SOLVING MULTI-WEAPON-TARGET ASSIGNMENT PROBLEM US!. ..

14th World Congress ofIFAC

Copyright ~I 1999 IFAC L-5a-04-6

14th Triennial World Congress~ Beijing, P.R. China

SOLVING MULTI.. WEAPON-TARGET ASSIGNMENT PROBLEM

USING A COMPOSITE OPTIMIZATION ALGORITHM

il
Shaoyong Hu to Yingping Zbeng ,..

Institute ofAutomation, Chinese Academy ofSciences,

P.O.Box,2728, Beijing, 100080, P.R. China
Tel: (86-10)82614438~ Fax: (86-10)62555383
* Email: [email protected] **Email: [email protected]

Abstract: The theories of two-state neural networks and genetic algorithms are
combined to form a new kind of composite stochastic optimization algorithm
with advantages of both. The large-scale multi-weapon-target assignment
(MWTA) problem is solved by using such a composite algorithm. Finally, An
example is showed. The simulation result demonstrates that the convergence
speed of the composite algorithm can meet the need of the real-time applications.
The result achieved by the composite algorithm is superior to both results
obtained via the two-state neural networks or genetic algorithms separately. It
will have a wide prospect if the composite algorithm is applied to the fire control
& command control system. Copyright@1999/FAC

Keywords: Resource allocation, multi-objective optimizations, neural networks,

genetic algorithms

1. INTRODUCTION significant component in the fire control &

command control system, so the research and
A proper and resolute decision will be made in a development of the decision methods and the
changeable battle only by the excellent commanders. algorithms becomes more and more important in the
These commanders not only possess rich experience, fITe control & command control system.
but also have a f"tre control & command control
system that can usually reflect the actual situations The MWTA problem is studied early by traditional
quickly and accurately. Reasonable decisions and algorithms (WaIler and Bruce, 1987) based on
right algorithms can provide the scientific basis for linear or nonlinear programming. Sometimes even
commanders' judgement in the fIre control & by the enumerative method. In recent years, the
command control system. Moreover, the quick neural networks is applied to solving MWT A
development of the computer techniques requires the problem (Wacholder, 1989; Hu and Song~ 1997). The
decision methods and the algorithms must be advantages and disadvantages of each algorithm were
improved in the fire control & command control discussed by Hu and Song (1997). For practical
systems. Whereas the large-scale multi-weapon- purpose, algorithms for solving nonlinear large-scale
target assignment (MWTA) problem is a MWTA NP-complete problem must satisfy both of
conditions simultaneously. (l) The convergence
speed of algoriThms must satisfy the real-time
This work is supported by the National Natural applications. (2) the results obtained by algorithms,
Sciences Foundation of China under grant No. if not the optimum solutions, should be the
69635030. satisfactory solutions at least.

5973
Copyright 1999 IFAC ISBN: 0 08 043248 4
SOLVING MULTI-WEAPON-TARGET ASSIGNMENT PROBLEM US!. .. 14th World Congress of IFAC

The theory of two-state neural networks is 1 N

investigated by Hopfield (1984) and Guo (1995). The p ... 1 ... p
large-scale MWTA problem was solved by this ~ ~ktP
theory (Hu and Song, 1997). The convergence speed k
was quite high and the assignment solution was also
satisfactory. Furthennore, the primary solution
obtained by the theory of two-state neural networks is
used as the initial value of genetic algorithms (GA),
the resulting solution win be superior to the one got
by v,vo-state neural netw"orks or GA separately. In yjli 1 ··~Yjlip .~.yjliN j

addition, with the rapid development of modem

computer techniques, the high-speed convergence k ~kil ..... Y;kip ..... ~kiNi
algoritlun by using the theory of two-state neural
networks will become quicker and quicker. That the
solution :1chieved by the theory of two-state neural
networks is used as the initial value of GA can
usually speed up the convergence of GA. And the
MWTA problem is a static assignment problem, M
\vhereas the speed that GA solves static problems is
high. So the speed of the composite algorithm that Fig. J. Decision Relation Figure of MWTA Problem
combines the two-state nemal networks and GA is Where, j=l, 2,. ~ M:t shows the j-th weapon
also high. The speed satisfies the real-time platfonn; k=l, 2, , Mj ~ shows the k.. th weapon in
applications in a battle and the final solution is also the j ..th weapon platform; i=l, 2, ... , N, shows the i-
very good. The advantages of the composite th group targets; p=l, 2, ... , Ni , shows the p-th
algorithm are further presented through simulating a target in the i-th group targets.
practical example in this paper9

The decision relation and the efficiency index about possible assignment schemes in all. This is a
MWTA problem are outlined in Section 2. In Section prohibited amount for an application. Set
3 the algorithm which solve MWTA problem by the
theory of two-state neural netviorks is discussed. GA
H= L y== 1 M j ,by the defInition, it is evident that:
and its application to MWTA problem are described
in Section 4. The simulation results and analyses of
the proposed composite algorithm are depicted in
Section 5, and the concluding remarks are delineated
in Section 6.

