic events generated in the previous stage has been T (t) represent the remaining number of weapons and targets
completed, the original blue targets were eliminated and dis- at stage t, respectively (W (1) = W , T (1) = T ). vj means
appeared, and the unknown gray targets were found to turn the threat value of target j. pij (s) denotes the probability that
blue. Afterward, it is necessary to continue to attack existing weapon i destroys target j at stage s, which is also called kill
targets according to the adjusted plan. probability. pij (s) and vj can be obtained in advance based on
Given the set of targets and the set of available weapons, the theory of shooting and performances of weapons.
our goal is to find the best weapon-to-target pairing to maxi-
mize the probability of damage, while minimizing operations B. CONSUMPTION
costs. Apart from satisfying the tactical requirement, a WTA deci-
Through the above description, the MWTA problem is sion should also cut down the operational costs. Therefore,
transformed into a bi-objective optimization model. The two when designing the second objective function, we would like
optimization goals are the overall damage probability and the to minimize ammunition consumption; besides, in the course
cost of operations. The objectives and related constraints will of military operations, each side struggles to preserve itself
be described below. and destroy the other. In the actual situation, considering
that the enemy may attack our weapon platform, we need to
A. EXPECTED DAMAGE OF ALL TARGETS minimize or even avoid enemy attacks on weapon platforms.
Based on the characteristics of the target-based model, When the second objective was designed, we comprehen-
the first objective is to maximize the total expected damage sively considered the above two requirements and combined

71834 VOLUME 9, 2021

X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

the minimization of ammunition consumption with platform actual situation. A minimum length is required to ensure that
losses to form a new goal. It can be presented as follows: the platforms can operate effectively at each stage. With the
S X
T X
W increase in the stage length, the computational complexity of
the model decreases, and when the length of a stage is equal to
X
F2 (t) =
 
vwi reji (s) + vai (s) xij (s) (2)
s=1 j=1 i=1 the total time on offense, the model becomes a static version
of the assignment model.
where S represents the total stage of the attack, T is the In the above model, constraint (6) is an important feature
number of targets, and W is the number of weapons. vai (s) that distinguishes MWTA from traditional SWTA. It reflects
denotes the ammunition consumption of weapon i at stage s, the impact of the time window of each stage on the avail-
and vwi is the value of weapons. reji (s) represents the ability ability of weapons. At the beginning of each stage, the Usij
of the target to resist our attacks at stage s, and its value can corresponding to weapon i needs to be updated.
be calculated by: The above model holds a critical assumption: after the
reji (s) = qji , (3) weapon is launched, its damage probability does not change.
This is reasonable since missiles usually have active terminal
where qji represents the probability of target j hitting i. guidance.
Through the above analysis, we get two optimization
objectives: III. DESIGN OF MOEA/D-NRSA
Since the MWTA problem to be solved in this paper is
max F1 (t) , min F2 (t) .
NP-hard [2], [28], it is impossible to obtain an accurate
solution through some polynomial time algorithm. Therefore,
C. CONSTRAINTS
we design a new algorithm based on the MOEA/D framework
W
X proposed by Zhang and Li et al. [22]. This framework decom-
xij (t) ≤ Wmj ∀j ∈ Ij , ∀t ∈ It (4) poses the multi-objective problem into multiple subproblems,
i=1 optimizes them separately, and then integrates the solutions of
T
X the subproblems to form a complete solution. Our algorithm
xij (t) ≤ Eui ∀i ∈ Ii , ∀t ∈ It (5) is called MOEA/D-NRSA. ‘‘N’’ means niche technology,
j=1 which we used for the evolution process of populations, and
T X
X S ‘‘RSA’’ stands for region self-adaptive aggregation method.
xij (t) ≤ Ni ∀i ∈ Ii (6) We divide the objective space into different regions, and
j=1 t=1 independently select the appropriate aggregation function to
Ii = {1, 2, . . . , W }; Ij = {1, 2, . . . , T }; guide the convergence direction of the population according
It = {1, 2, . . . , S} to the region’s characteristics. So MOEA/D-NRSA is mainly
xij (t) ≤ Usij (t) ∀i ∈ Ii , ∀j ∈ Ij , ∀t ∈ It (7) dedicated to improving the convergence to a certain extent
while maintaining the distribution of the solution. The frame-
Constraints (4) and (5) are the feasibility constraints of our work is given in Algorithm 1. We will discuss this in detail in
weapon platform. Constraint (4) indicates that the number of the following subsections.
weapons that can attack target j at each stage cannot exceed
Wmj . This constraint is linked to ammunition consumption A. REGIONAL DIVISION AND CLASSIFICATION
for each target at each stage. Constraint (5) means that the When designing algorithms based on the MOEA/D frame-
maximum number of units attacked by each weapon in a work to solve bi-objective optimization problems, two main
single stage is Eui . In most cases, a single weapon can only factors need to be considered. The first is to maintain the
attack one target, therefore, Eui is set to 1 in this paper. diversity of the evolutionary group, which is one of the most
Constraint (6) indicates the amount of available ammunitions important indicators to measure the pros and cons of the
of weapon i. The last constraint limits the use of weapon i, algorithm; the second is to improve the convergence speed
where Usi is a binary variable like xij , but Usij is used to of the algorithm, which is closely related to the efficiency of
constrain the availability of i at stage t. When weapon i cannot solving actual problems.
attack j (e.g., beyond the range or exhausted ammunition), let In maintaining the diversity of evolutionary populations,
Usij be 1; otherwise, Usij = 0. the clustering method is used by many experts and schol-
The deterministic model is given as follows: ars [29], but generally, this method is directly used to generate
( MOEA. This paper also adopts the clustering algorithm, but
max F1 (t) it is only used to classify the solutions into different regions
min F2 (t) as the basis for subsequent operations to maintain population
s.t. (4), (5), (6) and (7). diversity.

The proposed model divides the offense time interval into 1) CLUSTERING METHOD
several fixed stages. During the solution process, the length of In the iterative process of the evolutionary algorithm,
each stage is flexible and can be determined according to the as the mutation and replacement operations proceed,

VOLUME 9, 2021 71835

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

Algorithm 1 The Framework of MOEA/D-NRSA

Input:
• F1 , F2 : two objectives;
• N : population number (number of subproblems);
• Tm : size of mating neighborhood;
• EP: an elite population used to preserve the Pareto solution;
• MaxIter: maximum iteration;
• nReg: the number of regions generated according to the clustering results;
• thd: threshold of cluster density;
• Stopping criterion;
Output:
• EP: final elite population;
• FV : fitness;
Step 1. Initialization
1: Let EP = ∅;
2: Create an initial population Pop = x 1 , . . . , x N by uniformly randomly sampling. Evaluate the fitness value FV i of each


solution x i , FV i = F1 x i , F2 x i and set FV = FV 1 , . . . , FV N .



