www.scichina.

com
info.scichina.com
www.springer.com/scp
www.springerlink.com

Evolutionary decision-makings for the dynamic

weapon-target assignment problem
CHEN Jie1,2 , XIN Bin1,2† , PENG ZhiHong1,2 , DOU LiHua1,2 & ZHANG Juan1,2
1 School of Automation, Beijing Institute of Technology, Beijing 100081, China;
2 Key Laboratory of Complex System Intelligent Control and Decision, Ministry of Education, Beijing 100081, China

The dynamic weapon-target assignment (DWTA) problem is an important issue in the field of military
command and control. An asset-based DWTA optimization model was proposed with four kinds of con-
straints considered, including capability constraints, strategy constraints, resource constraints and
engagement feasibility constraints. A general “virtual” representation of decisions was presented to
facilitate the generation of feasible decisions. The representation is in essence the permutation of all
assignment pairs. A construction procedure converts the permutations into real feasible decisions.
In order to solve this problem, three evolutionary decision-making algorithms, including a genetic al-
gorithm and two memetic algorithms, were developed. Experimental results show that the memetic
algorithm based on greedy local search can generate obviously better DWTA decisions, especially for
large-scale problems, than the genetic algorithm and the memetic algorithm based on steepest local
search.

decision-making, dynamic weapon-target assignment (DWTA), military command and control, evolutionary computation, memetic
algorithms, constraints handling

1 Introduction be NP-Complete[2] . It has two versions — static

WTA (SWTA) and dynamic WTA (DWTA)[3,4] .
The dynamic weapon-target assignment (DWTA) In SWTA, all weapons engage targets in a single
problem is a typical constrained combinatorial op- stage, and all of the parameters for the problem
timization problem arising in the eld of military are known. Thus, the goal of SWTA is to nd
operations research. Its objective is to minimize the optimal assignment for a temporary defense
the expected damage of own-force assets by as- task. In contrast, DWTA is a multi-stage problem
signing available weapons to oensive targets at and the outcome of each engagement is assessed
appropriate occasions. In fact, the weapon-target for subsequent decisions. The goal of DWTA is to
assignment (WTA) problem is one of the crucial nd a global optimal assignment for the whole de-
issues for the automation of military command fense process in which the engagement occasion of
and control (C2) which is regarded as a challenge weapons must be taken into account. Intuitively,
to control science[1] . WTA has been proved to DWTA can be achieved by a series of SWTAs
Received January 29, 2009; accepted June 6, 2009
doi: 10.1007/s11432-009-0190-x
†
Corresponding author (email: [email protected])
Supported by the National Natural Science Foundation of China (Grant No. 60374069), and the Foundation of the Key Laboratory of Complex
Systems and Intelligent Science, Institute of Automation, Chinese Academy of Sciences (Grant No. 20060104)

Citation: Chen J, Xin B, Peng Z H, et al. Evolutionary decision-makings for the dynamic weapon-target assignment problem. Sci China Ser
F-Inf Sci, 2009, 52(11): 2006–2018, doi: 10.1007/s11432-009-0190-x
through all stages. However, although SWTA can target-based SWTA problems. Besides, Ahuja et
at best guarantee the optimality of the WTA deci- al.[19] developed several branch-and-bound algo-
sions for its corresponding defense stage, the com- rithms and a very large scale neighborhood search
bination of all SWTA decisions may not be optimal algorithm to solve target-based SWTA problems.
for the whole defense process. Besides, the actual Although nearly two decades have passed since
issue of time windows which limit the engagement DWTA was put forward[4] , there are in contrast to
of weapons is not involved in SWTA[3] . In addition SWTA only a few researches on DWTA[5,6,21−24] .
to the engagement constraints, the complexity of Cai et al.[6] introduced some basic concepts on
DWTA problems is also caused by resource con- DWTA and provided a systematic survey on WTA
straints and strategy constraints[5,6] . problem. Hosein and Athans[4] did an early re-
Previous researches on WTA mainly focus on search on a two-stage asset-based DWTA problem
SWTA[6−20] . In respect of SWTA models, Ho- and proposed a suboptimal algorithm for nding a
sein and Athans[3] proposed an asset-based SWTA good solution. Hosein et al.[21,22] also made some
model which was also used in refs. [7, 8]. typical empirical experiments and provided ana-
Karasakal[9] used the probability of shooting down lytical solutions to several simple cases of DWTA.
all incoming targets as the objective function of Khosla[23] used a hybrid genetic algorithm which
the air defense WTA model for a naval task incorporates a simulated annealing-type heuristic
group. Some scholars adopted target-based SWTA to solve a target-based DWTA problem. In partic-
models which do not employ the value of assets ular, a weighted combination of threat value and
directly[10−19] . Instead, each target in this case is option weight is employed in the objective function
assumed to have certain value of threat, and the of this DWTA model. Havens[24] models DWTA by
objective is to minimize the total threat of all tar- means of simulation; however, it is in fact the rep-
gets. In fact, the probability model in ref. [9] and etition of SWTA[5] . Li et al.[5] proposed a target-
the target-based models are just special cases of the based DWTA model with the objective of mini-
asset-based model in ref. [3]. Besides, the cost of mizing the total threat of the targets which sur-
weapons is also taken into account in some models vive the nal stage of air defense operation. As
like that in ref. [20]. A more complicated model far as we know, no algorithm has been proposed
which considers the function of special assets can in the literature to solve asset-based DWTA prob-
be found in ref. [21]. lems with complicated constraints caused by the
Based on the above models, varied algorithms issue of engagement feasibility (e.g., the limit of
have been proposed to solve SWTA since the time windows). The goal of this paper is to de-
middle of the last century. In the early stages, velop evolutionary decision-making algorithms to
the algorithms for SWTA were limited to tradi- solve asset-based DWTA problems incorporating
tional algorithms such as implicit enumeration al- engagement feasibility.
gorithms, branch-and-bound algorithms and dy- The remainder is organized as follows. In section
namic programming[6] . With the development 2, the mathematical model for asset-based DWTA
of computer technology, some novel algorithms problem is formulated. A general “virtual” rep-
such as neural networks[10] , genetic algorithms resentation of solutions (decisions) is proposed to
(GAs)[7,11,12] , taboo search (TS)[13] , simulated an- facilitate the generation of feasible solutions, and a
nealing algorithms (SA)[14] , ant colony optimiza- constructive procedure transforms the virtual rep-
tion (ACO)[15] , and particle swarm optimization resentation into a feasible solution. In section 3,
(PSO)[16] , have been developed. Some scholars also a genetic algorithm and two memetic algorithms
tried hybrid algorithms[8,17,18] . For example, Lee (MAs) for DWTA that are hybrid of genetic al-
et al.[17] designed a memetic algorithm which com- gorithms (GAs) and local improvement techniques
bines the advantages of global search (genetic algo- are proposed. The results of employing dierent
rithm) and local search (greedy eugenics) to solve algorithms to solve DWTA problems are presented

CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018 2007
and analyzed in section 4. Finally, section 5 con- target j in stage t.
cludes the paper.
2.2 Constraints
The following four categories of constraints are in-
2 Problem formulation cluded in our DWTA model.
DWTA models depend on many factors such as de- T

fense strategies, features of targets and weapons xij (t)  ni , ∀t ∈ {1, 2, . . . , S},
and actual combat situations. Dierent defense j=1

scenarios may require dierent models. The sce- ∀i ∈ {1, 2, . . . , W }, (2)

nario considered in this paper is narrated as fol- W
lows. At certain time, the defender detects T oen-

xij (t)  mj , ∀t ∈ {1, 2, . . . , S},
sive targets with their attack aims exposed, and K i=1
assets of the defender are threatened. There are W ∀j ∈ {1, 2, . . . , T }, (3)
weapons available to intercept the targets. Before
S 
T
these targets break through the defense, there are 
at most S stages in which the defender’s weapons xij (t)  Ni , ∀i ∈ {1, 2, . . . , W }, (4)
t=1 j=1
can be assigned to certain targets. The value of S
xij (t)  fij (t), ∀t ∈ {1, 2, . . . , S},
depends on the distance between targets and their
aims, target’s ight speed, weapon’s regulation & ∀i ∈ {1, 2, . . . , W }, ∀j ∈ {1, 2, . . . , T }. (5)
launch & ight time, the delay of data analysis The constraint set (2) reects the capability of
and decision-making, and so on[9] . A general en- weapons in ring at multiple targets at the same
gagement policy “Shoot-Look-Shoot (SLS)” is as- time. Most of actual weapons can shoot only one
sumed which is a tradeo between defense eect target at a time. Besides, a special weapon that
and defense cost[4,9,21−23] . can engage multiple targets simultaneously can be
2.1 Objective function viewed as multiple separate weapons. In view of
these facts, we set ni = 1 for ∀i ∈ {1, 2, . . . , W }.
The total expected value of surviving assets after
The constraint set (3) limits the weapon cost for
the nal stage is taken as the objective of WTA
each target in each stage. The setting of mj (j =
for the current stage. This objective is somewhat
1, 2, . . . , T ) usually depends on the combat per-
similar to that in ref. [5] which also considers the
formance of available weapons. In our research,
optimization of the whole defense eect. The for-
we suppose that mj = 1 for ∀j ∈ {1, 2, . . . , T }.
mulation of the objective function for stage t is
This is a reasonable setting for missile-based de-
shown as follows:
fense systems and the “Shoot-Look-Shoot” engage-
Jt (X t ) ment policy[5,9] . For artillery-based defense sys-
K(t) T (t) S W (t)
 
    tems, the value of mj (j = 1, 2, . . . , T ) may be
= vk 1 − qjk (1 − pij (h))xij (h) greatly increased under the same demand on de-
k=1 j=1 h=t i=1
fense strength. So the constraints in (3) can be
with t ∈ {1, 2, . . . , S}, (1) considered as strategy constraints. The constraint
where W (t), T (t) and K(t) are, respectively, the set (4) reects in essence the amount of ammuni-
total numbers of the remaining weapons, targets tion equipped for weapons. Ni (i = 1, 2, . . . , W ) is
and assets in stage t; h is, like t, also an in- the maximal number of times that weapon i can be
dex of stage; X t =[Xt , Xt+1 , . . . , XS ] with Xt = used due to the limit of its equipped ammunition.
[xij (t)]W ×T is the decision variable in stage t, and In the constraint set (5), fij (t) is the indication of
xij (t) = 1 if weapon i is assigned to target j in stage actual engagement feasibility for weapon i assigned
t, xij (t) = 0 otherwise; vk is the value of asset k; to target j in stage t. fij (t) = 0 if weapon i cannot
qjk is the probability that target j destroys asset shoot target j in stage t with any potential reason;
k; pij (t) is the probability that weapon i destroys fij (t) = 1 otherwise. The time windows of targets

