ARTICLE IN PRESS

Omega 38 (2010) 383392

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

Fire scheduling for planned artillery attack operations under time-dependent destruction probabilities
Young-Ho Cha, Yeong-Dae Kim
Department of Industrial Engineering, Korea Advanced Institute of Science and Technology, Yusong-gu, Daejon 305-701, Republic of Korea

a r t i c l e in fo
Article history: Received 2 July 2009 Accepted 10 October 2009 Processed by B. Lev Available online 31 October 2009 Keywords: Military Scheduling Branch and bound algorithm

abstract
In this paper, we consider the re scheduling problem (FSP) for eld artillery, which is the problem of scheduling operations of ring at given targets with a given set of weapons. We consider a situation in which the number of available weapons is smaller than the number of targets, the targets are assigned to the weapons already, and targets may move and hence the probability that a target is destroyed by a ring attack decreases as time passes. We present a branch and bound algorithm for the FSP with the objective of minimizing total threat of the targets, which is expressed as a function of the destruction probabilities of the targets. Results of computational tests show that the suggested algorithm solves problems of a medium size in a reasonable amount of computation time. & 2009 Elsevier Ltd. All rights reserved.

1. Introduction In this paper, we consider a scheduling problem arising in the military, especially in artillery operations. There are two types of artillery operations, planned artillery operations and responsive artillery operations. In a planned artillery attack operation, a set of targets, or enemy units, is to be red at and destroyed by weapons such as eld artillery units according to a predetermined operation plan. On the other hand, in responsive artillery attack operation, enemy units are red at (upon real-time requests of friendly units) in real-time combat situations. We consider the planned artillery attack operation. In this paper, a weapon represents an artillery battery, which is the basic unit to execute artillery attack operations and is typically composed of six howitzers, and a target represents enemy unit(s), i.e., armor unit(s), infantry unit(s), transportation unit(s), etc. In general, targets are identied by reconnaissance assets, such as human intelligence units, unmanned aerial vehicles and articial satellites, and it is assumed in this research that locations of the targets are known. The planned artillery attack operation consists of two steps of a sequential procedure: assignment of weapons to the targets; and scheduling ring operations against the targets that are assigned to each weapon. Note that when there are more targets than available weapons, one weapon may have to be assigned to more than one target. The decision problems of the rst step and the second step are called the weapon target assignment problem

Corresponding author. Tel.: + 82 42 350 3120; fax + 82 42 350 3110.

E-mail address: ydkim@kaist.ac.kr (Y.-D. Kim). 0305-0483/$ - see front matter & 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.omega.2009.10.003

(WTAP) as dened in Ahuja et al. [1] and the re scheduling problem (FSP), respectively. This paper focuses on the FSP, the problem of scheduling operations of ring at given targets with a given set of weapons for the objective of minimizing the threats from the enemies that survive the ring attack. In this problem, n targets are to be red at and destroyed by m weapons. We assume that ring operations are already assigned to the weapons. For a ring operation, a prespecied set of weapons should start ring at a target simultaneously. We consider a situation in which the number of weapons is smaller than the number of targets (m o n), and hence the sequence of targets to be red at should be determined. In such a situation, while a set of targets is under ring attack, other targets are usually alerted by a situation awareness system and these targets may move to avoid an anticipated incoming attack. Note that many types of targets, such as tank units, self-propelled artillery and combat support units, have mobility to hide themselves from attacks. This will decrease the probability that these targets are destroyed by the ring attack. In this paper, we assume the destruction probability decreases linearly as time passes. In the following, we briey review research results on the WTAP, FSP and other related problems. Reviews on earlier research articles on the WTAP are given in Matlin [2] and Eckler and Burr [3]. Manne [4] rst considers the WTAP and solves the problem after transforming the problem into a transportation problem, and Lloyd and Witsenhausen [5] show that the WTAP is NP-complete. Later, Hosein and Athans [6] consider a multi-stage or dynamic version of the WTAP, and present a method in which weapons are assigned to the targets in a period based on the assignment made in the previous period. In addition, Christ et al. [7] and Erdem and Ozdemirel [8] present a

ARTICLE IN PRESS
384 Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392

solution method based on neural networks and a genetic algorithm, respectively, while Lee and Lee [9] give a hybrid search algorithm composed of an ant colony optimization algorithm and a genetic algorithm. On the other hand, Ahuja et al. [1] propose exact and heuristic approaches to the WTAP. Unlike most of the research, in which weapons are to be allocated without consideration of the behavior of opponents or enemy troops, Malcolm [10] and Karasakal [11] present models in which the opponents behavior is taken into account by assuming friendly forces are equipped with a situation awareness system. The model for the WTAP is also used in the advertisement industry for media and budget allocation [12], and in the medicine area for cancer modeling [13]. Unlike the WTAP, research on the FSP is very rare. Introducing the FSP rst, Kwon et al. [14] consider the FSP with the objective of minimizing makespan under the assumption that the targets are xed (do not move) and the destruction probability is constant over time. Note that they focus on achieving a surprise attack effect through the quickest completion of the ring operation. However, their algorithm may not be applicable to cases in which the targets have mobility and are equipped with an alert system. In most machine scheduling problems considered in the literature, it is generally assumed that a job can be processed on one and only one machine [1517]. However, it is not the case with the problem considered in this study. The FSP is similar to the simultaneous resource scheduling problem [18], or the multiprocessor task scheduling problem (MPTSP) [19], in which each task requires one or more processors simultaneously. However, in the FSP, different weapons may require different ring durations, i.e., processing times for the ring operation, even for the same target, while an operation requires the same processing time on the machines in these scheduling problems. In other words, these scheduling problems are special cases of the FSP. Since it is proven that the MPTSP is NP-hard [19], the FSP is also NP-hard. As surveyed by Drozdowski [20], Lee et al. [21] and Allahverdi et al. [22], various solution methods have been developed for these scheduling problems [2328]. In this research, we consider the FSP with the objective of minimizing total threat of targets or opponents. The threat of a target is expressed as a function of the destruction probability, which depends on the weapon-target pair. It is assumed that the destruction probability decreases linearly as time passes. We develop several dominance properties and lower bounds for the FSP and present a branch and bound (B&B) algorithm using them. Due to the complexity of the problem and the requirement of prompt decisions in practice, we also suggest a heuristic algorithm that can give reasonably good solutions in a short time. The heuristic algorithm is also used in the B&B algorithm for an initial upper bound. In the next section, we describe the FSP considered in this study in more detail.