3: Generate N evenly distributed weight vectors λ = {λ1 , . . . , λN }. Find the Tm closest weight vectors to each weight vector
based on the Euclidean distances of any two weight vectors. Denote by NS(i) = i1 , . . . , iTm the neighbor set of the ith
subproblem.  
4: Initialize the ideal point z = (z1 , z2 )T , required by the Tchebysheff approach, set zi = minj FVji , and nadir point znad
i =
 
maxj FVji , i = 1, . . . , m.
Step 2. Update
1: while MaxIter is not reached do
2: nReg is the number of cluster centers. Use Algorithms 2 and 3 to cluster all the numbers in FV and generate nReg
regions.
3: for each region k do
4: The degree of aggregation in region k is calculated by Algorithm 4.
5: Update the flag fl k of region k according to the threshold thd of the degree of aggregation.
6: for each subproblem i and FV i in region k do
7: Two parents pa1 , pa2 are selected in the neighborhood by Algorithm 5.
8: Generate offspring of i through parents and
Algorithms 6 and 7. And we evaluate its fitness foi .
9: zi = min (zi , foi ) , znad
i = max z nad , fo , i = 1, . . . , m
i i
10: Through Algorithm 8 and the value of fl k , update the neighborhood of i.
11: end for
12: end for
13: Update EP.
14: end while

new offspring are produced in new positions. At this In Algorithm 1, we define nReg as the number of regions,
time, the solution distribution may be uneven. Most so in Algorithm 2, we randomly generate nReg cluster centers
of the existing processing methods are oriented to the cci . d(i, j) represents the degree of dissimilarity between i and
entire solution set. This paper uses clustering methods to j, expressed by the Euclidean distance between the two points.
divide individuals into regions with different aggregation q
2 2
degrees for processing separately, thereby maintaining the d(i, j) = xi1 − xj1 + · · · + xip − xjp (8)
group’s diversity. The specific clustering rules are given by
where i and j are both p-dimensional vectors.
Algorithm 2.
The evaluation value Ev can be calculated by:
Algorithm 2 first randomly selects a center point for each X
class, and the remaining individuals are merged into adjacent Evi = |cci − p|2 (9)
classes according to the distance from each class. Then re- p∈cli
select the clustering points, form a new class according to the nReg
X
minimum distance principle, and finally repeat this process Ev = Evi . (10)
until the clustering process is stable. i=1

71836 VOLUME 9, 2021

X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

Algorithm 2 Clustering Rules

Input: nReg, FV .
Output: cc, cl.
1: Set cluster center cc, cc = cc1 , cc2 , . . . , ccnReg .


2: Set the corresponding class as cli , i ∈ {1, 2, . . . , nReg}.

3: For  i  ∈
{1, 2,  . . . , N
}, let bo =
max max F1 x i , max F2 x i ;
4: On the corresponding coordinate axis, using bo as the
boundary, the solution space is equally divided into nReg
intervals.
5: Let Ev, Ev0 = 0.
6: while i ≤ nReg do
7: h Randomlyi select an individual in interval
(i−1)bo i·bo
nReg , nReg as cci ;
8: Calculate d (cci , p) ; cci , p ∈ FV , cci 6 = p;
9: Let some p with the most remarkable similarity join
cli ;
10: Calculate Ev, Ev = Ev + Evi ; FIGURE 3. Clustering scenario of solution space.
11: i = i + 1;
Algorithm 3 Boundary Calculation
12: end while
13: while Evmin is not reached do Input: pt, pd, pl, pr, nReg, cc.
14: Ev = Ev0 ; Output: lb1i , lb2i , ub1i , ub2i .
15: for each qj ∈ clj do 1: if i = 1 then
2: 1
xri−1 = −xli1 ;
16: Randomly choose qj , qj 6 = ccj ;
17: Calculate d(qj , p), qj ∈ cli , p ∈ Pop, qj 6 = p; 3: end if
18: Calculate Ev0 , Ev0 = Ev0 + Ev0j ; 4: if i = nReg then
5: 1
xli+1 = xri1 ;
19: end for
20: if Ev0 < Ev then 6: end if
1 ≤ x1
7: if xri
21: for k ≤ nReg do li+1 then

x1 −x 1
22: cck = qj ; 8: ub2i = li+12 ri ;
23: end for 9: ub1i = xti2 ;
24: end if 10: 1
if xri−1 ≤ x1li then

25: end while x 1 −xri−1
1
11: lb2i = li 2 ;
12: 2;
lb1i = xdi
13: else
After obtaining the clustering results, we divide the 14: lb2i = xli1 ;
operation area according to it. As shown in Fig. 3, 2 −x 2


xdi ti−1
15: lb1i = 2
we divide regions in the objective space, and each region has
16: end if
4 boundaries (lb1 ) , (lb2 ) , (ub1 ) and (ub2 ). Two diagonals
17: else
(lb1 , lb2 ) , (ub1 , ub2 ) are generally used to identify a region.
18: ub2i = xri1 ;
Suppose the four boundaries of region i are (lb1i ), x2 −x 2


(lb2i ), (ub1i ) and (ub2i ), and lb1i = di+1 ti
;

the farthest
points in
the
19: 2
1 , x 2 , pd x 1 , x 2 , pl x 1 , x 2 , and 1 1
if xri−1 ≤ xli then

four directions are pt i x ti ti i di di i li li
20:
pri xri1 , xri2 ,
respectively. The center point of the i-th cluster is x 1 −xri−1

1

1 , x 2 . The boundaries of each region can be obtained

21: lb2i = li 2 ;
cci xcci cci 22: 2;
lb1i = xdi
by Algorithm 3. 23: else
24: lb2i = xli1 ;
2) CLUSTER DENSITY CALCULATION 2 −x 2


xdi ti−1
25: lb1i = 2 ;
After completing the regional division, we introduce the con-
26: end if
cept of influence to calculate the aggregation density of each
27: end if
subspace.
Here, the influence of a certain solution j in the space on
the i-th solution is described as: Among them, lj→i represents the Euclidean distance of
individual j to i, and γ is the mapping function, which trans-
γ lj→i : R → R


(11) forms lj→i into the influence of j on i.