2008 CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018
and weapons are the primary factors that aect en- set (5). In order to avoid violating the con-
gagement feasibility[5,6] . Constraint set (5) is very straints in (5), we propose a virtual representa-
important to actual dynamic WTA problems since tion of DWTA decisions. The virtual representa-
it contains the inuence of time windows on the tion is the permutation of all available assignment
engagement feasibility of weapons. Besides, it also pairs. The denition of available assignment pairs
increases the complexity of DWTA problems and is shown as follows.
the diculty of generating feasible solutions. In Denition 1. Available assignment pair (AAP)
this case, it is hard to design a desirable operator If weapon i can be used to engage target j in
which can generate new solutions and guarantee stage t, then the assignment pair i − j − t is an
their feasibility at the same time. AAP. The pair i − j − t is an APP i fij (t) = 1.
2.3 Optimization model for DWTA Denote by SAAP the set of all AAPs. It is easy
to see that the engagement feasibility matrices
The optimization model for the dynamic WTA
F t = [fij (t)]W ×T (t = 1, 2, . . . , S) determine SAAP .
problem mentioned above is shown as follows:
So, F t (t = 1, 2, . . . , S) are utilized to get SAAP .
Maximize Jt (X t ) The following is an illustration for virtual repre-
s.t. (2), (3), (4) and (5), sentation.
with t ∈ {1, 2, . . . , S}. (6) Given
   
1 0 0 1
It is clear that the DWTA problem is a con-
F 1 = 1 0 , F 2 = 0 1
   
strained combinatorial optimization problem. In
the research eld of mathematical programming, it 0 1 1 0
can also be categorized as a 0-1 programming prob- and  
lem. The dynamic characteristic of this model is 0 0
mainly reected by the stochastic nature of damage F 3 = 1 0 ,
 
in each stage, the choice of engagement stages and 0 0
the change of engagement feasibility. In contrast to
target-based DWTA models[5,23] , this asset-based all AAPs with their indexes are shown as follows:
model stresses on the protection of own-force as- (1) i1 − j1 − t1; (2) i2 − j1 − t1; (3) i3 − j2 − t1;
sets especially those important assets, the ultimate (4) i1 − j2 − t2; (5) i2 − j2 − t2; (6) i3 − j1 − t2;
goal and essence of WTA, as shown by its objec- (7) i2 − j1 − t3.
tive. Besides, the threats of all targets are directly 3-2-7-4-1-5-6, for instance, is a permutation of
contained in this model while target-based models the seven AAPs that is a virtual representation of
depend on the evaluation of the threats of targets. certain DWTA decision. In fact, the permutation-
Note that we only consider the DWTA decision- based representation is common for the quadratic
making scenario corresponding to t = 1 for the assignment problem (QAP) which is a classical NP-
comparison of dierent algorithms in this paper. In hard combinatorial optimization problem[25] . How-
fact, decision-making will become relatively easier ever, the permutation employed for QAP is a di-
as t increases due to the destruction of targets and rect representation of solutions which indicates the
the consumption of weapons. There are no sub- correspondence between facilities and locations. In
stantial dierences between the DWTA decision- contrast, the permutation for DWTA does not pro-
making scenarios corresponding to dierent stages. duce feasible decisions directly since it cannot guar-
antee the satisfaction of the constraints in (2), (3)
2.4 Virtual representation and generation and (4). In fact, the foundation on which we em-
of feasible decisions ploy the virtual representation is that the genera-
The variable X t is a direct representation of tion of DWTA decisions depends on the assignment
DWTA decisions. However, it contains many im- order of all AAPs. Besides, it is convenient to pro-
mutable components restricted by the constraint duce feasible decisions with such a representation

CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018 2009
followed by a subsequent construction procedure. It should be noted that the increase of the use
In this procedure, we present the concept of con- of any eective weapon without the violation of
straint saturation to deal with the constraints in constraints will further improve the objective value
(2), (3) and (4). Note that a constraint c(X)  0 and thus lead to a better decision. Therefore, the
is said to be saturated i a decision X results in decision scheme in which all APPs are assigned will
c(X) = 0. be optimal if it does not cause any violation of con-
Initially, the states of all constraints in (2), straints. In general, however, the assignment of all
(3) and (4) are unsaturated. In the constructive APPs is infeasible due to the limitation of the con-
process, the states will be updated dynamically. straints in (2), (3) and (4). Although the above
The pseudocode for the construction procedure is construction procedure cannot ensure the optimal-
shown in Table 1. ity of generated decisions, it reduces the search
Since only AAPs are considered for the construc- scope of the DWTA problem greatly. This is be-
tion procedure, the constraints in (5) will not be vi- cause the generated decision with constraint satu-
olated. Thus, the generation of feasible decisions in rations is the best feasible one among all feasible
the above procedure concentrates on the constraint decisions generated by the same permutation with
satisfaction of (2), (3) and (4). It is evident that or without constraint saturations, and those infe-
no violation of constraints will occur due to the rior feasible decisions will not be produced.
mechanism of constraint saturation. Besides, each
permutation can be used to produce only one feasi-
3 Evolutionary decision-making algo-
ble decision because the construction procedure is
rithms
deterministic. All feasible decisions can be gener-
ated by certain permutations, but a feasible deci- In the eld of combinatorial optimization, it
sion may not correspond to a unique permutation. has been shown that combining evolutionary al-
In the following text on the design of evolution- gorithms, e.g., genetic algorithms (GAs), with
ary operators, we will elaborate some approaches problem-specic heuristics can lead to more e-
to avoid generating the same decision repeatedly. cient approaches[17,26] . These hybrid algorithms

Table 1 Pseudocode for the construction procedure

Procedure construction (Pe ) //Pe is a permutation of all AAPs (available assignment pairs; see Denition 1).

