arxiv: v1 [cs.ds] 15 Sep 2018

Transcription

1 Approximation algorithms for the three-machine proportionate mixed shop scheduling Longcheng Liu Yong Chen Jianming Dong Randy Goebel Guohui Lin Yue Luo Guanqun Ni Bing Su An Zhang arxiv: v1 [cs.ds] 15 Sep 2018 September 18, 2018 Abstract A mixed shop is a manufacturing infrastructure designed to process a mixture of a set of flowshop jobs and a set of open-shop jobs. Mixed shops are in general much more complex to schedule than flow-shops and open-shops, and have been studied since the 1980 s. We consider the three machine proportionate mixed shop problem denoted as M3 prpt C max, in which each job has equal processing times on all three machines. Koulamas and Kyparisis [European Journal of Operational Research, 243:70 74,2015] showed that the problem is solvable in polynomial time in some very special cases; for the non-solvable case, they proposed a 5/3-approximation algorithm. In this paper, we present an improved 4/3-approximation algorithm and show that this ratio of 4/3 is asymptotically tight; when the largest job is a flow-shop job, we present a fully polynomialtime approximation scheme (PTAS). On the negative side, while the 3 prpt C max problem is polynomial-time solvable, we show an interesting hardness result that adding one open-shop job to the job set makes the problem NP-hard if this open-shop job is larger than any flow-shop job. We are able to design an PTAS for this special case too. Keywords: Scheduling; mixed shop; proportionate; approximation algorithm; fully polynomialtime approximation scheme 1 Introduction We study the following three-machine proportionate mixed shop problem, denoted as M 3 prpt C max in the three-field notation [4]. Given three machines,, M 3 and a set J = O of jobs, where = {J 1, J 2,..., J l } and O = {J l+1, J l+2,..., }, each job J i needs to be processed non-preemptively through,, M 3 sequentially with a processing time p i on each machine and each job J i O needs to be processed non-preemptively on,, M 3 in any machine order, with a processing time q i on each machine. The scheduling constraint is usual in that at every time School of Mathematical Sciences, Xiamen University. Xiamen, China. longchengliu@xmu.edu.cn Department of Computing Science, University of Alberta. Edmonton, AB T6G 2E8, Canada. s: {rgoebel, guohui}@ualberta.ca Department of Mathematics, Hangzhou Dianzi University. Hangzhou, China. s: {chenyong, Department of Mathematics, Zhejiang Sci-Tech University. Hangzhou, China. djm226@163.com Correspondence author. ORCID: College of Management, ujian Agriculture and orestry University. uzhou, China. guanqunni@163.com School of Economics and Management, Xi an Technological University. Xi an, China. 1