VOLUME 9, 2021 71837

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

Therefore, the cluster density D(a) of the region a can degree as the weight of the individual. We use Si to represent
be defined as the mean value of all points in the region to the sharing degree of individual i in the group, then:
influence the center point. After getting D(a), we will classify X
the region a according to the density threshold thd. Si = sh[d(i, j)], (14)
Pm j∈NSi
γ (j, i)
D(a) = i=1 (12)
m where d(i, j) can be obtained by (8), sh is the sharing function
The specific algorithm is as follows: that measures the similarity level between individuals i and j,
which is defined as:
Algorithm 4 Cluster Density Calculation 
 0, d(i, j) > σshare
Input: a, cca , thd. sh[d(i, j)] = d(i, j) (15)
Output: D(a), fl a . 1 − , d(i, j) < σshare
σshare
1: Count the number of solutions in a as m;
2: Let the solution set in a be Sa , Sa = {Sa1 , . . . , Sam }; where σshare is a predefined niche radius [31]. We set
3: Let SumD(a) = 0. the value of σshare to Tm /2, and Tm is the size of the
4: while i ≤ m do neighborhood.
Calculate γ lSai →cca ;


5: The higher the sharing degree Si of an individual,
SumD(a) = SumD(a) + γ lSai →cca ;


6: the greater the similarity between it and all other individuals
7: i = i + 1; in the neighborhood, and the more hopeful it is to generate
8: end while high-quality offspring for the current subproblem. The weight
9: D(a) = SumD(a)/m; wi of individual i in the neighborhood can be defined as:
10: if D(a) ≤ thd then
fl a = 1; Si
11: wi = P , i ∈ NSi . (16)
12: else i Si
13: fl a = 0;
The specific algorithm is as follows:
14: end if

Algorithm 5 Select Parents

In Algorithm 4, γ lj→i can be calculated by Gaussian



influence function: Input: Sj , j ∈ NSi .
2
Output: pa1 , pa2 .
1 − r
γ lj→i = γ (r) = √ e 2σ 2 1: for k = 1 : 2 do


(13)
σ 2π 2:
S
wj = P jS , j ∈ NSi .
j j
where r represents Euclidean distance of individual j to i. The 3: Choose one subproblem  pak ∈ NSi based on the
threshold thd is the concentration density of the true PF. probability distribution wj .
4: Set wpak = 0.
B. MODIFICATION OF MOEA/D FRAMEWORK 5: end for
In this subsection, the Ideal-Nadir Tchebycheff approach that
can modify the general MOEA/D framework is designed. After getting the parents, we need to generate offspring.
In addition, a new neighborhood update and population evo- The specific details are given by Algorithm 6.
lution strategy has also been proposed to generate more good
offspring, thereby improving the overall performance of the
Algorithm 6 Genetic Operator
algorithm.
Input: xpa1 and xpa2 , wpa1 and wpa2 .
1) POPULATION EVOLUTION STRATEGY
Output: Offspring of
wpa
1: pc = w + 1w
In the basic MOEA/D framework, a subproblem and its pa1 pa2
2: for each genetic locus l do
neighbor have the exact equivalence when mating them to
generate a new solution. This approach will make the process 3: if rand < pc then
4: l
of (l) = xpa
of selecting parent solutions more random and affect the 1

convergence speed of the population. Therefore, when select- 5: else

6: l
of (l) = xpa
ing the parent class, we hope that the weights of different 2

individuals in the neighborhood are different, and the quality 7: end if

8: end for
of the solution can be improved through appropriate weight
settings.
Based on this idea, we propose a niche-guided matching The offsprings generated by Algorithm 6 may not satisfy
method. This method refers to the niche technology based on the constraints. In this case, we need to repair the solu-
the sharing mechanism [30], and introduces the group sharing tion [32]. The random repair mechanism used is as follows:

71838 VOLUME 9, 2021

X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

Algorithm 7 Random Repair Mechanism population convergence. Unlike the ideal-based Tchebycheff,
Input: An infeasible solution of , constraints. this method strives to make the optimization objective prin-
Output: A feasible solution of 0 ciple of each subproblem to the nadir point in the process of
1: Transform of into the corresponding 0-1 matrix X ; convergence. the scalar optimization subproblem i may also
2: Find the rows or columns that violate this constraint; be formulated by:
3: Calculate the number of redundant 1s in these   n o
rows or columns, denote as num. max gnte x | λ, znad = min λi znad i − FV i
(x) (18)
1≤i≤m
4: Randomly replace the num 1s in these rows or columns
with 0 to obtain X 0 . where FV is the objective function vector of a solution, znad
5: Retransform the 0-1 matrix X 0 into of 0 . is the nadir point. Under the guidance of the nadir point,
6: Regionk is a region where the target value FV i of sub- the population will have better sparsity.
problem i is located. (17) and (18) are often only suitable for MOPs with nor-
0
7: Calculate FV of . malized objective functions. Thus, when the ranges of the
0
8: if FV of is not in Regionk then objectives are on very different scales, we normalize the
9: Regenerate offspring of . objective value to avoid the influence of different dimensions.
10: end if The subproblem is defined in the following form:
For ideal point:
( )
FV i (x) − zi
2) IDEAL-NADIR TCHEBYCHEFF APPROACH min g (x | λ, z) = max λi
te
(19)
1≤i≤m znad
i − zi
For a long time in the past, many methods to decompose MOP
into scalar optimization subproblems have been proposed, For nadir point:
such as the weighted Tchebycheff approach [22], NPI-style ( )
Tchebycheff approach [33], ε-constraint approach [24],
  znad − FV i (x)
max g nte
x | λ, z nad
= min λi i
(20)
angle-based approach [34], etc. The most representative of 1≤i≤m znad
i − zi
the methods is the weighted Tchebycheff approach [22].
Let
These approaches have their advantages and disadvan-
i
tages, and they apply to different problems. This subsection ˜ ite = FV − Zi ,
FV
mainly discusses the Tchebycheff approach and its improve- nad
Zi − Zi
ment measures. A scalar optimization subproblem based on Zinad − FV i
the weighted Tchebycheff approach with the ideal point is ˜ inte =
FV ,
determined by: Zinad − Zi
i = 1, . . . , m, (21)
min gte (x | λ, z)
n o the optimization goal becomes:
= max λi FV i (x) − zi
1≤i≤m ˜ te | λ, 0).
min gte (FV (22)
s.t. x ∈ Pop (17) ˜ nte | λ, 0).
max gnte (FV (23)
where λ = (λ1 , . . . , λm )T is the weight vector of the scalar This section has designed an Ideal-Nadir Tchebycheff
optimization subproblem, and z = min FV i (x) | x ∈ Pop , approach (INT), which uses both the ideal point and the nadir
i ∈ {1, 2, . . . , m} is the ideal point. point to guide the convergence of the population. Its working
The problems of the Ideal-based Tchebycheff approach are principle is shown in Fig. 4. In the figure, according to the
as follows: clustering rules designed in Section III-A, there are different
(1) As the iterative process continues to advance, points regions in the objective space, and the individual aggregation
close to z will be retained, which will cause the solutions density of each region is different. We can judge its aggre-
of the subproblems to continue to gather in the direction gation density according to the signs fl i of each region, and
of z. When the distance between the solutions of the sub- then choose the appropriate reference point and aggregation
problems corresponding to different λ gradually shrinks, function to guide the convergence of the individual. For areas
it means that the neighborhood structure is destroyed. with high concentration density (RE3 and RE4 ), choose the
(2) The subproblem solution continues to converge in the ideal point as the reference point, otherwise (for RE1 and
direction of z will bring another problem. For a convex RE2 ) choose the nadir point as the reference point. In this
Pareto front, this trend will destroy the diversity of the way, it can be ensured that the solutions are distributed more
evolutionary population, thereby affecting the distribu- widely and more uniformly in the objective space.
tion of the solution and reducing the reliability of the In the figure, λ1 to λ4 are the weights corresponding to
solution [25]. the subproblems. In a two-dimensional space, the relation-
To avoid the above problems, we considered the nadir- ship between the weight λi and the essential direction λ0i of
subproblem i is: (λi1 , λi2 ) = λi2 , λi1 .
0 0