Initialize the saturation states of the constraints in (2), (3) and (4).

Let X = OW ×T ×S // a zero matrix

For k = 1 to size(SAAP )

If the assignment of the kth pair in the list L denoted by ik − jk − tk does not cause any violation of constraints

// L is the list formed by all AAPs in the order corresponding to the permutation Pe .

Let x(ik , jk , tk ) = 1

Update the saturation states of constraints related to the assigned AAP.

If all constraints in (2), (3) or (4) are saturated

Break //The procedure is over since any additional assignment will cause the violation of certain constraint.

End If

End If

End For

End Procedure

2010 CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018
are often termed memetic algorithms (MAs), hy- The utilization of local renement operations
brid genetic algorithms, Lamarckian GAs, etc.[27] . following initialization and all evolutionary opera-
In the framework of MAs, evolutionary algo- tions except for selection, i.e., crossover and muta-
rithms are expected to perform the global search tion, is the primary dierence between MAs and
since they are usually competent for the global ex- their GA counterparts. So, the development of
ploration of problem space. In contrast, problem- MAs mainly includes the design of local renement
specic heuristics are responsible for local improve- approaches and that of GAs. In the following, we
ment. Thus, the synergy of global search and lo- will present the content of the GA for DWTA and
cal renement, as a primary eect of hybridiza- two local renement approaches. Then, the resul-
tion, contributes to the superiority of MAs. It has tant MAs can be constructed easily according to
been approved that MA is one of the most pop- the above template for MA procedures. So, the de-
ular and eective algorithms to solve combinato- tails for MAs, for concision, will not be presented
rial optimization problems like QAP[25,26] . So, it is later. We will dierentiate the MAs proposed in
rational to adopt MAs to solve the DWTA prob- this paper by denominations later.
lem since it is also, as narrated above, a typical 3.1 Local renement approaches for
combinatorial optimization problem. In addition, DWTA
we also employ genetic algorithms for performance
comparisons. Typically, the common procedure of The goal of local renement is to improve the
MAs developed on the framework of GAs can be current solution locally and even nd a local op-
described as follows: timum. Heuristic algorithms are often used to
achieve local renement, and there are mainly
Step 1. Initialize population (denote by P op
three kinds of heuristics — constructive methods,
the initial population).
enumeration or limited enumeration, and improve-
Step 2. Apply certain local renement ment methods[25] . Constructive methods produce
method to improve the initial population: new solutions directly according to some predened
P op = Local Ref inement(P op). rules which usually contain some problem-specic
knowledge. Enumeration can guarantee the opti-
Step 3. Apply crossover operations on mality of generated solutions within certain neigh-
the population and use certain local renement borhood. However, it will generally take a long
method to improve the solutions generated by time to guarantee optimality when the search space
crossover. Add the improved solutions into the is very large. In this case, a limited enumera-
population. tion will be employed instead. The limited enu-

P op = P op Local Ref inement(Crossover(P op)). meration can reduce the execution time greatly;
howbeit, it loses the guarantee of optimality. Im-
Step 4. Apply mutation operations on provement methods correspond to local search al-
the population and use certain local renement gorithms which are also termed as neighborhood
method to improve the solutions generated by mu- search. An improvement method begins with a
tation. Add the improved solutions into the popu- feasible solution and tries to improve it, search-
lation. ing for other solutions in its neighborhood. In
the context of MAs, improvement methods are fre-

P op = P op Local Ref inement(M utate(P op)).
quently applied. A very eective local search oper-
Step 5. Apply selection operations on the ator for many combinatorial optimization problems
population to eliminate some bad solutions: like QAP is the 2-opt operator which exchanges two
P op = Select(P op). selected elements in a permutation[26] . In our re-
search, two local search methods based on a variant
Step 6. If termination criteria are satised, of the 2-opt heuristic are employed respectively for
stop the algorithm; otherwise, go to step 3. local renement.

CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018 2011
 As a rst step at introducing the proposed local measured by the product of 1) the value of the as-
renement approaches, an illustrative example is set threatened by target j, 2) the probability that
shown in Figure 1 to describe the rationale of the target j destroys the asset, and 3) the probabil-
basic operations in these approaches. As shown in ity that weapon i destroys target j. For example,
Figure 1, weapon i1 is assigned, as a result of the the AAP8 in Figure 1 can exchange with AAP5 to
assignment of AAP5, to target j1 in stage t2 . Thus, produce a new decision. The neighborhood of the
AAP8, if assigned, will make weapon i1 assigned to selective 2-opt operation is dened as the set of all
two targets in the same stage and cause the viola- permutations that can be formed by the current
tion of a constraint in (2). So it was discarded in permutation using this operation. The operation
the process of construction procedure. All assigned is implemented iteratively according to two dier-
AAPs, abbr. AAAPs, form a feasible DWTA deci- ent local search strategies as described below.
sion. 1) Steepest strategy[27] : Find the best decision
Any 2-opt operation within the set of AAAPs in the neighborhood of the current decision, and
will not produce a new decision since any two replace the current decision with it. A new local
AAAPs have no conicts against each other in re- search proceeds in the neighborhood of the new de-
spect of constraint satisfaction. In fact, only when cision.
unassigned AAPs, abbr. UAAPs, are exchanged 2) Greedy strategy[27] : Once a better decision is
with AAAPs, is a new decision likely to be gener- detected in the neighborhood of the current deci-
ated. However, the AAAPs for exchange have to sion, replace the current decision with it. A new
be carefully chosen. For the purpose of ensuring local search proceeds in the neighborhood of the
the birth of new decisions, a UAAP should be ex- new decision.
changed with the AAAPs which collide with the The procedures of the two local search methods
UAAP regarding some constraints. These AAAPs based on the above operator are shown in Table 2
can be regarded as the competitors of the UAAP and Table 3, respectively.
to become the component of a feasible decision. In contrast to the greedy local search procedure,
All AAPs are categorized into two sets — SAAAP the input variables of the steepest local search pro-
and SUAAP which, respectively, denote the set of cedure, i.e., Pe and X, are not changed during its
AAAPs and that of UAAPs. Each 2-opt operation implementation. It is easy to see that all solutions
will select a UAAP from SUAAP at rst. Then, a generated by the above two local search approaches
competitive AAAP of the UAAP will be chosen for are local optima. However, in our research, the
exchange. Further, it is required that the value of neighborhood of a permutation will not be com-
the AAAP is lower than that of the UAAP, which is pletely searched since it is usually very large. In
another means to curtail computation time. Note order to reduce the complexity of local search ap-
that the value of an AAP denoted by i − j − t is proaches and make a better tradeo between global

Figure 1 An illustration for the rational of basic operations.

2012 CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018
Table 2 Pseudocode for the greedy local search procedure

Procedure Greedy local search(P e , X)

//The input parameter Pe is a permutation of all AAPs, and X is the decision w.r.t. Pe .

//Denote by N (Pe ) the neighborhood of Pe .

Repeat
′
Implement a selective 2-opt operation on Pe to get a new permutation Pe .
′ ′
Apply the construction procedure to get the new decision X w.r.t. Pe .
′ ′
If J(X ) > J(X) Then X = X , P e =P ′e End If
′
Until ∀ P ′e ∈ N (P e ): J(X )  J(X)

Return Pe and X

End Procedure

Table 3 Pseudocode for the steepest local search procedure

Procedure Steepest local search(P e ,X)

//Record the currently best decision and its corresponding permutation: X∗ = X, Pe∗ =Pe

Repeat
′
Implement a selective 2-opt operation on Pe to get a new permutation Pe .
′ ′
Apply the construction procedure to get a new decision X w.r.t. Pe .
′ ′ ′
If J(X ) > J(X ∗ ) Then X ∗ = X , Pe∗ =Pe End If
′ ′
Until ∀ Pe ∈ N (Pe ): J(X )  J(X∗ )

Return X ∗ and Pe∗

End Procedure

exploration and local renement[28] , we employ 1) Initialization. The initial population is gen-
truncated local searches. In the implementation erated randomly without any heuristic informa-
of the selective 2-opt operations above, all UAAPs tion. This is mainly because eective constructive
are sorted according to their values, and only 10% heuristics for DWTA may have a high complexity,
of all UAAPs with larger values will be considered and random starting solutions in combination with
for exchange. local renement approaches, as shown above in the
We call the memetic algorithms corresponding common procedure of MAs, can produce relatively
to the greedy local search (GLS) and the steepest good solutions.
local search (SLS), MA-GLS and MA-SLS, respec- 2) Evolutionary operator 1 — Mutation. The
tively. function of mutation in GAs is to randomly gener-
3.2 A genetic algorithm for DWTA ate a new solution from an individual in the current
population. It is regarded as an important mech-
The design of genetic algorithms mainly encom-
anism to keep population diversity. The mutation
passes the encoding of solutions (i.e., the repre-
operation for DWTA is described as follows:
sentation of solutions), the initialization method,
First, randomly select m UAAPs from SUAAP .
the design of evolutionary operators and the set-
The number of selected UAAPs, i.e., m, is
ting of parameters. Since a suitable representation
randomly selected from the set {1, 2, . . . , size
has been proposed above for DWTA solutions, the
(SUAAP )}. Then, move the selected UAAPs in seq-
following will focus on the latter three contents.

CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018 2013
 Figure 2 A mutation example.