2 2 Liu et al. /v:september 18, 2018 point a job can be processed by at most one machine and a machine can process at most one job. The objective is to minimize the maximum job completion time, i.e., the makespan. The jobs of are referred to as flow-shop jobs and the jobs of O are called open-shop jobs. The mixed shop problem is to process such a mixture of a set of flow-shop jobs and a set of open-shop jobs. We assume without loss of generality that p 1 p 2... p l and q l+1 q l+2... q n. Mixed shops have many real-life applications and have been studied since the 1980 s. The scheduling of medical tests in an outpatient health care facility and the scheduling of classes/exams in an academic institution are two typical examples, where the patients (students, respectively) must complete a number of medical tests (academic activities, respectively); some of these activities must be done in a specified sequential order while the others can be finished in any order; and the time-spans for all these activities should not overlap with each other. The proportionate shops were also introduced in the 1980 s [10] and they are one of the most specialized shops with respect to the job processing times which have received many studies [11]. Masuda et al. [9] and Strusevich [15] considered the two-machine mixed shop problem to minimize the makespan, i.e., M2 C max ; they both showed that the problem is polynomialtime solvable. Shakhlevich and Sotskov [12] studied mixed shops for processing two jobs with an arbitrary regular objective function. Brucker [1] surveyed the known results on the mixed shop problems either with two machines or for processing two jobs. Shakhlevich et al. [14] studied the mixed shop problems with more than two machines for processing more than two jobs, with or without preemption. Shakhlevich et al. [13] reviewed the complexity results on the mixed shop problems with three or more machines for processing a constant number of jobs. When O =, the M3 prpt C max problem reduces to the 3 prpt C max problem, which is solvable in polynomial time [2]. When =, the problem reduces to the O3 prpt C max problem, which is ordinary NP-hard [8]. It follows that the M3 prpt C max problem is at least ordinary NP-hard. Recently, Koulamas and Kyparisis [7] showed that for some very special cases, the M3 prpt C max problem is solvable in polynomial time; for the non-solvable case, they showed an absolute performance bound of 2 max{p 1, q l+1 } and presented a 5/3-approximation algorithm. In this paper, we design an improved 4/3-approximation algorithm for (the non-solvable case of) the M3 prpt C max problem, and show that the performance ratio of 4/3 is asymptotically tight. When the largest job is a flow-shop job, that is p 1 q l+1, we present a fully polynomialtime approximation scheme (PTAS). On the negative side, while the 3 prpt C max problem is polynomial-time solvable, we show an interesting hardness result that adding one single open-shop job to the job set makes the problem NP-hard if this open-shop job is larger than any flow-shop job (that is, = {J 1, J 2,..., 1 } and O = { }, and q n > p 1 ). We construct the reduction from the well-known Partition problem [3]. Denote the special case in which = n 1 and O = 1 as M3 prpt, (n 1, 1) C max. We propose an PTAS for this special case M3 prpt, (n 1, 1) C max. The rest of the paper is organized as follows. In Section 2, we introduce some notation and present a lower bound on the optimal makespan C max. We present in Section 3 the PTAS for the M3 prpt C max problem when p 1 q l+1. The 4/3-approximation algorithm for the case where p 1 < q l+1 is presented in Section 4, and the performance ratio of 4/3 is shown to be asymptotically tight. We show in Section 5 that, when the open-shop job is the only largest, the special case M3 prpt, (n 1, 1) C max is NP-hard, through a reduction from the Partition problem. Section 6 contains an PTAS for the special case M3 prpt, (n 1, 1) C max. We conclude the paper with some remarks in Section 7.

3 Approximating M3 prpt C max 3 2 Preliminaries or any subset of jobs X, the total processing time of the jobs of X on one machine is denoted as P (X ) = p i. or any subset of jobs Y O, the total processing time of the jobs of Y on one machine is denoted as Q(Y) = q i. J i Y J i X The set minus operation J \ {J} for a single job J J is abbreviated as J \ J throughout the paper. In a schedule π, we use S i j and Ci j to denote the start time and the finish time of the job J j on the machine M i, respectively, for i = 1, 2, 3 and j = 1, 2,..., n. Given that the load (i.e., the total job processing time) of each machine is P () + Q(O), the job J l+1 has to be processed by all three machines, and one needs to process all the flow-shop jobs of, the following lower bound on the optimum C max is established [2, 7]: C max max{p () + Q(O), 3q l+1, 2p 1 + P ()}. (1) 3 An PTAS for the case where p 1 q l+1 In this section, we design an approximation algorithm A(ɛ) for the M3 prpt C max problem when p 1 q l+1, for any given ɛ > 0. The algorithm A(ɛ) produces a schedule π with its makespan Cmax π < (1 + ɛ)cmax, and its running time polynomial in both n and 1/ɛ. Consider a bipartition {A, B} of the job set O = {J l+1, J l+2,..., }, i.e., A B = O and A B =. Throughout the paper, a part of the bipartition is allowed to be empty. The following six-step procedure Proc(A, B, ) produces a schedule π: 1. the jobs of are processed in the same longest processing time (LPT) order on all three machines, and every job is processed first on, then on, lastly on M 3 ; 2. the jobs of A are processed in the same LPT order on all three machines, and every one is processed first on, then on M 3, lastly on ; 3. the jobs of B are processed in the same LPT order on all three machines, and every one is processed first on M 3, then on, lastly on ; and 4. the machine processes (the jobs of) first, then B, lastly A, denoted as, B, A ; 5. the machine processes A first, then, lastly B, denoted as A,, B ; 6. the machine M 3 processes B first, then A, lastly, denoted as B, A,. Proc(A, B, ) runs in O(n log n) time to produce the schedule π, of which an illustration is shown in igure 3.1. Lemma 3.1 If both Q(A) p 1 and Q(B) p 1, then the schedule π produced by Proc(A, B, ) is optimal, with its makespan C π max = C max = 2p 1 + P ().