based Tchebycheff proposed by [35] in the process of

VOLUME 9, 2021 71839

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

FIGURE 6. Neighborhood update strategy.

FIGURE 4. Ideal-Nadir Tchebycheff approach.

λN +1−i
Na = λNI
+1−i
, the different weight vectors correspond-
ing to subproblems i can all be represented by cluster λI , and
(20) becomes:
 
max gnte x | λ, znad
( )
znad − FV i (x)
= min λN −i+1 i
(24)
1≤i≤m znad
i − zi

3) NEIGHBORHOOD UPDATE RULES

The idea of the traditional MOEA/D neighborhood update
strategy is as follows: optimize each subproblem by using
the information of several adjacent subproblems, make these
problems develop in a favorable direction, and finally con-
verge to the Pareto frontier.
The neighborhood update strategy of MOEA/D generally
takes the aggregate function value of the solution and the
subproblem as the fitness value, and replaces all inferior
FIGURE 5. Correspondence between subproblems and weight vectors. solutions in the neighborhood by comparing the fitness value
of the new solution with other solutions in its neighborhood.
In the process of using INT, there will also be a problem- As shown in Fig. 6, the aggregate function value of FV 0 is
the unification of the weight vector. When the subproblem i calculated by (17). Through comparison in the neighborhood,
updates the neighborhood based on the ideal point z or the it can be seen that x3 , x4 , and x5 are all inferior solutions
nadir point znad , it corresponds to different weight vectors compared to FV 0 and need to be replaced by FV 0 . Although
λi or λ0i . What we need to do is to integrate λiI and λiNa into this operation will speed up the convergence rate of the
the same weight vector cluster λ. population, it will also destroy the diversity of the population.
Fig. 5 shows the objective space, the weight vectors λ1I , As the iteration progresses, the number of effective solutions
λI , λ1Na and λN
N
Na are set to (,1-), (1-, ), (,1-) and (1-, in other neighborhoods will decrease sharply, which reduces
), where  is a very small number to avoid abnormal selec- the number of parent samples that can be selected during
tion. In the objective space, generate N uniformly distributed the evolution of some subproblems and affects the evolution
weight vectors along the clockwise direction to correspond to efficiency of the population, making the algorithm fall into a
N subproblems, respectively. Because λ1I = λ1Na and λN I = local optimum.
λN ,
Na Iλ i = λi can be obtained.
Na Therefore, when we update the neighborhood, we need to
For the subproblem i in the figure, when the ideal point selectively eliminate inferior solutions instead of replacing
is selected, its corresponding weight vector is λiI ; when the them all. In this way, the diversity of the population can be
nadir point is selected, its weight vector is λN Na
+1−i
. Since maintained to a certain extent.

71840 VOLUME 9, 2021

X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

As an important part of MOEA/D, the neighborhood Algorithm 8 Neighborhood Update Strategy

structure plays an important role in iteration and popula- Input: NSi , of i , λofi , fo, fl k .
tion update. In the basic MOEA/D framework, the solu- Output: Pop, FV .

tions to adjacent subproblems are considered similar to each 1: zi = min zi , FV i ;
2: znad = max znadi , FV ;
i