neously to achieve tactical surprise, and this type of ring is called the time on target ring. As stated earlier, a weapon represents an artillery battery which is equipped with howitzers. In general, the enemy under a ring attack cannot see the weapon, and hence immediate counter ring from the enemy against the weapon is almost impossible. In addition, the maximum ranges of the howitzers are between 11 and 30 km while the aggressive targets such as tanks have a shorter range of less than 10 km. Therefore, we assume that the threat from a direct counter ring attack from the enemy is negligibly small during the planning horizon, which is less than 1 h in our problem as in real planned artillery operations. In this study, a ring operation is specied by a pair of a weapon and a target associated with the operation and the duration for the operation. A set of ring operations against one target is called a job. A job for a single-weapon target is composed of only one ring operation, while a job for a multiple-weapon target is composed of two or more ring operations. Note that a job corresponds to a target and that an operation corresponds to a pair of a target and a weapon. We use the period as the unit of time lengths, that is, the ring duration for a ring operation is given as an integer multiple of the length of a period, as in real situations. By scheduling the ring operations in terms of the periods, one can synchronize ring operations of multiple weapons more easily. In typical situations, a period represents 1 min. Before a ring attack from a weapon is started against a target, a setup operation needs to be performed on the weapon for adjustment of azimuth, ring range, etc. In general, setup times (times required for setup operations) on different weapons may be different, but the setup times on the same weapon for ring operations against different targets are the same. Since the sequence of the ring operations are predetermined (through the solution of the FSP considered in this study) in planned artillery attack operations, the setup operations for the rst ring operation on each weapon is assumed to have been done by the commencement of the attack, i.e., by time 0 in the planning horizon. Therefore, in this study, we assume the setup time for the rst ring operation of each weapon is 0. The following are additional assumptions made in this study. (1) The locations and types of the targets are known, and threats from the targets may be different for different targets. (2) The targets are already allocated to the weapons. (3) The duration of the ring operation for each weapon-target pair as well as the setup time on each weapon is given, and they are integer multiples of the length of a period. (4) Each weapon can re at no more than one target in a single period. (5) Preemption of a ring operation is not allowed. (6) The destruction probability associated with a ring attack decreases linearly as time passes after the rst ring operation is started (since the targets, which are movable and equipped with a situation awareness system, hide themselves from anticipated attack). (7) Counter ring attacks from the enemy is not expected during the planning horizon. (8) Firing operations against targets assigned to more than one weapon, such as multiple-weapon targets, should be started simultaneously from the weapons (to achieve surprise effect). The end times of the operations from different weapons may be different. In the following, we present a mathematical formulation of the FSP. In the formulation, the following notation is used.

2. Problem description The re scheduling problem (FSP) considered in this study is the problem of scheduling a set of ring operations with the objective of minimizing total threat from the targets, or opponents, that survive the ring attack. We consider two types of targets in this research, single- and multiple-weapon targets. Single-weapon targets, also called point targets, have small target areas and hence a single weapon is assigned to a point target. On the other hand, multiple-weapon targets, or area targets, cannot be handled by a single weapon, and hence multiple weapons are assigned to an area target. That is, each area target is partitioned into sub-areas and a weapon is assigned to each sub-area. In general, these partitioned targets should be red at simulta-

ARTICLE IN PRESS
Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392 385

Indices and parameters i j Vj Ej dij index for weapons (i = 1,y,m) index for targets (j = 1,y,n) initial threat from target j set of weapons assigned to target j duration (in periods) of the ring operation from weapon i against target j initial destruction probability, i.e., the probability that pij target j is destroyed by the assigned periods f the ring operation from weapon i setup time (in periods) required on weapon i before a si ring operation aj decreasing rate of the destruction probability of the attack against target j W(j,k) set of weapons that are needed for both of the ring operations against targets j and k M a large number Decision variables xjk binary variable that is equal to 1 if the ring operation against target j precedes the ring operation against target k, and 0 otherwise start time of the ring operation(s) against target j

constraint that ring operations of the same job should be started simultaneously. We give a simple example in which a schedule can be represented by a sequence of jobs. See Fig. 1 for example, in which three jobs are scheduled on three weapons. Jobs 1 and 2 represent ring attacks against (multiple-weapon) targets 1 and 2, respectively, while job 3 represents a ring attack against a singleweapon target, target 3. In the gure, Tj denotes the ring operation against target j and Si denotes the setup operation on weapon i. Upon completion of a ring operation, a setup operation is performed for a subsequent ring operation on each weapon. Note that there may be idle time after a setup operation since other weapon(s) for the ring attack against the same target may not be ready. The schedule given in the gure can be represented by job sequence (132). With this solution representation method, one can always generate a unique non-delay schedule from a sequence of the jobs. However, there may be multiple sequences representing the same schedule. Note that the schedule given in Fig. 1 can also be represented by sequence (312). In this study, we use the following scheme to avoid such redundant representation. 2.1. Solution representation scheme Schedules are represented with a sequence of jobs. On each weapon, (setup for) a ring operation of a job in an earlier position of the sequence is scheduled earlier. If two targets require ring operations that are assigned to two disjoint sets of weapons, we only consider a sequence in which the job with an operation assigned to the lowest-indexed weapon is placed earlier. With this schedule representation scheme, the schedule given in Fig. 1 is represented with sequence (132) only, and the other sequence, (312), is not considered in the branch and bound procedure to be presented in this paper. For a better description of the problem and the solution representation scheme used in this study, we give another example, in which there are three weapons, four targets, and six ring operations, i.e., weapontarget pairs. Table 1 shows relevant data including the weapontarget allocation, durations of the ring operations and setup times on the weapons. In the example, targets 1 and 4 are

tj

Now, a non-linear integer programming formulation is given. X Y Vj f1 maxpij tj aj ; 0g Minimize

j1 i A Ej

subject to tk tj M 1 xjk Z max dij si

i A W j;k

8i A W j; k; j; k 8i A W j; k; j; k 8j 8j; k j o k

1 2 3 4

tj tk Mxjk Z max dik si

i A W j;k

tj Z 0 xjk A f0; 1g

The beginning time of period 1 is time 0 and the beginning time of period k is k1. Note that pij is the probability that target j is destroyed by the ring operation from weapon i if the ring operation is started at time 0, and the destruction probability decreases linearly (at the rate of aj) as time passes. Hence, max(pijtjaj, 0) denotes the probability that target j is destroyed by a ring attack from weapon i that is started at tj, and the value of {1 max(pij tjaj,0)} denotes the probability that target j survives the ring operation from weapon i that is started at time tj. Note that the objective function is a non-decreasing function of the start times (and completion times) of the ring operations. Hence, if a sequence of the ring operations is given, the start times and completion times of the operations can be easily determined. Constraints (1) and (2) are disjunctive constraints for the precedence relationship of each pair of two targets. Here, M should be large enough to make constraints (1) and (2) are always satised when xjk =0 and xjk = 1, respectively, say P M maxi;j dij maxi si n j 1 maxi dij . In this paper, schedules and partial schedules are represented with a sequence of jobs. On each weapon, a ring operation of a job in a former position of the sequence is scheduled earlier than that of a job placed in a latter position. In other words, the sequence of ring operations on each weapon is obtained from the sequence of jobs. Note that a job in a former position in the sequence is not necessarily started earlier than a job in a latter position if the two jobs are assigned to different weapons. When constructing a schedule from a sequence, one should consider the

W1 W2 W3

T1 T1 T3
S3 S2

T2 T2

Fig. 1. A simple example for solution representation.