2. THE DESCRIPTION OF MWTA PROBLEM In order to realize the algorithm concerning the
theory of two-state neural networks, supposing that
not more than Rip weapons are assigned to attack the
2.1. The decision relation ofMWTAproblem p-th target in the i-th group targets. So formula (2) is
turned into (3).
As showed in Fig. J, there are N-group targets in all.
Ni (i=1, 2, . ~.~ N) is the number of targets in the i-th
group targets. M is the number of weapon platfonns.
M j G=l~ 2, ... , M) is the number of weapons in the j-
th weapon platform. When the p-th target in the i-th 2.2. The efficiency index Dj AIWTA problem
group targets is attacked by the k-th weapon in the j-
th weapon platform, set Yjkip =1, otherwise Yjldp =0. Supposing U jkip is the killing probability that the k-th
So the assignment matrix of weapon-target is fonned weapon in the j-th weapon platfonn kills the p-th
with Y jkip . That is, Yjkip constitutes the decision matrix target in the i-th group targets (Each weapon can
of the large scale MWTA problem. There are attack each target independently in practi ce.)~ The
(L .t:= 1 M j) x ('L.F= 1 Ni) elements in the decision efficiency index of MWTA problem is showed in
formula (4).

matrix. Thus, there are (Lf:, 1 Ni )(L ~ 1 M j )

5974
Copyright 1999 IFAC ISBN: 0 08 043248 4
SOLVING MULTI-WEAPON-TARGET ASSIGNMENT PROBLEM US!. .. 14th World Congress of IFAC

N Nr
r :: 1 s :: 1 (g (Ujkip • U jkrs )) + w7
(r'=Fi,s*p)
M Mr
Where, Q shows the sum of the killing probability
Ujkip g (0, ~ L Yrsip ) (7)
that all the weapons engage all the targets~ The r=19=1
MWTA problem is to find the decision matrices to "*
(r *- j ,S k)
maximize Q.
Where, function g(x, y) is defined as:
3. SOLVING MWTA PROBLEM USING TWO-
STATE NEURAL NETWORKS l X~y
g (x,Y)=
{O others
In Fig. 1, Supposing X jkip is the inputing value of the
two-state neuron corresponding to the p-th target in W 1 , W 2 , W 3 , W 4 , W s ~ W 6 , W 7 are posItIve
the i-th group targets and the k-th weapon in the j-th numbers. In the fotIDula (7), the first term is an
weapon platform. Yjlcip is the outputing value of the effectively restrained one when the number of targets
tw'o-state neuron. The relationship between X jkip and which the j-th weapon platform attacks is more than
Yjkip is dermed as: the number Mj of weapons in the j-th weapon
platfonn; the second tenn is an effectively restrained
XoJ k lp
· >0 one when the number of weapons that attacks the
(5) same one target is more than Rip ; the third tenn is an
X jkip -< 0 effectively restrained one when one weapon attacks
two or more targets at the same time; the fourth
tenn shows ~jkip /6.t is proportionate to the killing
That Y jkip equals 1 shows that assigned weapon
probability U jkip when the outputing value Y jkip of the
attacks the corresponding targets; That YjJrip equals 0
corresponding judgement neuron equals 0; the fifth
shows that no assigned weapon attacks the
tenn shows 6.Xjkip l.6.t will increase when Ujkip is not
corresponding targets. The constrained formula of
less than Ursip., otheIVIise, L\Xjkip fAt is not influenced
X jkip is showed as follows:
by this term; the six term is a reinforcement one when
a certain weapon chooses one target which the
(6)
weapon has the maximum probability to attack; the
seventh term is a reinforcement one when the number
Where, X jkip (t+8.t): the inputing value at the time
of all the weapons is more than that of all the targets,
(t+~t). Xjkip(t): the inputing value at the time t.
and any target must be attacked by one weapon at
~Xjkjp : the changing value from time t to time (t+~t). least.
Set .6t= 1 in the simulation process. AXjkip is derIDed
by the fonnula (7). The simulation of an example using the theory of
two-state neural networks is discussed by Ru and
AV M. N Song (1997), and the simulation results for
UJ'). ·k~ J N r
~ lp = -w] g (L L L Yj1rs ' M j ) comparison are presented at section 5 in this paper.
t l=lr=ls=l
(I ~ k ,r i ,S "* p) "*
M Mr 4. GA AND ITS APPLICATION TO MWTA
PROBLEM
-W2 g (L L Yrsip ' H) -
r = ls:=:)
(r =# j, S *" k) A main practical field of GA is the optimization
problem (Goldberg, 1989; Eberhart and Simpson,
N Nr
1996). Especially some overall optimum solution ofa
w3 g (L :2:. Yjkrs ,2) complicated NP pToblem may be solved conveniently
r = 1s == 1 by GA. Two problems must be usually settled before
(r =F i ,s -=F p)
GA is used. (1) Coding individuals, (2) Detennining
M Mr the objective functions. The MWTA probJem has
+W4UjJdp(1 - Y jkip)+ W s L L. (g (Ujkip , some constrained conditions. The individual codes,
r=ls=l the individual mutation, the individual reproduction,
(r*j,s*k) and crossover benveen individuals must comply with
the constrained conditions in order to make GA
converge to a satisfactory solution, or an optimum
solution.