other. A further consideration is that the closer the two i
subproblems are, the more similar their solutions will be. 3: Let the set of inferior solutions be Iso , Iso = ∅;
Based on this idea, a neighborhood screening strategy is pro- 4: if fl k = 1 then
posed here. We define the relationship between subproblems i 5: for each subproblem
  do
 j ∈ NS(i) 
and j as [36]: 6: ˜ | λj , 0 < gte FV
if gte fo ˜ jte | λj , 0 then
2
Rij = e−δ kλi −λj k , (25) 7: Let x j merge into Iso ;
8: end if
where λi and λj are the weight vectors of subproblems i and 9: end for
j, respectively. δ is a parameter used to adjust the size of Rij . 10: else
If we count all problems j related to subproblems i, and sum for each subproblem
11:  j j ∈ NS(i) do 
all Rij . Then the proportion rij of any subproblem j in all
12: if g nte ˜
FV | λN +1−j , 0 <
relations of i can be defined as:   nte
Rij g nte ˜ | λN +1−j , 0 then
fo
rij = P , j = 1, . . . , IS. (26)
j Rij 13: Let x j merge into Iso ;
14: end if
Combining the contents of the previous subsection and this
15: end for
subsection, we can obtain a complete neighborhood update
16: end if
strategy, and the details are given in Algorithm 8.
17: Count the number of elements contained in Iso and record
Algorithm 8 combines the contents of the two subsec-
it as m.
tions. First, select the individual’s convergence mode (ideal
18: for each inferior solution q do
point or nadir point) according to the flag fl k of the region k.
19: Calculate Rjq .
Then, compare the offspring of i with the points in the neigh-
20: end for
borhood NS(i), and count the number of inferior solutions. R
21: rjq = P jqR , q = 1, . . . , m.
Finally, the probability distribution can be obtained according q jq
to (26). Choose the solution that needs to be eliminated. 22: for each inferior solution q do
According to (25) and (26), we can understand that the 23: Choose inferior solution based on the probability
closer the inferior solution is to the offspring, the easier it distribution rjq .
is to be eliminated; otherwise, the corresponding subproblem 24: if Inferior solution q is selected then
is more likely to be retained. Through this rule, the diversity 25: x q = of i , FV q = fo;
of the solution set can be maintained. 26: end if
27: end for
C. COMPLEXITY ANALYSIS
In this subsection, We use N to denote the population size
(Pop), m to denotes the number of objectives.
There are two parts in MOEA/D-NRSA, the clustering resource allocation (MOEA/D-DRA) [39]. For the original
process and the process of generating non-dominated solu- versions of MOEA/D and NSGA-II [40], we apply the unified
tions. Suppose the number of iterations of the clustering crossover and random mutation operators to the algorithm
process is Irc , then the time complexity of clustering algo- and compare them with MOEAD-NRSA.
rithm is O (m · N · nReg · Irc ). The process of generating non-
dominated solutions is consistent with MOEA/D which is B. ENCODING
O (m · N · Tm ). So, the time complexity of MOEA/D-NRSA This paper adopts decimal encoding. The length of the chro-
is O [m · N · (nReg · Irc + Tm )]. mosome is the total number of different types of weapon
platforms. Each weapon platform is regarded as a gene locus,
IV. EXPERIMENT AND RESULT ANALYSIS and the gene value on it indicates the number of targets that
A. COMPARISON ALGORITHM the weapon is assigned to. The specific coding form and
In order to prove the effectiveness of the improve- operation process are the same as those introduced in [41].
ment measures, we compared MOEA/D-NRSA with five Different types of weapons are not coded separately but
other algorithms, including multi-objective particle swarm are integrated into the same chromosome. Their difference
optimization algorithm (MOPSO) [37], non-dominated sort- lies only in the probability of destruction, the value of the
ing genetic algorithm II (NSGA-II), and two improved ammunition fired, and the value of the weapon itself.
MOEA/D algorithms: MOEA/D with adaptive weight adjust- Such an encoding method can guarantee that every solu-
ment (MOEA/D-AWA) [38] and MOEA/D with dynamical tion satisfies the constraint (5) naturally. As to the other

VOLUME 9, 2021 71841

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

constraints, each solution will be randomly repaired by TABLE 1. Public parameters of algorithms.
Algorithm 7 to satisfy them.

C. PERFORMANCE METRICS
So far, there is no single performance indicator that can
comprehensively measure the performance of an MOEA.
Therefore, we have introduced several famous metrics to
compare the performance of different algorithms. TABLE 2. Parameters of different scenarios.

1) INVERTED GENERATIONAL DISTANCE (IGD) [42]

The IGD metric can measure the diversity and convergence of
solutions simultaneously, and it can give the average distance
from a given set of non-dominated solutions to the true Pareto
front. The smaller the value of IGD (P, P∗ ), the better the
performance of P. The formulation is as follows:
P
∗ d(v, P)
IGD P, P = v∈P ∗
∗


(27)
|P |
P∗ denotes a set of uniformly distributed points in the
objective space along the true Pareto front (PF) or nearly
true PF when it is hard or impossible to get the true PF; The
set of non-dominated solutions obtained by all comparison For all algorithms, we first set their public parameters of
algorithms is used as the true Pareto front in this paper. d(v, P) algorithms in Table 1:
is the minimum Euclidean distance between v and elements Pop represents the population size, and Gen is the number
in P. of iterations.
The unique parameter settings in the comparison algorithm
2) GENERATIONAL DISTANCE (GD) [42] are as follows:
The GD metric measures the average distance from an (1) Parameter settings in MOPSO adopted here are the
inverted perspective. This metric is more sensitive to the con- same as those claimed in [44].
vergence of the solution. The smaller the value of GD (P, P∗ ), • The inertia weight: w = 0.4.
the better the convergence of P. Its formulation is as follows: • The acceleration constants: c1 = c2 = 1.4962.
P
d̃ (v, P∗ ) (2) Parameter settings in NSGA-II [40]:
GD P, P∗ = v∈P


(28) • Probability for crossover: cr = 0.9.
|P|
• Probability for mutation: mr = 1/n. n represents the
3) DIVERSIFICATION METRIC (DM) [43] number of decision variables.
This metric measures the spread of non-dominated solutions (3) Parameter Settings in MOEA/D, MOEA/D-AWA, and
on the Pareto front. The larger the value of DM, the wider MOEA/D-DRA adopted here are the same as those claimed
the solution of this method is distributed in the target space, in [22], [38], and [39].
achieving a better approximation of the Pareto front. It is • Neighborhood size: Tm = b0.1 N c.
calculated as follows: • Probability of selecting mate solutions: δ = 0.9.
v
u m
uX • Maximal number of replacement: nr = b0.01 N c.
DM = t (min fi − max fi )2 (29) The Table 2 gives the relevant parameters of each scenario.
i=1 Table 3 shows the settings of the public simulation param-
where min fi and max fi are the minimum and the maximum eters required for the simulation, where the value expressed
value of each fitness function among all non-dominated solu- in the form of an interval represents that the value of the cor-
tions obtained by the algorithms. responding parameter is randomly generated in this interval.

D. PARAMETER SETTINGS AND TEST SCENARIOS E. RESULT ANALYSIS

In the numerical experiment, we simulated six differ- In this subsection, we perform numerical experiments to com-
ent offense scenarios, each of which contains an instance pare six different algorithms under six different scenarios and
to test the algorithm’s performance in different problem analyze the results. Each scenario corresponds to an instance.
scales. Through simulation results and algorithm comparison, Independent runs are performed on each instance 25 times for
we can judge whether MOEA/D-NRSA can generate distri- all algorithms. The program is implemented using MATLAB
bution plans that are highly differentiated and meet the needs 2016b software and run on a desktop with 3.4 GHz Core
of different decision makers. i5-7500 CPU and 8.00 GB RAM. Fig. 7 shows the PF of

71842 VOLUME 9, 2021

X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

TABLE 3. Setting of public simulation parameters.