Table 1 Data for the example. Target 1 2 3 4 Weapona 1 3 2 3 1 2

a

Duration 2 periods 1 1 2 1 1

(2) (1) (1) (1) (2) (1)

Setup times (in periods) for the weapons are given in parentheses.

ARTICLE IN PRESS
386 Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392

W1 W2

T1 T2 S2

S1

T4 T4

contradicts the optimality of the current schedule. Thus, an upper bound on the sum of idle times is (|Fi|2) (eij1). Therefore, the ring operation against any single-weapon target can be started at no later than Dieij +(|Fi|2)(eij1). & The above proposition reduces to the following corollary if only single-weapon targets are assigned to a weapon. Corollary 1. If only single-weapon targets are assigned to weapon i, then the ring operation against target j can be started no later than Dieij in an optimal solution. Proof. The targets that are assigned to weapon i can be considered independently without considering other targets assigned to other weapons. Since the rst ring operation begins at time 0 and there is no idle time in an optimal solution, the ring operations can be started at no later than Didijsi = Di eij. & We can develop dominance properties by using the above upper bound on the start time of a ring operation against a singleweapon target. The following proposition gives a dominance property related to single-weapon targets assigned to the same weapon. Proposition 2. If only single-weapon targets are assigned to weapon i and pij(Dieij)aj Z 0 is satised for all j, then in an optimal schedule, ring operations against these single-weapon targets are ordered in a non-increasing order of Vjaj/eij. Proof. Since only single-weapon targets are assigned to the weapon, the scheduling problem for the targets on the weapon can be considered independently. If pij (Di eij)aj Z 0, we have max(pij tjaj,0)= pij tjaj for all jAFi, since tj r Di eij. Therefore, the contribution of weapon i to the solution value is SVj{1 max(pij tjaj,0)}= {V1(1 pi1 + t1a1)+ V2(1 pi2 + t2a2) + ? + Vn(1 pin + tnan)}, which can be minimized by minimizing {V1a1t1 + V2a2t2 + ? + Vnantn} since Vj(1 pij) is a constant. Next, we prove the dominance of the sequence obtained in a non-increasing order of Vjaj/eij. Assume there is a sequence, S, that does not satisfy this sequencing rule. Then, in the sequence there is an adjacent pair of ring operations against targets j and k such that j precedes k but Vjaj/eij o Vkak/eik. Let t0 = tj and t00 = tk for tj and tk in S (t0 o t00 ). Consider a sequence, S0 , that is identical to S except that the positions of j and k are interchanged, and let g( ) be the value of {V1a1t1 + V2a2t2 + ? + Vnantn} of sequence . Since the two operations are adjacent, the starting time of none of the other operations changes by this interchange. Therefore, g S0 g S fVj aj t 00 eik eij Vk ak t 0 g fVk ak t 0 Vk ak t 00 g Vj aj t 00 eik eij t 0 Vk ak t 0 t 00 Vj aj eik Vk ak eij : The last equality comes from t00 t0 = eij. Since Vjaj/eij o Vkak/eik, we have g(S0 ) g(S) o 0, that is, S0 is better than S. By repeating this pairwise interchange argument, we can nd an optimal schedule in which ring operations are sequenced in a non-increasing order of Vjaj/eij. & In many real situations, only single-weapon targets are assigned to a weapon. In such cases, the above dominance property will be very effective in nding optimal schedules, especially if the weapon has large pij and the targets have small aj. A similar dominance property can be obtained for a more general case, in which single-weapon targets are assigned to a weapon along with multiple-weapon targets. The following proposition gives a dominance property related to a sequence of

W3

T1

S3

T3

Fig. 2. A feasible schedule.

multiple-weapon targets. Fig. 2 shows a feasible schedule for this FSP. According to the solution representation scheme, this schedule is represented with one sequence, (1243). Note that other sequences such as (1234), (1324), (2134), and (2143) result in the same schedule, but these sequences are not considered in the branch and bound algorithm to be developed in this study.

3. Dominance properties In this section, we present dominance properties that can be used in a branch and bound (B&B) algorithm for the re scheduling problem (FSP) considered in this paper. We use the following additional notation. Fi eij Di hj set of ring operations that are assigned to weapon i sum of the setup time and the duration for the ring operation from weapon i against target j, i.e., eij = dij + si the sum of the setup times and the durations for ring P operations assigned to weapon i, i.e., Di j A Fi eij maxi A Ej pij =aj , threshold start time of the ring operation against target j when a ring attack against target j has no effect on the threat from the target, that is, the smallest value of tj that satises pijtjaj r 0 for all iAEj sequences of ring operations objective function value of sequence

S, S0 f( )

First, we obtain an upper bound on the start time of a ring operation on a weapon in an optimal solution. Proposition 1. In an optimal schedule, the ring operation from weapon i against an arbitrary single-weapon target, say target j, starts at no later than Di eij jFi j 2eij 1. Proof. Without loss of generality, we assume the ring operation against target j is scheduled last in an optimal schedule. The start time of the ring operation against target j, i.e., tj, is given as the sum of three factors: (1) ring durations for preceding ring operations on the weapon; (2) setup times for those ring operations; and (3) idle time occurring because of multipleweapon targets. The sum of the rst two factors for the upper bound can be given as Dieij since the rst ring operation begins at time 0 (recall that the setup operation for the rst ring operation is assumed to have been done by time 0). We show an upper bound on the last factor (idle time) can be given as (|Fi|2)(eij1). Since the number of ring operations that precede the ring operation against target j is |Fi|1, there may be at most |Fi|2 positions for idle periods on the schedule for the weapon. Also, the length of idle time in each idle period is less than eij. Note that if there is an idle period with idle time longer than or equal to eij, we can improve the schedule by moving the ring operation against target j into the idle period, which