5975
Copyright 1999 IFAC ISBN: 0 08 043248 4
SOLVING MULTI-WEAPON-TARGET ASSIGNMENT PROBLEM US!. .. 14th World Congress of IFAC

The individual codes of the proposed GA based on

the constrained conditions will be presented. Showed
in Fig.2~ Assuming each individual has
(1: J1= I M j ) chromosomes (Each chromosome
Chromosome 1 ... Chromosome (L:, M j
)

Fig. 2. Individual Code for MWTA Problem

expresses a weapon.). And each chromosome has
(L 1': I Ni) -bit 0 or 1. So each individual has the other initial individuals is at least more than
that of all the other initial individuals.
(Lf= I Ni) X(L~ I MJ -bit 0 or 1. The number
of 1 in each chromosome is not more than one,. which Stopping conditions of the program of the composite
satisfies the constrained condition of the formula (1). algorithm:
It is obvious that the above things satisfies the (1) When the average value Q of the objective
constrained condition of the formula (2). The fonnula function has rarely been changed and the efficiency
(4) is used as the objective function Q. According to index of MWTA has not been improved.
the Fig.2~ the bit 1 or 0 indicates the assignment (2) The recursive times of the program is more than
relation of MWTA problem. the maximum generation number. In the end, the
individual that has the maximum Q is selected as the
In order to ensure not to violate all the constrained fmal assignment solution.
conditions, the reproduction of GA is the same as that
of the traditional GA in this paper; during the
crossover~ the corresponding chromosomes of two 5. THE SIMULATION RESULTS AND
individuals are exchanged; during the mutation, if a ANALYSES OF THE COMPOSITE ALGORITHM
certain bit of a chromosome varies from 0 to 1, set
any other bit of the chromosome be 0, if a certain bit The example in this paper is the same as the one in
of a chromosome varies from I to 0, set any other bit the reference (Rn and Song, 1997).
of the chromosome be 1, other bits are all O.
For example: Supposing a naval fleet has four vessels
In order to ensure not to damage the optimum which possess two, three, four, and five weapons
individual, the individual possessing the maxinum Q, respectively, that is~ M=4, M 1 =2, M 2 =3, M 3 =4, M 4
in every generation of population, should be reserved. =5. So there are fourteen weapons in the fleet. Each
At the same time, it must participate in the genetic weapon can attack a target independently at the same
operations. time. There are five-group targets which possess
three) three, four, four, four targets respectively, that
in order to use the solution obtained by the theory of is, N=5, N J =N2 =3,. N 3 =N4 =Ns =4. There are
two-state neural networks as the initial value of an eighteen targets in alL The probability that each
individual, the initial value of a certain individual weapon kills targets is given in Fig.3
(supposing individual 1) is set to be the solution in
the population. At fITst we must retain individual I, The number of individuals is 120 in the population.
then some individuals are got through mutating The ratio of crossover is 0.70. The ratio of mutation
individual 1. Generally speaking, the number of is 0.01. The maximum generation number is 30. The
individuals achieved through mutating individual 1 solution achieved by the theory of two-state neural
shouldn~t be more than 15% of all the individuals. networks is used as the initial value of individual 1.
The other initial individuals in the population are So the coding configuration, (Vu, 1995), of
achieved by the heuristic method (Eberhart and individual 1 has fourteen chromosomes and each
Simpson, 1996; Wang and Fang, 1991). If the chromosome has eighteen-bit I or o~ Showed in Fig.
number of bits in each individual codes, 4.
(L.t: 1NJ x CL. ~ 1 M j ), is not more than that Fifteen initial individuals are achieved through
of all the other initial individuals, it is attempted mutating individual 1. The other 104 initial
to ensure the Boolean add of the corresponding individuals are achieved by the heuristic method. In
bits of all the other intial individuals must be one; the end, the results of the composite algorithm are
showed in Fig.5~ The efficiency index Q3 , solved by
if (:E f'= 1 Ni) x ( :E1= 1 M j ) is more than the the composite algorithm, equals 11.78. The efficiency
number of all the other initial individuals, it is index Q2, solved only by GA, equals 11.53. The
attempted to ensure the number of bits 1 got via efficiency index Ql , solved only by the theory of
the Boolean add of the corresponding bits of all two-state neural networks, equals 11.30. So as far as
the efficiency index is concerned,. the composite