TABLE 4. Comparison of IGD.

a single run with minimum IGD value on six instances in metrics IGD, GD, and DM are presented in Tables 4, 5, and 6,
different scenarios. respectively. Fig. 7 shows the PF of a single run with mini-
For the instances corresponding to each scenario, we give mum IGD value on 6 instances in different scenarios. The
the mean values of the three metrics under different average ranking of the metrics and the statistics of Wilcoxon’s
algorithms. In addition, Wilcoxon’s rank-sum test with a test results are given in Table 7. Fig. 8 shows the comparison
5% significance level was performed to compare whether of IGD box plots of different algorithms in 6 instances.
the difference between the mean values of the metrics of Through the simulation results, we can understand that
MOEAD-NRSA and other algorithms is significant. The when solving small-scale problems, NSGA-II ranks first in
symbols †, §, and ≈ indicate that the performance of the three metrics of IGD, GD, and DM; and the algorithm
MOEAD-NRSA is better than, worse than, or similar to that based on the MOEA/D framework is better in Scenario 1.
of the comparison algorithm according to Wilcoxon’s rank- Performance is unsatisfactory. The main reason is that the
sum test, respectively. We set it in bold for the best average small size of the problem means that the search space is
metric value in each scenario; the standard deviation is in small, and there is no noticeable difference in the weight λ
parentheses below the average. The statistical results for the corresponding to different subproblems, which will cause the

VOLUME 9, 2021 71843

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

FIGURE 7. (a)–(f): PF of a single run with minimum IGD value on 6 instances in different scenarios.

offspring to lose diversity and make the problem fall into the The advantages of MOEA/D-NRSA are undeniable in
local optimum. And NSGA-II, which includes the principle the distribution of target space points. In Fig. 7, except for
of random matching, has more advantages. As the problem instance 1, in other medium-scale and large-scale problems
scale expands, the advantages of MOEA/D algorithms begin (Scenario 2 - 6), MOEA/D-NRSA can obtain solutions that
to manifest, especially MOEA/D-NRSA, which ranks first are closer to the true Pareto front. Especially at the tail of the
among multiple metrics. Pareto front, MOEA/D-NRSA obtains more non-dominated

71844 VOLUME 9, 2021

X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

TABLE 5. Comparison of GD.

TABLE 6. Comparison of DM.

solutions than other algorithms. The IGD and DM values in Table 3. The GD measurement value of MOEA/D-NRSA
in Tables 4 and 6 can also illustrate this advantage, which in Table 5 does not have a great advantage. After statistics,
reflects the effectiveness of the INT method. The INT method the GD value of MOEA/D-AWA is better. However, accord-
can obtain a better approximation of the whole Pareto front ing to the results of Wilcoxon’s rank-sum test, it can be
and make the distribution of the solution more even, maintain- seen that the GD measures of MOEA/D-NRSA and MOEA/
ing the diversity of the evolutionary population to a certain D-AWA are similar, there is no significant difference between
extent. the GD measures of MOEA/D-NRSA and MOEA/D-AWA,
However, the improvement of DM metrics means the indicating that the two are at the same level. The numer-
decline of GD metrics. When the metric value of DM ical approximation of the GD metric also shows that the
increases, the area equivalent to P that needs to be calculated convergence of MOEA/D-NRSA is only slightly inferior to
by (28) becomes more extensive, so the metric value of MOEA/D-AWA. This result proves the effectiveness of our
GD will become worse. This phenomenon is also reflected proposed population evolution strategy to a certain extent.

VOLUME 9, 2021 71845

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

FIGURE 8. (a)–(f): Comparison of IGD box plots of different algorithms in 6 instances.

TABLE 7. Overall performance of four algorithms on the six instances in The mean rank and the total count of †/§/≈ presented
terms of IGD, GD, and DM metrics.
in Table 7 shows the statistical results of the three metrics,
MOEA/D-NRSA ranks first among all algorithms. NSGA-II
has the same performance as MOEA/D, and they have certain
advantages in solving small-scale and large-scale problems,
respectively. MOEA/D-AWA ranks second with its excellent
performance in IGD and GD metrics, and its GD metrics
are even better than MOEA/D-NRSA on medium-scale
issues.

V. CONCLUSION AND FUTURE WORK

This paper mainly studies the modeling and solving methods
When the advantages of IGD and DM metrics are apparent, of heterogeneous weapon platforms’ cooperative task assign-
the GD metrics can be maintained at a high level. ment problem. Firstly, we propose a bi-objective optimiza-
Fig. 8 shows the comparison of IGD box plots of different tion model that can describe the multi-stage weapon-target
algorithms in 6 instances. These box plots cover the mean assignment problem. The model considers the probabil-
value and variation interval of the IGD metric values obtained ity of damage, ammunition consumption, and the loss
by 6 algorithms in a total of 150 simulations in 6 examples. of the platform itself. Secondly, to effectively solve the
By comparison, we can find that the change range of IGD of model, the MOEA/D-NRSA method obtained by improving
MOEA/D-NRSA is relatively tiny in the medium-scale and the MOEA/D framework is proposed. In MOEA/D-NRSA,
large-scale problems. Only MOEA/D-AWA has less fluctu- the convergence and distribution of the solution are consid-
ations in large-scale issues than MOEA/D-NRSA. It shows ered at the same time. In terms of convergence, a population
that the algorithm’s stability is better, and it is easier to obtain evolution mechanism based on niche technology is pro-
ideal results during operation. posed. This mechanism can make better use of neighborhood