ARTICLE IN PRESS
Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392 387

ring operations against single-weapon targets regardless of adjacency of the ring operations. Proposition 3. Consider a weapon, weapon i, to which singleweapon targets as well as multiple-weapon targets are assigned. If ring operations against two or more single-weapon targets have the same ring duration, and if pij aj fDi eij jFi j 2eij 1g Z 0 is satised for all of those ring operations, js, then in an optimal schedule, those ring operations are sequenced in a non-increasing order of Vjaj. Proof. This proposition can be proven with a pairwise interchange argument as well. Since sequences of the ring operations against the single-weapon targets with the same ring duration can be changed without inuencing the sequence and start times of the other ring operations, sequences of those ring operations with the same duration can be considered independently. Assume there is a sequence, S, that does not satisfy this sequencing rule. Then, in S, there is an adjacent pair (without considering the other ring operations) of ring operations against targets, j and k, such that j precedes k but Vjaj o Vkak. Let t0 and t00 denote the start times of the ring operations against targets j and k in S, respectively (t0 o t00 ). Let S0 be the sequence that is identical to S except that the positions of j and k are interchanged and let f( ) be the objective function value of sequence . Since the starting time of none of the other operations changes by this interchange, we have f(S0 ) f(S) =(t00 t0 )(VjajVkak) o 0, that is, S0 is better than S. By repeating this pairwise interchange argument on the sequences of the ring operations against single-weapon targets with the same duration, we can nd an optimal schedule in which those ring operations are sequenced in a non-increasing order of Vjaj. & In the following ve propositions (Propositions 48), we present precedence relationships between two ring operations against single-weapon targets when the two ring operations can or should be started within a certain time interval. Proposition 4. Suppose single-weapon targets as well as multipleweapon targets are assigned to weapon i and pij aj {(Di eij)+(|Fi| 2)(eij 1)} Z 0 is satised for all single-weapon targets js. If the inequality Vjaj/eij Z Vkak/eik is satised for singleweapon targets j and k, then a sequence in which the ring operation against target k immediately precedes the ring operation against target j is dominated. Proof. Let S be a sequence in which k immediately precedes j, and let t0 and t00 be the start times of the ring operations against targets j and k in S, respectively (t00 o t0 ). Consider a sequence, S0 , that is identical to S except that the positions of j and k are interchanged. Note that the start time of none of the other operations changes by this interchange. Since contributions of ring operations against targets j and k to the solution value in S and S0 are Vj{1 max(pij t0 aj, 0)}+ Vk{1 max(pik t00 ak, 0)} and Vj{1 max(pij t00 aj, 0)}+ Vk[1 max{pik (t0 + eij eik)ak,0}], respectively, f S0 f S Vj fmaxpij t 0 aj ; 0 maxpij t 00 aj ; 0g Vk fmaxpik t 00 ak ; 0 maxpik t0 eij eik ak ; 0g: Since t00 o t0 r Di eij + (|Fi| 2)(eij 1) pij t0 aj Z 0 and pik t00 ak Z 0. Hence, from Proposition 1,

Proposition 5. Suppose ring operations against two single-weapon targets with the same ring duration, targets j and k, that are assigned to the same weapon, weapon i, are to be started no later than min(hj, hk). If Vjaj 4 Vkak, then it is better to start the ring operation against target j earlier than the operation against target k. Proof. Consider a sequence, S, in which k precedes j. We will prove this proposition by showing that interchanging the positions of the ring operations against the two targets decreases the solution value. Let t0 and t00 denote the start times of the ring operations against targets j and k in S, respectively (t00 o t0 ). Consider a sequence, S0 , that is identical to S except that the positions of j and k are interchanged. Since the two operations require the same ring duration, the start time of none of the other operations changes by this interchange. Since contributions of ring operations against targets j and k to the solution value in S and S0 are Vj{1 max(pij t0 aj, 0)}+ Vk{1 max(pik t00 ak, 0)} and Vj{1 max(pij t00 aj, 0)} + Vk{1 max(pik t0 ak,0)}, respectively, f S0 f S Vj fmaxpij t 0 aj ; 0 maxpij t 00 aj ; 0g Vk fmaxpik t 00 ak ; 0 maxpik t 0 ak ; 0g t 0 t 00 Vk ak Vj aj o 0: Therefore, interchanging the positions of the ring operations against the two targets decreases the solution value. & Proposition 6. Suppose ring operations against two single-weapon targets, targets j and k, that are assigned to the same weapon, weapon i, are to be started no later than min{hj, hk, hk eij + eik}. If Vkak/eik o Vjaj/eij, then sequences in which the ring operation against target k immediately precedes the operation against target j is dominated. Proof. Let S be a sequence in which k immediately precedes j, and let t0 and t00 be the start times of the ring operations against targets j and k in S, respectively (t00 o t0 ). Consider a sequence, S0 , that is identical to S except that the positions of j and k are interchanged. Note that the start time of none of the other operations changes by this interchange. Since contributions of ring operations against targets j and k to the solution value in S and S0 are Vj{1 max(pij t0 aj, 0)} + Vk{1 max(pik t00 ak, 0)} and Vj{1 max(pij t00 aj, 0)}+ Vk{1 max(pik (t0 + eij eik)ak,0)}, respectively, f S0 f S Vj fmaxpij t 0 aj ; 0 maxpij t 00 aj ; 0g Vk fmaxpik t 00 ak ; 0 maxpik t 0 eij eik ak ; 0g Vk ak eij Vj aj eik o 0: In the above, the second equality is valid because t00 o t0 o min{hj, hk, hkeij + eik} and the inequality is valid because Vkak/eik o Vjaj/eij. Therefore, S is dominated by S0 . & Proposition 7. Suppose two single-weapon targets with the same ring duration, targets j and k, are assigned to the same weapon, weapon i. If the ring operation against target k are to be started within time interval [hk, hj), then it is better to start the ring operation against target j earlier than the operation against target k. Proof. Consider a sequence, S, in which k precedes j and inequalities hk r tk o hj are satised. Let t0 and t00 denote the start times of the ring operations against targets j and k in S, respectively (t00 o t0 ). We show that interchanging the positions of the ring operations against the two targets decreases the solution value. Let S0 be the sequence that is identical to S except that the positions of j and k are interchanged. Since the two operations require the same ring duration, the start times of all other operations remain unchanged after this interchange. Since hkptk in both sequences, the contribution of target k to the

f S0 f S Vj t 00 t 0 aj Vk t 0 eij eik t 00 ak Vk ak eij Vj aj eik r 0: In the above, the second equality is valid because t0 t00 = eik, and the inequality is valid because Vkak/eik r Vjaj/eij. Since interchanging the positions (in S) of the ring operations against the two targets decreases the solution value, S is dominated. &