4 4 Liu et al. /v:september 18, 2018 A M 3 B A 0 p 1 2p 1 B A B igure 3.1: An illustration of the schedule π produced by Proc(A, B, ), where {A, B} is a bipartition of the set O and the jobs of each of A, B, are processed in the same LPT order on all three machines. Proof. The schedule depicted in igure 3.2 is feasible since we have the following inequalities: Q(A) p 1, Q(B) p 1. The makespan is achieved on M 3 and Cmax π = 2p 1 + P (), which meets the lower bound in Eq. (1). A B A B M 3 B A 0 p1 2p1 p1 + P () 2p 1 + P () igure 3.2: An illustration of the schedule π produced by Proc(A, B, ) when both Q(A) p 1 and Q(B) p 1. Lemma 3.2 If both Q(A) p 1 and Q(B) p 1, then the schedule π produced by Proc(A, B, ) is optimal, with its makespan C π max = C max = P () + Q(O). Proof. In this case, in the schedule π produced by Proc(A, B, ), the machine does not idle since p 1 q l+1 and the jobs of A, B and are processed in the LPT order. The machine does not idle either since Q(A) p 1 and the jobs of A, B and are processed in the LPT order. The machine M 3 does not idle since Q(B) p 1 q l+1, Q(A) + Q(B) 2p 1, and the jobs of A, B and are processed in the LPT order. Therefore, the makespan of this schedule is Cmax π = P () + Q(O), which meets the lower bound in Eq. (1) (see for an illustration in igure 3.3). This finishes the proof of the lemma. Now we are ready to present the approximation algorithm A(ɛ), for any ɛ > 0. In the first step, we check whether Q(O) p 1 or not. If Q(O) p 1, then we run Proc(O,, ) to construct a schedule π and terminate the algorithm. The schedule π is optimal by Lemma 3.1. In the second step, the algorithm A(ɛ) constructs an instance of the Knapsack problem [3], in which there is an item corresponding to the job J i O, also denoted as J i. The item J i has a profit q i and a size q i. The capacity of the knapsack is p 1. The Min-Knapsack problem is to find a subset of items of minimum profit that cannot be packed into the knapsack, and it admits an PTAS [6]. The algorithm A(ɛ) runs a (1 + ɛ)-approximation algorithm for the Min-Knapsack problem to obtain a job subset A. It then runs Proc(A, O \A, ) to construct a schedule, denoted as π 1.

5 Approximating M3 prpt C max 5 B A A B M 3 B A 0 p 1 2p 1 P () + Q(O) igure 3.3: An illustration of the schedule π produced by Proc(A, B, ) when both Q(A) p 1 and Q(B) p 1. The Max-Knapsack problem is to find a subset of items of maximum profit that can be packed into the knapsack, and it admits an PTAS, too [5]. In the third step, the algorithm A(ɛ) runs a (1 ɛ)-approximation algorithm for the Max-Knapsack problem to obtain a job subset B. Then it runs Proc(O \ B, B, ) to construct a schedule, denoted as π 2. The algorithm A(ɛ) outputs the schedule with a smaller makespan between π 1 and π 2. A high-level description of the algorithm A(ɛ) is provided in igure 3.4. Algorithm A(ɛ): 1. If Q(O) p 1, then run Proc(O,, ) to produce a schedule π; output the schedule π. 2. Construct an instance of Knapsack, where an item J i corresponds to the job J i O; J i has a profit q i and a size q i ; the capacity of the knapsack is p Run a (1 + ɛ)-approximation for Min-Knapsack to obtain a job subset A Run Proc(A, O \ A, ) to construct a schedule π Run a (1 ɛ)-approximation for Max-Knapsack to obtain a job subset B Run Proc(O \ B, B, ) to construct a schedule π Output the schedule with a smaller makespan between π 1 and π 2. igure 3.4: A high-level description of the algorithm A(ɛ). In the following performance analysis, we assume without of loss of generality that Q(O) > p 1. We have the following (in-)equalities inside the algorithm A(ɛ): OPT 1 = min{q(x ) X O, Q(X ) > p 1 }; (2) p 1 < Q(A) (1 + ɛ)opt 1 ; (3) OPT 2 = max{q(y) Y O, Q(Y) p 1 }; (4) p 1 Q(B) (1 ɛ)opt 2, (5) where OPT 1 (OPT 2, respectively) is the optimum to the constructed Min-Knapsack (Max- Knapsack, respectively) problem.