71846 VOLUME 9, 2021

X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

information and spontaneously select suitable parents and [10] W. Yanxia, Q. Longjun, G. Zhi, and M. Lifeng, ‘‘Weapon target assignment
produce offspring through the sharing of individuals. problem satisfying expected damage probabilities based on ant colony
algorithm,’’ J. Syst. Eng. Electron., vol. 19, no. 5, pp. 939–944, Oct. 2008.
On the other hand, the introduction of the INT method [11] G. Shang, ‘‘Solving weapon-target assignment problems by a new ant
achieves a better approximation to the Pareto front and colony algorithm,’’ in Proc. Int. Symp. Comput. Intell. Design, vol. 1,
maintains the diversity of solutions. Finally, the simulation Oct. 2008, pp. 221–224.
[12] R. R. Hill and E. A. Pohl, ‘‘Heuristics and their use in military modeling,’’
results show the advantages of MOEA/D-NRSA over other in Wiley Encyclopedia of Operations Research and Management Science.
comparative algorithms. It does not perform well on small- Hoboken, NJ, USA: Wiley, 2010.
scale problems, but on medium-scale and large-scale prob- [13] A. Tokgöz and S. Bulkan, ‘‘Weapon target assignment with combinatorial
optimization techniques,’’ Int. J. Adv. Res. Artif. Intell., vol. 2, no. 7,
lems, the distribution of MOEA/D-NRSA is better than other pp. 39–50, 2013.
algorithms. We know that each solution represents a kind of [14] Y. Zhou, X. Li, Y. Zhu, and W. Wang, ‘‘A discrete particle swarm opti-
offensive plan in the objective space, and the better distribu- mization algorithm applied in constrained static weapon-target assignment
tion of the solution on the Pareto front is equivalent to obtain- problem,’’ in Proc. 12th World Congr. Intell. Control Automat. (WCICA),
Jun. 2016, pp. 3118–3123.
ing a more different plan. The advantage of MOEA/D-NRSA [15] B. Xin, J. Chen, Z. Peng, L. Dou, and J. Zhang, ‘‘An efficient rule-
in terms of distribution proves that more decision-makers can based constructive heuristic to solve dynamic weapon-target assignment
use it to obtain offensive schemes that meet their preferences. problem,’’ IEEE Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 41, no. 3,
pp. 598–606, May 2011.
Judging from the current research results, the dynamics, [16] B. Xin, J. Chen, J. Zhang, L. Dou, and Z. Peng, ‘‘Efficient decision makings
uncertainty, and coordination of the actual combat process for dynamic weapon-target assignment by virtual permutation and tabu
have not been well described in the model. The multi-stage search heuristics,’’ IEEE Trans. Syst., Man, Cybern. C, Appl. Rev., vol. 40,
no. 6, pp. 649–662, Nov. 2010.
model designed in this paper is only a product between SWTA [17] C. Leboucher, H.-S. Shin, P. Siarry, R. Chelouah, S. L. Ménec, and
and DWTA, and cannot fully describe the dynamics of con- A. Tsourdos, ‘‘A two-step optimisation method for dynamic weapon target
frontation. Regarding the uncertainty in the battlefield, this assignment problem,’’ Recent Advances on Meta-Heuristics and Their
Application to Real Scenarios. Rijeka, Croatia: InTech, 2013, pp. 109–129.
paper only caters to this feature by randomly generating some [18] A. Zhou, B.-Y. Qu, H. Li, S.-Z. Zhao, P. N. Suganthan, and Q. Zhang,
parameter values in a specific interval, rather than describing ‘‘Multiobjective evolutionary algorithms: A survey of the state of the art,’’
and processing uncertain factors through models. The limita- Swarm Evol. Comput., vol. 1, no. 1, pp. 32–49, Mar. 2011.
[19] C. A. C. Coello, G. B. Lamont, and D. A. Van Veldhuizen, Evolutionary
tions mentioned above in this paper lead to the direction of our
Algorithms for Solving Multi-Objective Problems, vol. 5. Boston, MA,
future research. In terms of problem models, we can change USA: Springer, 2007.
our thinking and design models with individuals rather than [20] B. Xin, Y. Wang, and J. Chen, ‘‘An efficient marginal-return-based
timelines. The timeline can be used as a standard for sorting constructive heuristic to solve the sensor–weapon–target assignment
problem,’’ IEEE Trans. Syst., Man, Cybern., Syst., vol. 49, no. 12,
out the logic of actions. This modeling method can control the pp. 2536–2547, Dec. 2019.
actions of each individual and simulate the dynamics of the [21] J. D. Schaffer, ‘‘Multiple objective optimization with vector evaluated
entire combat area. In dealing with uncertain factors, it is nec- genetic algorithms,’’ in Proc. 1st Int. Conf. Genet. Algorithms Appl.
Hillsdale, NJ, USA: Lawrence Erlbaum, 1985, pp. 93–100.
essary to find a quantifiable parameter to describe them and [22] Q. Zhang and H. Li, ‘‘MOEA/D: A multiobjective evolutionary algorithm
design a treatment method for uncertain factors. Finally, it is based on decomposition,’’ IEEE Trans. Evol. Comput., vol. 11, no. 6,
possible to study the specific collaborative actions between pp. 712–731, Dec. 2007.
[23] L. Ke, Q. Zhang, and R. Battiti, ‘‘MOEA/D-ACO: A multiobjective evo-
individuals in terms of collaboration instead of describing lutionary algorithm using decomposition and AntColony,’’ IEEE Trans.
the collaboration through simple numerical stacking or some Cybern., vol. 43, no. 6, pp. 1845–1859, Dec. 2013.
connection relationships between individuals. [24] J. Chen, J. Li, and B. Xin, ‘‘DMOEA-εC: Decomposition-based multi-
objective evolutionary algorithm with the ε-constraint framework,’’ IEEE
Trans. Evol. Comput., vol. 21, no. 5, pp. 714–730, Feb. 2017.
REFERENCES
[1] R. K. Ahuja, A. Kumar, K. C. Jha, and J. B. Orlin, ‘‘Exact and heuristic [25] X. Cai, Z. Mei, Z. Fan, and Q. Zhang, ‘‘A constrained decomposition
algorithms for the weapon-target assignment problem,’’ Oper. Res., vol. 55, approach with grids for evolutionary multiobjective optimization,’’ IEEE
no. 6, pp. 1136–1146, Dec. 2007. Trans. Evol. Comput., vol. 22, no. 4, pp. 564–577, Aug. 2018.
[2] S. P. Lloyd and H. S. Witsenhausen, ‘‘Weapons allocation is NP-complete,’’ [26] Y. Li, Y. Kou, and Z. Li, ‘‘An improved nondominated sorting genetic
in Proc. Summer Comput. Simul. Conf., 1986, pp. 1054–1058. algorithm III method for solving multiobjective weapon-target assignment
[3] A. Kline, D. Ahner, and R. Hill, ‘‘The weapon-target assignment problem,’’ part I: The value of fighter combat,’’ Int. J. Aerosp. Eng., vol. 2018,
Comput. Oper. Res., vol. 105, pp. 226–236, May 2019. pp. 1–23, Jun. 2018.
[4] A. S. Manne, ‘‘A target-assignment problem,’’ Oper. Res., vol. 6, no. 3, [27] S. Ding, C. Chen, Q. Zhang, B. Xin, and P. M. Pardalos, Metaheuristics
pp. 346–351, 1957. for Resource Deployment under Uncertainty in Complex Systems, 1st ed.
[5] Z.-J. Lee, C.-Y. Lee, and S.-F. Su, ‘‘An immunity-based ant colony opti- Boca Raton FL, USA: CRC Press, 2021.
mization algorithm for solving weapon–target assignment problem,’’ Appl. [28] K. Miettinen, Nonlinear Multiobjective Optimization, vol. 12. Boston, MA,
Soft Comput., vol. 2, no. 1, pp. 39–47, Aug. 2002. USA: Springer, 1998,
[6] O. Karasakal, ‘‘Air defense missile-target allocation models for a naval task [29] E. Zitzler and L. Thiele, ‘‘Multiobjective evolutionary algorithms: A com-
group,’’ Comput. Oper. Res., vol. 35, no. 6, pp. 1759–1770, Jun. 2008. parative case study and the strength Pareto approach,’’ IEEE Trans. Evol.
[7] A. G. Kline, D. K. Ahner, and B. J. Lunday, ‘‘Real-time heuristic algo- Comput., vol. 3, no. 4, pp. 257–271, Nov. 1999.
rithms for the static weapon target assignment problem,’’ J. Heuristics, [30] D. E. Goldberg, J. Richardson, and Others, ‘‘Genetic algorithms with
vol. 25, no. 3, pp. 377–397, Jun. 2019. sharing for multimodal function optimization,’’ in Genetic Algorithms and
[8] S. Zhihua, F. Zhu, and Z. Duolin, ‘‘A heuristic genetic algorithm for solving Their Applications Proceedings of the Second International Conference
constrained weapon-target assignment problem,’’ in Proc. IEEE Int. Conf. on Genetic Algorithms. Hillsdale, NJ, USA: Lawrence Erlbaum, 1987,
Intell. Comput. Intell. Syst., vol. 1, Nov. 2009, pp. 336–341. pp. 41–49.
[9] Z. R. Bogdanowicz, A. Tolano, K. Patel, and N. P. Coleman, ‘‘Optimiza- [31] J. Horn, N. Nafpliotis, and D. E. Goldberg, ‘‘A Niched Pareto genetic
tion of weapon–target pairings based on kill probabilities,’’ IEEE Trans. algorithm for multiobjective optimization,’’ in Proc. 1st IEEE Conf. Evol.
Cybern., vol. 43, no. 6, pp. 1835–1844, Dec. 2013. Comput. IEEE World Congr. Comput. Intell., Jun. 1994, pp. 82–87.