ARTICLE IN PRESS
388 Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392

solution value is Vk in both sequences. We consider two cases related to the value of t0 in S: t0 o hj and t0 Z hj. (Case 1: t0 o hj) Since contributions of the ring operations against targets j and k to the solution value in S and S0 are Vj{1 max(pij t0 aj,0)}+ Vk and Vj{1max(pij t00 aj,0)} + Vk, respectively, we have f(S0 ) f(S) = Vj{max(pijt0 aj, 0)max(pij t00 aj, 0)} o 0. (Case 2: t0 Z hj) Since the contributions in S and S0 are Vj + Vk and Vj{1 max(pij t00 aj, 0)} + Vk, respectively, we have f(S0 ) f(S)= Vj{max(pij t00 aj, 0)} o 0. Therefore, interchanging the positions of the ring operations against the two targets decreases the solution value. This completes the proof. & Proposition 8. Suppose two single-weapon targets, targets j and k, are assigned to the same weapon, weapon i. If the ring operation against target k are to be started within time interval [hk, hj), then sequences in which the ring operation against target k immediately precedes the operation against target j is dominated. Proof. Let S be a sequence in which k immediately precedes j, and let t0 and t00 be the start times of the ring operations against targets j and k in S, respectively (hk r t00 o t0 ). Consider a sequence, S0 , that is identical to S except that the positions of j and k are interchanged. Since the two operations are adjacent, the start time of none of the other operations changes by this interchange. Since hk r t00 in both sequences, the contribution of target k to the solution value is Vk in both sequences. We consider two cases related to the value of t0 in S: t0 o hj and t0 Z hj. (Case 1: t0 o hj) Since contributions of the ring operations against targets j and k to the solution value in S and S0 are Vj{1 max(pij t0 aj,0)}+ Vk and Vj{1 max(pij t00 aj,0)} + Vk, respectively, f(S0 ) f(S)= Vj{max(pij t0 aj, 0) max(pij t00 aj, 0)} o 0. (Case 2: t0 Z hj) Since the contributions in S and S0 are Vj + Vk and Vj{1 max(pij t00 aj, 0)} + Vk, respectively, we have f(S0 ) f(S) = Vj{max(pij t00 aj, 0)} o 0. Therefore, interchanging the positions of the ring operations against the two targets decreases the solution value. This completes the proof. & The above propositions can be used in reducing the number of subproblems to be considered in a B&B algorithm. For example, the dominance condition given in Proposition 6 can be applied when the ring operation against target k is already sequenced at the last position of the partial sequence, r, currently being considered in the branch and bound procedure, and another ring operation is to be appended after r. If there is a ring operation against target j (not in r) that is to be started no later than min(hj, hk, hk eij + eik) once appended after r and that satises the inequality, Vkak/eik o Vjaj/eij, then this partial sequence with j appended (and its corresponding B&B node) can be deleted from further consideration in the B&B algorithm. Also, the dominance condition presented in Proposition 8 can be applied when the ring operation against target k included in the current partial sequence, r, is scheduled to be started within time interval [hk, hj). If there is a ring operation assigned to the same weapon (as the one for the ring operation against target k), then r can be deleted from further consideration in the B&B algorithm. In the following two propositions, we present conditions under which solution values are not affected by sequences of two ring operations against single-weapon targets. Proposition 9. Suppose two single-weapon targets with the same ring duration, targets j and k, are assigned to the same weapon, weapon i. If neither of the ring operations against the two targets can be started earlier than max(hj, hk), then the positions of the ring

operations against the two targets in the sequence can be interchanged without affecting the solution value. Proof. Consider a sequence in which the ring operation against target j precedes the one against target k. We show that interchanging the positions of the ring operations against the two targets does not change the solution value. Since the start times of the other operations remain the same after the interchange of the two operations, we compare the contributions of the two ring operations to the solution value. Since hj r tj and hk r tk in both cases, i.e., before and after the interchange, the contributions of the two operations are the same and equal to Vj + Vk in both cases. This completes the proof. & Proposition 10. Suppose two single-weapon targets, targets j and k, are assigned to the same weapon, weapon i. If ring operations against the targets are placed adjacent to each other and neither can be started earlier than max(hj, hk) in a partial sequence, then the positions of the two ring operations can be interchanged without affecting the solution value. Proof. This can be proven easily as in the proof for Proposition 9. & Propositions 9 and 10 imply that if a pair of ring operations against single-weapon targets assigned to the same weapon is to be started after a specic time point, one of two sequences of the two operations can be deleted from consideration in a B&B algorithm.

4. A lower bound In this section, we describe the method used in this study for obtaining a lower bound on the solution value for a partial schedule. By the solution representation scheme described earlier, a partial schedule is represented with a sequence, r, of jobs or targets. Recall that a job denotes a set of ring (and setup) operations against a target. The lower bound associated with r is computed by adding the solution value of the schedule for jobs included in r and a lower bound on the solution value for jobs not included in r. The former can be easily obtained from r. For the computation of the latter, we classify the jobs not in r into two sets: the set of jobs whose contributions to the solution value are affected by upcoming scheduling decisions and the set of the other jobs. The rst set is composed of jobs corresponding to targets of which the value of the threshold start time (hj) is greater than the earliest possible start time of those jobs, while the second set is composed of the other jobs that are not included in r. Jobs in the second set can be sequenced arbitrarily (or may be skipped) and contributions of the jobs to the solution value can be easily computed as the sum of the initial threat values of the targets associated with the jobs. A lower bound on the solution values associated with the rst set of jobs is obtained by partitioning the jobs into a number of groups and obtaining an optimal schedule for each group independently through full enumeration. The sum of the solution values of these optimal schedules can be used as a lower bound. However, if the number of jobs included in a group is not small enough, it will take a long time to obtain an optimal schedule for the jobs. Thus, we limit the maximum number of jobs in a group to 8 (after tests on a few values for it). In the method suggested in this study, jobs are partitioned into the groups as follows. First, a weapon with the largest number of ring operations assigned to it is selected, and then a group is constituted with jobs associated with the ring operations assigned to the selected

ARTICLE IN PRESS
Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392 389

weapon. If there are more than eight jobs in the group, the jobs are divided into a number of groups of no more than eight jobs and this division is done arbitrarily. After formation of a group or groups for the selected weapon, a next group of jobs is identied by applying the same procedure to the jobs that have not been included in a group yet. This procedure for partitioning is repeated until there is no more job to be partitioned. For each group of jobs, an optimal schedule is obtained through full enumeration of possible sequences under the assumption that the jobs are sequenced after the jobs in r.

executed sequentially, and the one (ICIS) in which IC and IS are executed sequentially.

6. A branch and bound algorithm In the branch and bound (B&B) algorithm developed in this research, each node in the B&B tree corresponds to a partial sequence, r, and a node in the ith level of the tree represents a partial sequence in which i jobs are scheduled. According to the solution representation scheme used in this study, the B&B tree is constructed as follows. A job included in a partial sequence corresponding to a node closer to the root node are placed in an earlier position in the sequence than those corresponding to its child nodes farther from the root node. For branching, i.e., for selecting a node to generate branches from, the depth-rst rule is employed in the B&B algorithm. That is, a node with the most jobs included in the associated partial sequence is selected for branching. In case of ties, a node with the lowest lower bound is selected. When child nodes are generated from a selected parent node, we check redundancy of (partial) sequences as described in the solution representation scheme. That is, if two targets require ring operations that are assigned to two disjoint sets of weapons, we only consider a sequence in which the job with an operation assigned to the lowest-indexed weapon is placed earlier. If a node is related with a sequence that does not satisfy this rule, the node is deleted from consideration (pruned) in the B&B procedure. When a node is generated, the dominance conditions given in Propositions 210 are checked, and nodes corresponding to dominated sequences are also pruned. For each child node that is not pruned, a lower bound is computed with the method presented earlier in this paper. Nodes with lower bounds greater than or equal to the current incumbent solution are deleted from further consideration. We obtain an initial upper bound (incumbent solution) using the heuristic algorithms presented in the previous section. The upper bound is updated whenever a complete schedule that is better than the current upper bound is found.