5976
Copyright 1999 IFAC ISBN: 0 08 043248 4
SOLVING MULTI-WEAPON-TARGET ASSIGNMENT PROBLEM US!. .. 14th World Congress ofIFAC

~
1 2 3 4 5

j k 1 2 3 1 2 3 1 2 3 4 1 2 3 4 1 2 3 4

1 1 0.10 0.15 0.20 0.560.320.10 0.00 0.80 0.50 0.40 0.55 0.20 0.03 0.67 0.100.200.100.40
2 0.20 0.75 0.45 0.56 0.00 0.20 0.74 0.25 0.20 0.00 0.30 0.20 0.42 0.23 0.09 0.30 0.05 0.67
1 0.40 0.86 0.40 0.00 0.40 0.45 0.65 0.400.750.23 0.40 0.30 0.40 0.04 0.40 0.23 0.21 0.34
2 2 0.70 0.80 0.70 0.060.540.15 0.10 0.20 0.35 0.80 0.60 0.40 0.30 0.72 0.37 0.67 0.79 0.54
3 0.90 0.00 0.65 0.40 0.90 0.95 0.08 0.45 0.87 0.32 0.75 0.00 0.00 0.10 0.50 0.45 0.65 0.65
1 0.10 0.00 0.20 0.00 0.20 0.00 0.00 0.60 0.00 0.50 0.00 0.00 0.90 0.00 0.40 0.54 0.34 0.23
3 2 0.80 0.35 0.10 0.50 0.70 0.90 0.85 0.30 0.50 0.80 0.50 0.40 0.30 0.20 0.10 0.76 0.45 0.90
3 0.700.45 0.10 0.40 0.60 0.70 0.43 0.50 0.65 0.30 0.56 0.70 0.55 0.00 0.00 0.35 0.45 0.92
4 0.60 0.50 0.65 0.50 0.20 0.30 0.50 0.60 0.70 0.60 0.43 0.60 0.55 0.60 0.06 0.45 0.54 0.56
] 0.50 0.23 0.45 0.67 0.65 0.00 0.30 0.40 0.30 0.10 0.43 0.87 0.25 0.96 0.00 0.00 0.23 0.45
2 0.20 0.12 0.34 0.87 0.10 0.50 0.45 0.500.100.45 0.54 0.67 0.35 0.00 0.56 0.12 0.34 0.65
4 3 0.34 0.43 0.54 0.95 0.08 0.40 0.67 0.70 0.70 0.54 0.65 0.90 0.45 0.30 0.23 0.08 0.45 0.76
4 0.78 0.56 0.76 0.45 0.65 0.50 0.79 0.80 0.56 0.65 0.76 0.45 0.65 0.63 0.45 0.10 0.23 0.78
5 0.65 0.74 0.34 0.63 0.05 0.04 0.40 0.40 0.34 0.76 0.34 0.34 0.45 0.20 0.31 0.34 0.76 0.34