VOLUME 9, 2021 71847

 X. Wu et al.: Modified MOEA/D Algorithm for Solving Bi-Objective MWTA Problem

[32] L. Juan, C. Jie, and X. Bin, ‘‘Efficiently solving multi-objective dynamic CHEN CHEN received the B.S. degree in automa-
weapon-target assignment problems by NSGA-II,’’ in Proc. 34th Chin. tion and the Ph.D. degree in control science
Control Conf. (CCC), Jul. 2015, pp. 2556–2561. and engineering from the Beijing Institute of
[33] Q. Zhang, H. Li, D. Maringer, and E. Tsang, ‘‘MOEA/D with NBI-style Technology, Beijing, China, in 2004 and 2009,
Tchebycheff approach for portfolio management,’’ in Proc. IEEE Congr. respectively.
Evol. Comput., Jul. 2010, pp. 1–8. She is currently a Professor with the School of
[34] Z. Fan, Y. Fang, W. Li, X. Cai, C. Wei, and E. Goodman, ‘‘MOEA/D Automation, Beijing Institute of Technology. Her
with angle-based constrained dominance principle for constrained
current research interests include complicated sys-
multi-objective optimization problems,’’ Appl. Soft Comput., vol. 74,
tems, multi-objective optimization, and distributed
pp. 621–633, Jan. 2019.
[35] S. Jiang and S. Yang, ‘‘An improved multiobjective optimization evolution- simulation.
ary algorithm based on decomposition for complex Pareto fronts,’’ IEEE
Trans. Cybern., vol. 46, no. 2, pp. 421–437, Feb. 2016.
[36] W. Xu, C. Chen, S. Ding, and P. M. Pardalos, ‘‘A bi-objective dynamic
collaborative task assignment under uncertainty using modified MOEA/D
with heuristic initialization,’’ Expert Syst. Appl., vol. 140, Feb. 2020,
Art. no. 112844.
[37] S. Ding, C. Chen, B. Xin, and P. M. Pardalos, ‘‘A bi-objective load balanc-
ing model in a distributed simulation system using NSGA-II and MOPSO
approaches,’’ Appl. Soft Comput., vol. 63, pp. 249–267, Feb. 2018.
[38] Y. Qi, X. Ma, F. Liu, L. Jiao, J. Sun, and J. Wu, ‘‘MOEA/D with adaptive
weight adjustment,’’ Evol. Comput., vol. 22, no. 2, pp. 231–264, Jun. 2014.
[39] W. Khan and Q. Zhang, ‘‘MOEA/D-DRA with two crossover operators,’’
in Proc. U.K. Workshop Comput. Intell. (UKCI), Sep. 2010, pp. 1–6.
[40] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, ‘‘A fast and elitist
multiobjective genetic algorithm: NSGA-II,’’ IEEE Trans. Evol. Comput.,
vol. 6, no. 2, pp. 182–197, Apr. 2002.
[41] J. Li, B. Xin, P. M. Pardalos, and J. Chen, ‘‘Solving bi-objective uncertain
stochastic resource allocation problems by the CVaR-based risk measure
and decomposition-based multi-objective evolutionary algorithms,’’ Ann.
Oper. Res., vol. 296, nos. 1–2, pp. 639–666, Jan. 2021.
[42] S. Zapotecas-Martínez, B. Derbel, A. Liefooghe, D. Brockhoff,
H. E. Aguirre, and K. Tanaka, ‘‘Injecting CMA-ES into MOEA/D,’’
in Proc. Annu. Conf. Genetic Evol. Comput., Jul. 2015, pp. 783–790.
[43] M. Afzalirad and J. Rezaeian, ‘‘A realistic variant of bi-objective unre-
lated parallel machine scheduling problem: NSGA-II and MOACO
approaches,’’ Appl. Soft Comput., vol. 50, pp. 109–123, Jan. 2017.
[44] C. R. Raquel and P. C. Naval, ‘‘An effective use of crowding distance in
multiobjective particle swarm optimization,’’ in Proc. Conf. Genetic Evol.
Comput. (GECCO), Jun. 2005, pp. 257–264.
SHUXIN DING received the B.E. degree in
automation and the Ph.D. degree in control sci-
ence and engineering from the Beijing Institute
of Technology, Beijing, China, in 2012 and 2019,
XIAOCHEN WU received the B.S. degree in respectively.
automation and the M.Eng. degree in control engi- From 2016 to 2017, he was a Visiting Scholar of
neering from the Beijing Institute of Technology, industrial and systems engineering with the Uni-
Beijing, China, in 2013 and 2016, respectively, versity of Florida, Gainesville, FL, USA. He is
where he is currently pursuing the Ph.D. degree in currently an Assistant Researcher with the Signal
control science and engineering. and Communication Research Institute, China
His current research interests include multi- Academy of Railway Sciences Corporation Ltd. His current research inter-
route planning, damage assessment, dynamic ests include railway scheduling, evolutionary computation, multi-objective
self-organization of agents, and multi-objective optimization, and optimization under uncertainty.
optimization.

71848 VOLUME 9, 2021