5. Heuristic algorithms To obtain an initial upper bound in the branch and bound (B&B) algorithm, we develop a heuristic algorithm. This heuristic algorithm can also be used in situations in which the B&B algorithm may not give solutions within a reasonably short time. Note that in certain circumstances the re scheduling problem should be solved within a short time, say within 5 min. The heuristic algorithm suggested here is expected to give a reasonably good solution in such a short computation time. We present two types of algorithms, a construction algorithm and local search algorithms. In all algorithms, we nd a sequence of jobs. Note that each sequence species a schedule for ring operations (according to the solution representation scheme given earlier). 5.1. Modied NEH algorithm (MNEH) NEH is an algorithm developed by Nawaz et al. [29] for mmachine permutation ow shop scheduling problems. In NEH, jobs are initially sorted in a non-increasing order of the sum of the processing times on the machines, and then a nal solution (sequence) is obtained in a constructive way, adding at each iteration a new job in that order, and then inserting it in the best position, i.e., the one that results in the best partial solution. In this study, NEH is modied based on the idea that the initial threat values and the decreasing rates of the destruction probabilities are important factors to be considered when the priorities or emergencies of the ring operations are determined. In the modied NEH algorithm, we use the sum of the initial threat (Vj) and decreasing rate (aj) of each target (or job) for the initial sorting of the jobs instead of the sum of processing times, and apply the same schedule-construction procedure as that of NEH. 5.2. Local search algorithms Four local search algorithms are suggested in this study. In these algorithms, starting from an initial solution, a set of neighborhood solutions is generated for the current solution and the best neighborhood solution is selected as the new current solution. In all of the four algorithms, we use the solution obtained by MNEH as the initial solution for the improvement procedure. To generate neighborhood solutions (sequences) from the current solution, we use two methods, insertion and interchange. In the insertion method, a set of neighborhood solutions is generated by selecting a job in the current sequence and inserting it into each of all other positions in the sequence. On the other hand, in the interchange method, a set of neighborhood solutions is generated by selecting a job in the current sequence and exchanging its position with each of all other jobs. The four algorithms tested in this study are the one (IS) with the insertion method for neighborhood generation, the one (IC) with the interchange method, the one (ISIC) in which IS and IC are

7. Computational experiments To evaluate performance of the suggested branch and bound (B&B) algorithm, we performed a series of tests on problem instances that were generated randomly in a way that the resulting problems reect real situations in Korea relatively well. First, to select a heuristic algorithm to be used in the B&B algorithm, we test the heuristic algorithms presented earlier. For this test, we generated 135 problem instances, ve instances for each of all combination of three levels for the number of weapons (3, 6, and 9), three levels for the number of targets (22, 24, and 26), and three levels for the percentage of single-weapon targets among all targets (20%, 50%, and 80%). The number of weapons needed for a multiple-weapon target was selected from 2 and 3 with an equal probability. The index(es) of weapon(s) assigned to each target was randomly selected. The durations (dij) of the ring operations were generated as follows: for a single-weapon target, the duration was set to either 2 or 3 with an equal probability; for a multiple-weapon target, the duration for a randomly selected weapon among those assigned to the target was set to 1 and the durations for the other weapons were set to either 2 or 3 with an equal probability. The setup times (si) on the weapons were generated from the discrete uniform distribution with range [1, 4]. The initial threat values (Vj) were generated from the discrete uniform distribution with range

ARTICLE IN PRESS
390 Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392

[2, 100], while the initial destruction probabilities (pij) and the decreasing rates of destruction probabilities (aj) were randomly generated from the uniform distributions with ranges [0.6, 0.8] and [0.01, 0.1], respectively. A military manual called the joint munitions effectiveness manual (JMEM) provides the destruction probabilities, which can be determined according to the types of munitions, postures of targets, impact angles, radii of targets, etc, and the JMEM can be used for the generation of the values of pij. However, since it is classied information, we arbitrarily set the ranges for pij and aj here for the tests. All the algorithms tested in this study were coded in C, and computational experiments were performed on a personal computer with a Pentium 4 processor operating at 2.8 GHz clock speed. Since optimal solutions cannot be obtained in a reasonable amount of time, we compare the solutions with the relative deviation index (RDI), which is dened as (Ta TB)/(TW TB) for algorithm a, where Ta, TB, and TW are the solutions of algorithm a, the algorithm which gave the best solution value, and the algorithm which gave the worst solution value, respectively. Results of the test are given in Table 2, which shows the average and standard deviation of RDI values and the number of problems for which each algorithm found the best solutions. As can be seen from the table, the local search algorithms signicantly improved the solutions obtained from MNEH. Among the local search algorithms, ICIS worked best. All the algorithms required less than 1 s for a problem of the largest size. Since it takes very short time to solve a problem with the heuristic algorithm, we execute both of ICIS and ISIC and select the better solution as the initial upper bound in our B&B algorithm.

Before evaluating the performance of the B&B algorithm, we test the effectiveness of the dominance properties, the lower bound, and the heuristic algorithm used in the B&B algorithm. For this test, we compare four B&B algorithms: BB, the B&B algorithm with all of these included; BBDP, the one without dominance properties, i.e., the one with the lower bound and the heuristic algorithm only; BBLB, the one without the lower bound; and BB H, the one without the heuristic algorithm. For this test, 135 problem instances, ve instances for each of all combinations of three levels for the number of weapons (3,6, and 9), three levels for the number of targets (16, 17, and 18), and three levels for the percentage of single-weapon targets among all targets (20%, 50%, and 80%). To avoid excessive computation time for the test, we set the upper limit of the CPU time for a problem to 3600 s. Results of the test are given in Table 3, which shows (a lower bound on) the average CPU time, the number of problem instances that were not solved to the optimality in 3600 s, and the average ratio of the number of nodes generated to that of nodes that would have been generated if none of the dominance properties, the lower bound and the heuristic algorithm had been used in the B&B algorithm. None of the problem instances were solved by BB LB within the time limit, and hence results of BBLB are not given in the table. Note that the lower bound on the average CPU time was computed by using 3600 s if the algorithm could not solve the