uence to the head of the mutated permutation. An Freisleben[26] to set P S = 40, and the number of
illustrative example for the mutation is shown in crossovers and that of mutations per generation are
Figure 2. both set at 10.
3) Evolutionary operator 2 — Crossover. The
main idea behind crossover is the inheritance of
4 Computational experiments
“good” genes from parents. As often as it is,
two parent individuals (solutions) are randomly se- 4.1 Test-case generator
lected for each crossover operation to generate an
ospring (new solution). Besides, a rule of thumb Given W , T , K and S, the generator will pro-
for crossover is to preserve the alleles that are iden- vide the essential parameters for a DWTA in-
tical for the same genes in both parents[26] . That stance which include V = [vk ]1×K , Q = [qjk ]T ×K ,
is, it is checked whether each AAAP in the SAAAP P t = [pij (t)]W ×T , Ni (i = 1, 2, . . . , W ) and F t =
for one of the two parent permutations exists in the [fij (t)]W ×T (t = 1, 2, . . . , S). The instructions for
SAAAP for the other parent. The common AAAPs these parameters can be found in sections 2.1 and
in both parents will be inherited and placed at the 2.2.
head of their ospring. Particularly, the order of The values of assets, i.e., vk (k = 1, 2, . . . , K),
the common AAAPs in the ospring will be the are randomly generated in the interval (10, 100).
same as that in the parent whose tness (i.e., ob- Since the number of targets is not less than that
jective value) is better. The remaining AAPs for of threatened assets, that is T  K, we assume
the ospring will be arranged randomly in its per- that the aim of the kth target is the kth asset
mutation. (k = 1, 2, . . . , K) to ensure that each asset is
4) Evolutionary operator 3 — Selection. The threatened by at least one target. The aims of
widely applied selection strategy, (µ + λ)-selection, the remaining targets will be randomly selected
is adopted since it has been proved to have good from K assets. Any target has no threat to the
performance in many researches[17,26,27] . assets which are not its aims. The parameters
5) Parameter settings. For GAs, the primary qjk (j = 1, 2, . . . , T ; k = 1, 2, . . . , K) are randomly
parameters include population size (P S), cross- generated in the interval (0.6, 0.99). The param-
over probability (Pc ) and mutation probability eters pij (t)(i = 1, 2, . . . , W ; j = 1, 2, . . . , T ; t =
(Pm ). However, in the context of MAs[26,27] , 1, 2, . . . , S) are randomly generated in the interval
the latter two parameters are not necessary. In- (0.4, 0.9). The engagement feasibility parameters
stead, the number of crossover and mutation oper- F t (t = 1, 2, . . . , S) are also generated in a similar
ations per generation is adopted. Considering the way. However, more zeros will appear in F t cor-
computational complexity of the local search ap- responding to later stages, which accords with the
proaches above, we follow the idea from Merz and fact that fewer weapons can be used in later sta-

2014 CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018
ges. Note that fij (t) = 0 implies that weapon i 4.2 Experimental results
cannot engage target j in stage t. The generation The DWTA decision-making algorithms under test
of F t (t = 1, 2, . . . , S) is shown as follows: include the GA, MA-GLS, MA-SLS proposed in
ratio(t) = fL + (fH − fL )(t − 1)/(S − 1) section 3. The maximal number of function evalua-
tions is set as N F Emax = 5W T S. In other words,
for t = 1, 2, . . . , S; any algorithm will be terminated once N F Emax
fij (t) = [sign(rand − ratio(t)) + 1]/2 DWTA decisions have been analyzed. This setting
is benecial to observe the performance dierences
for i = 1, 2, . . . , W ; j = 1, 2, . . . , T ; t = 1, 2, . . . , S, of the algorithms under test. Each algorithm was
where ratio(t) is the probability that fij (t) is set run 30 times for each test case with the results sta-
to 0, fL and fH are predened constants with tistically analyzed. All algorithms were performed
0 < fL < fH < 1 which reect the lower and upper on a PC with Intel(R) 2.0 GHz CPU and 2.0 GB
bounds of ratio(t), respectively, rand represents a internal memory.
random number in (0,1), and the function sign() The statistical results are presented in Table 4.
is equal to 1 if its argument is positive, −1 other- In columns 2–4 of Table 4, the data in the rst
wise. In our research, fL and fH are set at 0.1 and and second rows for each case correspond to the
0.9, respectively. mean and standard deviation of nally discovered
The generation of Ni (i = 1, 2, . . . , W ) is cat- best objective values (DBOV) and that of running
egorized into three cases so as to cover dierent time, respectively. The best result in each case
defense scenarios. is highlighted in bold. The results of two-tailed
t-tests (signicance level: 0.05) for paired compar-
Case 1. One weapon, one shot. Ni = 1 for
isons between DWTA algorithms are also provided
i = 1, 2, . . . , W .
as shown in columns 5–7 of Table 4. A value of one
Case 2. No weapon can be used through all indicates that the rst algorithm for a paired com-
stages. parison performs with 95% certainty better than
Ni = ⌈S  rand/2⌉ for i = 1, 2, . . . , W ; the second algorithm. A value of −1 indicates that
Case 3. All weapons are available in all stages. the second algorithm outperforms the rst one in
Ni = S for i = 1, 2, . . . , W . the aspect of optimization result or computation
Case 1 represents a scenario of “insucient” am- time. A value of zero implies that the performances
munition since each weapon has at most one “bul- are not statistically dierent. For example, in the
let”. In contrast, Case 3 corresponds to a scenario case labeled as “No. 1, Case 1”, MA-GLS (A2)
of “adequate” ammunition and each weapon can is superior to GA (A1) in respect of DBOV with
be used through all defense stages. Case 2 stands 95% certainty since the corresponding t-test result
for a moderate one between the two extremes rep- is one.
resented by Case 1 and Case 3. The three typical The convergence curves for the three algorithms
cases are considered for a comprehensive compar- in each case are shown in Figure 3. Note that these
ison of the DWTA algorithms under test. In our curves are the average of 30 runs. The three tested
simulation, the number of stages (S) is xed at algorithms perform comparatively when solving
4. The setting of the other three primary param- the two simple instances of DWTA problem cor-
eters W , T and K includes the following cases: responding to “No. 1, Case 2” and “No. 1, Case
W 10T 10K10 (No. 1), W 100T 50K50 (No. 2), 3”. The numbers of both weapons and targets
W 50T 100K50 (No. 3), W 100T 100K50 (No. 4). in the two instances are relatively small and the
For practical military decision-making, the latter equipped ammunition is sucient, especially for
three cases with four defense stages, i.e., No. 2, “No. 1, Case 3”, to guarantee a strong defense,
No. 3 and No. 4, represent DWTA scenarios of so it is easy for the three algorithms to nd a de-
large enough scale. sirable high-quality decision. As shown in Figure

CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018 2015
Table 4 Statistical results about dierent algorithms

Algorithms t-test
Case
A1 :GA A2 :MA-GLS A3 :MA-SLS A2 vs. A1 A2 vs. A3 A3 vs. A1

No. 1 337.0±1.4 345.2±2.1 343.6±2.5 1 0 1

Case 1 0.64±0.03(s) 0.68±0.03(s) 0.70±0.03(s) 0 0 −1

No. 1 382.7±1.2 383.1±2.8 382.9±1.5 0 0 0

Case 2 0.61±0.01(s) 0.65±0.02(s) 0.68±0.02(s) −1 0 −1

No. 1 389.3±1.0 391.3±2.2 389.5±2.3 0 0 0

Case 3 0.63±0.01(s) 0.68±0.02(s) 0.70±0.02(s) −1 0 −1

No. 2 1707.3±4.8 1743.7±5.2 1721.8±6.7 1 1 1

Case 1 118.2±0.2(s) 120.3±0.4(s) 122.7±0.5(s) −1 1 −1

No. 2 1739.4±4.4 1753.5±5.1 1744.2±6.3 1 1 0

Case 2 116.3±0.1(s) 118.1±0.3(s) 121.8±0.3(s) −1 1 −1

No. 2 1751.8±4.2 1777.2±4.9 1753.3±6.3 1 1 0

Case 3 112.7±0.1(s) 117.6±0.3(s) 121.3±0.3(s) −1 1 −1

No. 3 1441.6±6.2 1522.1±6.3 1467.5±7.2 1 1 1

Case 1 124.4±0.2(s) 128.2±0.4(s) 129.5±0.5(s) −1 1 −1

No. 3 1463.2±5.5 1541.9±6.1 1491.2±6.9 1 1 1

Case 2 120.1±0.1(s) 124.3±0.2(s) 126.9±0.3(s) −1 1 −1

No. 3 1471.8±5.3 1548.3±5.9 1527.1±6.7 1 1 1

Case 3 121.3±0.1(s) 123.7±0.2(s) 125.2±0.4(s) −1 1 −1

No. 4 1492.3±4.8 1583.1±5.2 1547.4±5.8 1 1 1

Case 1 213.8±0.3(s) 220.6±0.4(s) 225.3±0.6(s) −1 1 −1

No. 4 1537.5±4.2 1615.3±5.0 1587.6±5.3 1 1 1

Case 2 210.2±0.2(s) 217.2±0.3(s) 221.7±0.4(s) −1 1 −1

No. 4 1547.2±4.1 1621.7±4.9 1591.3±5.3 1 1 1

Case 3 210.3±0.2(s) 218.5±0.3(s) 222.5±0.4(s) −1 1 −1

3, the performance dierences between the three SLS especially in the decision-making of large-scale
algorithms increase with the scale of DWTA prob- DWTA problems. It should be noted that MA-
lems. MA-GLS prominently outperforms GA and GLS and MA-SLS are a little more time-consuming
MA-SLS in solving large-scale DWTA instances than GA in all cases. However, the trivial advan-
from No. 2 through No. 3 to No. 4. MA-SLS also tage of GA in computation time, as shown in Table
has obvious advantages over GA in all large-scale 4, may be neglected in contrast to its optimization
cases except “No. 2, Case 3”. The results shown quality. Compared with GA, MA-SLS performs
in Table 4 also indicate that the algorithm MA- better in the aspect of optimization quality. The
GLS has obvious advantages over GA and MA- advantage of MA-GLS and MA-SLS over GA dem-

2016 CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018
Figure 3 Convergence curves for GA, MA-GLS and MA-SLS. (a) No. 1, Case 1; (b) No. 1, Case 2; (c) No. 1, Case 3; (d) No. 2,
Case 1; (e) No. 2, Case 2; (f) No. 2, Case 3; (g) No. 3, Case 1; (h) No. 3, Case 2; (i) No. 3, Case 3; (j) No. 4, Case 1; (k) No. 4, Case
2; (l) No. 4, Case 3.

onstrates that the cooperation of local search meth- 5 Conclusions

ods and global search implemented by GA can
In order to solve asset-based DWTA problems with
improve the quality of DWTA decision-making
complicated constraints, a general “virtual” rep-
greatly. The advantage of MA-GLS over MA-SLS
resentation of decisions which is in essence the
shows that the utilization of greedy local search, in
permutation of all assignment pairs was proposed
contrast to steepest local search, can achieve better
to facilitate the generation of feasible decisions.
tradeo between the exploration and exploitation
A construction procedure which operates on the
of DWTA decision space.
virtual representation can guarantee the satisfac-

CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018 2017
 tion of all constraints. Based on the construction ter DWTA decisions without the cost of overmuch
technique, three evolutionary decision-making al- extra computation time than genetic algorithm
gorithms including a genetic algorithm and two and the memetic algorithm based on steepest lo-
memetic algorithms were proposed to solve DWTA cal search.
problems. The proposed memetic algorithm based
on greedy local search can generate obviously bet- The authors would like to thank the anonymous reviewers for
their constructive comments on this paper