Fig.3. The Probability that the fourteen weapons attack the eighteen targets

o ... Y1132=1 •.. 0 o ... Y1231=1 ... 0 o ... Y21t2=1 .. ,0 o ... Y2213=1 '" 0 o ... Y2322=1 ••• 0

chromosome I chromosome 2 chromosome 3 chromosome 4 chromosome 5

o ... y 3143 = 1 ... 0 o .,. y 3223= 1 ... 0 o ... Y3342=1 ... 0 o ... Y3433=1 •.. 0 o ... Y4144=1 ... 0

chromosome 6 chromosome 7 chromosome 8 chromosome 9 chromosome 10

o .-. Y4254=1 •.. 0 o ... Y4323=1 ..• 0 o ... Y4411=1 '" 0 o ... y 4534=1 ... 0 Q=11.30

chromosome 11 chromosome 12 chromosome 13 chromosome 14

Fig. 4. The coding configuration of the individual 1

~
1 2 3 4 5

1 2 3 1 2 3 1 2 3 4 1 2 3 4 1 2 3 4

It1
)2 CBA
CBA

1 CBA
2 2 A CB
3 CBA
1 CBA
3 2 eBA
3 A CB
4 CBA
1 CBA
2 CB A
4 ~
.:> A B C
4 CBA
5 CBA

Fig.5. The Results of the three algrithms

A: results of the theory of the two-state neural networks; B: results of GA; C: results of the composite

5977
Copyright 1999 IFAC ISBN: 0 08 043248 4
SOLVING MULTI-WEAPON-TARGET ASSIGNMENT PROBLEM US!. .. 14th World Congress of IFAC

algorjthm is superior to the other two algorithms. The Hopfield J.J. (1984). Neurons with Graded Response
convergence of the composite algorithm is also very Have Collec'dve Computational Properties Like
good because GA owns good convergence. In Those of Two-state Neurons. Proc Natl Acad Set,
addition, The convergence speed of the composite USA, 81~ 3088.
algorithm is a little slower than that of the theory of
two-state neural networks, but it still satisfies the real- R.Eberhart and P. Simpson (1996).
time applications as usual. Computational Intelligence PC Tools. AP
Professional: A Division of Academic Press
In actual battles, if the time is very limited, the Inc ..
solution solved only by the theory of two-state neural
networks may be used as the final result of MWTA; if Shaoyong, Hu and Fuxiang, Song (1997). The Theory
the time is enough to run the prorgam of the of The Judgement Neuron and Its Application to
composite algorithm, of course, the solution of Weapon-target. Fire Control &Command
MWTA should be the one solved by the composite Control, 22 (4}, 21. China~
algorithm. The simulation time of the example is not
more than half a second in Pentium MMX 166. With Wacholder,E. (1989). A Neural Nenvorks-Based
the development of the concurrent computational Optimization Algorithm for the Static Weapon-
techniques, the convergence speed of two-state neural target Assignment Problem~ ORSA Joural on
netvlorks and GA will become higher and higher, the Computing, 1 (4), 232.
speed of the composite algorithm will become higher
and higher, too. So the practical prospect of the Waller,D.C. and Bruce,J.T. III (1987). The Strategic
composite algorithm will become wider and wider. Defence Initiative Progress and Challenges.
Regina Book, CA.

6.CONCLUSIONS Wen-Deng,Yu (1995). An Adaptive Analytic

Hierarchy Process for Determining the Fuzzy
Firstly, the decision relation figure concerning the Membership Values in Multi-Criterion
large-scale MWTA problem is established in this Decision Problem. Int. Joint Conf. of
paper. According to the decision relation figure, an CFSAjIFISISOFT~95 on Fuzzy Theory and
efficiency index is detennined. The index can Application, 308. Taipei.
evaluate whether or not the assignment schemes of
weapon-target are good. Then the algoritlnns are
presented which use the theories of the two-state
neural networks and GA to solve large-scale MWTA
problem. A new composite optimization algorithm is
achieved through synthesizing the two algorithms.
Finally, an example is simulated. The superiority of
the proposed composite algorithm is expressed via
comparing and analyzing the performances of the
three algorithms. It is safe to say that the composite
algorithm has its theoretical and practical value to a
certain degree.

REFERENCES

D.Goldberg (1989). Genetic Algorithm in Search,

Optimization} and Maching Learning.
Addision Wesley PUblishing, New York.

Dingwei,Wang and Shu-Cherng,Fang (1997). A

Genetics-based Approach for Aggregated
production Planning in a Fuzzy Environment.
IEEE Trans. on System, Man and Cybernetics,
27,636.

Guo L. (1995). The Logical Theory of Two-state

Neural Networks. Journal of Northwestern
University afTechnology, 13, 629. XPan, China.

5978
Copyright 1999 IFAC ISBN: 0 08 043248 4