Table 4 Results of the comparison of the algorithms. Wa Tb Average ratio of CPU timesc TBBDP/TBB TBBH/TBB Average ratio of nodes generatedd NBBDP/NBB NBBH/ NBB 102.737 85.470 4.189 1.458 1.170 1.700 2.541 6.450 16.501

Table 2 Performance of the heuristic algorithms. 3 Methods Relative deviation index Mean Standard deviation 0.237 0.338 0.254 0.232 0.092 8 48 45 69 114 NBSa 6 16 17 18 16 17 18 16 17 18
a,b

1.113 1.210 1.291 1.893 1.228 1.229 1.435 1.193 1.792

3.970 6.480 3.740 1.126 1.064 1.151 1.611 1.199 4.433

1.256 1.219 1.377 2.053 1.176 1.144 1.435 1.534 1.828

MNEH Insertion (IS) Interchange (IC) ISIC ICIS

a

0.941 0.254 0.173 0.149 0.026

Number of problems (among 135 problems) for which each algorithm found the best solutions.

c d

See the footnotes of Table 3. TX denotes the CPU time required to solve a problem using algorithm X. NX denotes the number of nodes considered for a problem in algorithm X.

Table 3 Performance of the B&B algorithms. Wa Tb BB CPUT 3 16 17 18 16 17 18 16 17 18

a c

BBDP NPNS 0 0 0 0 0 0 0 0 0
d

BBH NPNS 0 0 0 0 0 0 0 0 0 RNN 6.4 10 1.6 1011 5.3 1011 5.6 1010 7.0 1011 5.7 1012 4.6 1010 7.7 1011 1.1 1011
11

RNN

CPUT
11

CPUT 18.8 64.9 159.7 11.0 27.8 72.0 23.9 143.8 348.6

NPNS 0 0 0 0 0 0 0 0 1

RNN 5.3 109 1.1 109 1.6 1010 4.0 1010 6.9 1011 8.5 1012 8.2 1010 3.2 1010 9.7 1011

4.7 10.0 42.7 9.7 26.1 62.6 14.9 120.0 78.6

5.1 10 1.3 10 11 3.8 1011 2.7 1010 5.9 1011 5.0 1012 3.2 1010 5.0 1011 5.9 1012

5.3 12.1 55.2 18.4 32.0 76.9 21.3 143.1 140.9

Number of weapons. Number of targets. c (Lower bound on the) average CPU time (in seconds) required to solve a problem. d Number of problems (among ve problems) that were not solved to the optimality in 3600 s. e Average ratio of the number of nodes considered in the B&B procedure to that of nodes that would have been generated if no nodes had been fathomed.
b

ARTICLE IN PRESS
Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392 391

Table 5 Results of CPLEX and BB. Ta CPLEX CPUTb 11 13 15 17 19

a

Table 6 Results of the main test. BB NPNSc 0 0 0 2 3 CPUT 0.200 1.490 28.970 280.420 1758.017 NPNS 0 0 0 0 1 3 20 21 22 20 21 22 20 21 22
a,b c

Wa

Tb

BB CPUTc 228.4 602.8 360.4 236.3 563.3 934.5 895.8 652.0 1078.9 NPNSd 0 2 0 0 0 2 3 2 3

Heuristic algorithm APGe (%) 4.3 3.7 3.6 1.9 0.6 3.0 1.4 2.4 4.3 NOBSf 4 3 5 6 10 5 7 8 7

2.068 225.078 650.036 2492.024 3600.033

Number of targets. Average CPU time (in seconds) required to solve a problem. Number of problems (among three problems) that were not solved to the optimality in 3600 s.
b c

problem within 3600 s. For a better exposition of the results, we prepared another table, Table 4, which shows the ratios of the CPU times and the ratios of the numbers of nodes generated in the B&B procedures. The lower bound was more effective than the dominance properties and the heuristic in reducing the computation time as well as in fathoming nodes in the B&B algorithm. Since most of the dominance properties used in the algorithm are related with single-weapon targets only, they were not very effective, especially for the problem instances in which the percentage of singleweapon targets was small. However, use of the dominance properties and the heuristic algorithm reduced CPU time, although the reduction was not as impressive as that obtained from the lower bound. This means that the benet of using the dominance properties, i.e., pruning dominated partial solutions, exceeds the cost of using them, i.e., the time required for checking the dominance conditions or for obtaining a good initial upper bound. In the following, we further test BB, the B&B algorithm with the dominance properties and heuristic algorithm as well as the lower bound. We compare BB with CPLEX 11.0, a commercial software package for linear optimization problems and several extensions including quadratic programming problems. Note that CPLEX cannot handle all non-linear objective functions but just quadratic objective functions. Therefore, in this comparison, the maximum number of weapons needed for a multiple-weapon target was limited to 2 so that the objective function becomes quadratic. In addition, the re scheduling problem (FSP) considered in this study is a non-convex programming problem (due to the max term in the objective function), which is known to be very difcult to solve. To make the problem solvable with CPLEX, we considered a special case of the FSP by setting the values of pij and aj to make pij tjaj be positive. The value of pij was set to be large (high destruction probabilities) while the value of aj was set to be small (low decreasing rate), which represents a case in which the targets do not move much such as commanding posts, combat support facilities, etc. We generated pij and aj from the uniform distributions with ranges [0.7, 0.9] and [0.001, 0.007], respectively. For this comparison, we generated 15 3-weapon problem instances, one instance for each combination of ve levels for the number of targets (11, 13, 15, 17, and 19) and three levels for the percentage of single-weapon targets. Note that problems with larger sizes were not included in this test since they are not expected to be solved by CPLEX. Results of the test are given in Table 5, which shows the average computation time and the number of problems that were not solved to the optimality within the CPU time limit of 1 h. As can be seen from the table, the B&B algorithm outperformed CPLEX. Two problem instances (out of 3) with 17 targets were not solved to the optimality by CPLEX within 3600 s for problems, and none of the 19-target problems were

See the footnotes of Table 3. (Lower bound on the) average CPU time (in seconds) required to solve a problem. d Number of problems (among 15 problems) that were not solved to the optimality in 3600 s. e Average percentage gap of the heuristic solutions from the solutions of BB. f Number of problems (among 15 problems) for which the heuristic algorithm gave the same solutions as those of BB.

solved to the optimality by CPLEX, while all problem instances except for one 19-target instance were solved by the B&B algorithm. We further test the performance of the B&B algorithm on larger size problem instances. For this test, 135 problem instances were generated, ve instances for each of all combinations of three levels for the number of weapons (3, 6, and 9), three levels for the number of targets (20, 21, and 22), and three levels for the percentage of single-weapon targets among all targets (20%, 50%, and 80%). As in the previous test, the upper limit of 3600 s is given for each problem. Results of the test are given in Table 6. The table shows (a lower bound on) the average CPU time and the number of problems which were not solved to optimality within the time limit as well as the average percentage gap of the heuristic solutions from the solutions of BB and the number of problems for which the heuristic algorithm gave the same solutions as those obtained from BB. The B&B algorithm solved (to the optimality) 123 problem instances out of 135 within 3600 s. The CPU time was affected more by the number of targets than the number of weapons. Also, it took longer time to solve the problems or the problems became more difcult to solve with the B&B algorithm, when the number of weapons is larger. This may be because, if there are more weapons (when the number of target is given), the number of ring operations assigned to each weapon is larger and there are more possible sequences for those ring operations. In general, problems with a larger percentage of single-weapon targets required shorter time. This may be because most dominance properties used in the B&B algorithm are related with singleweapon targets and they become more effective when the percentage of single-weapon targets is larger. The heuristic algorithm worked relatively well. In 55 problem instances it gave the same solutions as those obtained from the B&B algorithm, and the overall average percentage gap of the heuristic solution (from the solutions of BB) was 2.8%.