1 Athans M. Command and control (C2) theory: a challenge optimization algorithm for solving weapon-target assignment
to control science. IEEE Trans Automat Contr, 1987, 32(4): problem. Appl Soft Comput, 2002, 2: 39–47
286–293 16 Zeng X, Zhu Y, Nan L, et al. Solving weapon-target assign-
2 Lloyd S P, Witsenhausen H S. Weapon allocation is NP- ment problem using discrete particle swarm optimization. In:
complete. In: Proceedings of the IEEE Summer Computer Proceedings of the 6th World Congress on Intelligent Control
Simulation Conference, Nevada, USA, 1986. 1054–1058 & Automation, Dalian, China, 2006. 3562–3565
3 Hosein P A, Athans M. Preferential defense strategies. Part I: 17 Lee Z J, Su S F, Lee C Y. Eciently solving general weapon-
The static case. Report LIPS-P-2002. 1990 target assignment problem by genetic algorithms with greedy
4 Hosein P A, Athans M. Preferential defense strategies. Part eugenics. IEEE Trans Syst Man Cybern-B, 2003, 33(1): 113–
II: The dynamic case. Report LIPS-P-2003. 1990 121
5 Li J, Cong R, Xiong J. Dynamic WTA optimization model 18 Fu T, Liu Y, Chen J. Improved genetic & ant colony optimiza-
of air defense operation of warships’ formation. J Syst Eng tion for regional air defense WTA problem. In: Proceedings
Electr, 2006, 7(1): 126–131 of the 1st International Conference on Innovative Computing,
6 Cai H, Liu J, Chen Y, et al. Survey of the research on dynamic Information and Control (ICICIC’06), Dalian, China, 2006.
weapon-target assignment problem. J Syst Eng Electr, 2006, 226–229
17(3): 559–565 19 Ahuja R K, Kumar A, Jha K C, et al. Exact and heuristic
7 Malhotra A, Jain R K. Genetic algorithm for optimal weapon algorithms for the weapon-target assignment problem. Oper
allocation in multilayer defence scenario. Defence Sci J, 2001, Res, 2007, 55(6): 1136–1146
51(3): 285–293 20 Kwon O, Lee K, Kang D, et al. A branch-and-price algorithm
8 Bisht S. Hybrid genetic-simulated annealing algorithm for op- for a targeting problem. Naval Res Logist, 2007, 54: 732–741
timal weapon allocation in multilayer defence scenario. De- 21 Hosein P A, Walton J T, Athans M. Dynamic weapon-target
fence Sci J, 2004, 54(3): 395–405 assignment problems with vulnerable C2 nodes. Report LIDS-
9 Karasakal O. Air defense missile-target allocation models for P-1786. 1988
a naval task group. Comput Oper Res, 2008, 35: 1759–1770 22 Hosein P A, Athans M. Some analytical results for the dynamic
10 Wacholder E. A neural network-based optimization algorithm weapon-target allocation problem. Report LIDS-P-1944. 1990
for the static weapon-target assignment problem. ORSA J 23 Khosla D. Hybrid genetic approach for the dynamic weapon-
Comput, 1989, 1(4): 232–246 target allocation problem. In: Proceedings of SPIE, 2001,
11 Grant K E. Optimal resource allocation using genetic algo- 4396: 244–259
rithms. Naval Review, Naval Research Laboratory, Washing- 24 Havens M E. Dynamic allocation of res and sensors. Master
ton, DC, 1993. 174–175 Thesis, Naval Postgraduate School, OMB No. 0704-0188, 2002
12 Lu H, Zhang H, Zhang X, et al. An improved genetic algorithm 25 Loiola E M, Maia de Abreu N M, Boaventura Netto P O, et
for target assignment optimization of naval eet air defense. al. A survey for the quadratic assignment problem. Euro J
In: Proceedings of the 6th World Congress on Intelligent Con- Oper Res, 2007, 176: 657–690
trol & Automation, Dalian, China, 2006. 3401–3405 26 Merz P, Freisleben B. Fitness landscape analysis and memetic
13 Cullenbine A C. A taboo search approach to the weapon as- algorithms for the quadratic assignment problem. IEEE Trans
signment model. Master Thesis, Department of Operational Evol Comput, 2000, 4(4): 337–352
Sciences, Air Force Institute of Technology, 2000 27 Krasnogor N, Smith J. A tutorial for competent memetic al-
14 Li H R, Miao Y. WTA with the maximum kill probability gorithms: model, taxonomy, and design issues. IEEE Trans
based on simulated annealing algorithms. In: Proceedings of Evol Comput, 2005, 9(5): 474–488
the Conference on Special Committee of C2 and Computer 28 Chen J, Xin B, Peng Z H, et al. Optimal contraction theorem
of the Electronic Technology Academic Committee of China, for exploration-exploitation tradeo in search and optimiza-
Ship Engineering Society, 2000. 436–440 tion. IEEE Trans Syst Man Cybern-A, 2009, 39(3): 680–691
15 Lee Z J, Lee C Y, Su S F. An immunity-based ant colony

2018 CHEN J et al. Sci China Ser F-Inf Sci | Nov. 2009 | vol. 52 | no. 11 | 2006-2018