8. Concluding remarks In this paper, we presented a branch and bound (B&B) algorithm for the problem of scheduling ring operations against movable targets with the objective of minimizing total threat from the targets

ARTICLE IN PRESS
392 Y.-H. Cha, Y.-D. Kim / Omega 38 (2010) 383392

after the ring attack is completed. We developed dominance properties and a lower bound as well as heuristic algorithms, which can be used for obtaining an initial upper bound in the B&B algorithm as well as for obtaining good solutions in a short time. Results of computational experiments showed that the suggested B&B algorithm can nd optimal solutions for medium-size problems in a reasonable amount of CPU time and the heuristic algorithm can give good solutions within a very short time. This research can be extended in several directions. For instance, one can develop more efcient and/or effective dominance properties since the dominance properties developed in this study are not very effective, especially when a large portion of the targets are multiple-weapon targets, and practical problems in real planned artillery attack operations may be larger than what can be solved with the suggested B&B algorithm. Also, metaheuristics such as the tabu search, simulated annealing and genetic algorithms may be developed for better heuristic solutions, since one can use a little more time than what is needed for the heuristic algorithm developed in this study. In addition, we may have to consider the weapon-target assignment problem (WTAP) and the re scheduling problem (FSP) simultaneously, since the two problems are closely related with each other. In such a combined problem, the FSP may have to be solved many times for each feasible solution for WTAP to obtain an overall optimal or good solution, and hence very efcient solution methods are needed for each problem.

References
[1] Ahuja RK, Kumar A, Krishna JB, Orlin. Exact and heuristic algorithms for the weapon-target assignment problem. Operations Research 2007;55:113646. [2] Matlin S. A review of the literature on the missile-allocation problem. Operations Research 1970;18:33473. [3] Eckler AR, Burr SA. Mathematical models of target coverage and missile allocation. Alexandria, VA, USA: Military Operations Research Society; 1972. [4] Manne A. A target assignment problem. Operations Research 1958;6:34651. [5] Lloyd SP, Witsenhausen HS. Weapons allocation is NP-complete. In: Proceedings of the 1986 summer conference on simulation. NV, USA; 1986. p. 10548. [6] Hosein PA, Athans M. An asymptotic result for the multi-stage weapon-target allocation problem. In: Proceedings of the 29th conference on decision and control. HI, USA; 1990. p. 2405. [7] Christ JF, Page EW, Tagliarini GA. Performance evaluation of a neural network for weapon-to-target assignment. In: Proceedings of SPIEthe International society for optical engineering. FL, USA; 1993. p. 67681.

[8] Erdem E, Ozdemirel NE. An evolutionary approach for the target allocation problem. Journal of the Operational Research Society 2003;54:95869. [9] Lee ZJ, Lee WL. A hybrid search algorithm of ant colony optimization and genetic algorithm applied to weapon-target assignment problems. Lecture Notes in Computer Science 2003;2690:27885. [10] Malcolm WP. On the character and complexity of certain defensive resource allocation problems. DSTO-TR-1570, Australian Government Department of Defence, 2004. [11] Karasakal W, Orhan R. Air defense missile-target allocation models for a naval task group. Computers and Operations Research 2008;35: 17591770. [12] C - etin E, Esen ST. A weapon-target assignment approach to media allocation. Applied Mathematics and Computation 2006;175:126675. , C [13] Esen O - etin E, Esen ST. A mathematical immunochemoradiotherapy model: a multiobjective approach. Nonlinear Analysis: Real World Applications 2008;9:5117. [14] Kwon O, Lee K, Park S. Targeting and scheduling problem for eld artillery. Computers and Industrial Engineering 1997;33:6936. [15] Chen JS. Optimization models for the tool change scheduling problem. Omega 2008;36:88894. [16] Balas E, Simonetti N, Vazacopoulos A. Job shop scheduling with setup times, deadlines and precedence constraints. Journal of Scheduling 2008;11:25362. [17] Laha D, Sarin SC. A heuristic to minimize total ow time in permutation ow shop. Omega 2009;37:7349. [18] Dobson G, Karmarkar US. Simultaneous resource scheduling to minimize weighted ow times. Operations Research 1989;37:592600. [19] Hoogeveen JA, Van de Velde SL, Veltman B. Complexity of scheduling multiprocessor tasks with prespecied processor allocations. Discrete Applied Mathematics 1994;55:25972. [20] Drozdowski M. Scheduling multiprocessor tasksan overview. European Journal of Operational Research 1996;94:21530. [21] Lee CY, Lei L, Pinedo M. Current trends in deterministic scheduling. Annals of Operations Research 1997;70:141. [22] Allahverdi A, Gupta NDJ, Aldowaisan T. A review of scheduling research involving setup considerations. Omega 1999;27:21939. [23] Chen J, Lee CY. General multiprocessor tasks scheduling. Naval Research Logistics 1999;46:5774. [24] Guan Y, Xiao WQ, Cheung RK, Li CL. A multiprocessor task scheduling model for berth allocation: heuristic and worst-case analysis. Operations Research Letters 2002;30:34350. [25] Serifoglu FS, Ulusoy G. Multiprocessor task scheduling in multistage hybrid ow-shops: a genetic algorithm approach. Journal of the Operational Research Society 2004;55:50412. [26] Huang J, Chen J, Chen S, Wang J. A simple linear time approximation algorithm for multi-processor job scheduling on four processors. Journal of Combinatorial Optimization 2007;13:3345. [27] Philippe B. A note on scheduling multiprocessor tasks with identical processing times. Computers and Operations Research 2003;30:20718. [28] Kramer A. Branch and bound methods for scheduling problems with multiprocessor tasks on dedicated processors. OR Spektrum 1997;19:21927. [29] Nawaz M, Enscore Jr EE, Ham I. A heuristic algorithm for the m-machine, njob ow-shop sequencing problem. Omega 1983;11:915.