6 6 Liu et al. /v:september 18, 2018 Lemma 3.3 In the algorithm A(ɛ), if Q(O \ A) p 1 ɛopt 1, then for any bipartition {X, Y} of the job set O, Q(X ) > p 1 implies Q(Y) p 1. Proof. Note that the job subset A is computed in Step 2.1 of the algorithm A(ɛ), and it satisfies Eq. (3). By the definition of OPT 1 in Eq. (2) and using Eq. (3), we have Q(X ) OPT 1 Q(A) ɛopt 1. urthermore, from the fact that Q(O) = Q(X ) + Q(Y) = Q(A) + Q(O \ A) and the assumption that Q(O \ A) p 1 ɛopt 1, we have Q(Y) = Q(A) + Q(O \ A) Q(X ) Q(A) + Q(O \ A) (Q(A) ɛopt 1 ) = Q(O \ A) + ɛopt 1 p 1 ɛopt 1 + ɛopt 1 = p 1. This finishes the proof of the lemma. Lemma 3.4 In the algorithm A(ɛ), if Q(O \ A) p 1 ɛopt 1, then C max P () + Q(O) + p 1 OPT 2. Proof. Consider an arbitrary optimal schedule π that achieves the makespan Cmax. Note that the flow-shop job J 1 is first processed on the machine, then on machine, and last on machine M 3. On the machine, let J 1 = O 1 1 denote the subset of jobs processed before J 1, and J 2 = O 2 2 denote the subset of jobs processed after J 1, where {O 1, O 2 } is a bipartition of the job set O and { 1, 2 } is a bipartition of the job set \ J 1. Also, let δ 1 and δ 2 denote the total amount of machine idle time for before processing J 1 and after processing J 1, respectively (see igure 3.5 for an illustration). J 1 J 1 δ 1 J 1 δ J 2 2 M 3 J 1 0 S1 1 C1 1 S1 2 C1 2 S1 3 C1 3 C max igure 3.5: An illustration of an optimal schedule π, in which J 1 and J 2 are the subsets of jobs processed on before J 1 and after J 1, respectively; δ 1 and δ 2 are the total amount of machine idle time for before processing J 1 and after processing J 1, respectively. Note that = J is the set of flow-shop jobs. The job J 1 and the jobs of 1 should be finished on the machine before time S1 2, and the job J 1 and the jobs of 2 can only be started on the machine M 3 after time C1 2. That is, p 1 + P ( 1 ) S 2 1 (6) and p 1 + P ( 2 ) C max C 2 1. (7)

7 Approximating M3 prpt C max 7 If Q(O 1 ) p 1, then we have Q(O 1 ) OPT 2 by the definition of OPT 2 in Eq. (4). Combining this with Eq. (6), we achieve that δ 1 = S1 2 P ( 1 ) Q(O 1 ) p 1 OPT 2. If Q(O 1 ) > p 1, then we have Q(O 2 ) p 1 by Lemma 3.3. Hence, Q(O 2 ) OPT 2 by the definition of OPT 2 in Eq. (4). Combining this with Eq. (7), we achieve that δ 2 = Cmax C1 2 P ( 2 ) Q(O 2 ) p 1 OPT 2. The last two paragraphs prove that δ 1 + δ 2 p 1 OPT 2. Therefore, C max = Q(O 1 ) + P ( 1 ) + δ 1 + p 1 + Q(O 2 ) + P ( 2 ) + δ 2 = P () + Q(O) + δ 1 + δ 2 P () + Q(O) + p 1 OPT 2. This finishes the proof of the lemma. Lemma 3.5 In the algorithm A(ɛ), if Q(O \ A) p 1 ɛopt 1, then C π2 max < (1 + ɛ)c max. Proof. Denote B = O \ B. Note that the job set B computed in Step 3.1 of the algorithm A(ɛ) satisfies p 1 Q(B) (1 ɛ)opt 2, and the schedule π 2 is constructed by Proc(B, B, ). We distinguish the following two cases according to the value of Q(B). Case 1. Q(B) p 1. In this case, the schedule π 2 is optimal by Lemma 3.1. Case 2. Q(B) > p 1. The schedule π 2 constructed by Proc(B, B, ) has the following properties (see igure 3.6 for an illustration): B B B M 3 B B 0 p 1 p 1 + Q(B) P () + p 1 + Q(B) B igure 3.6: An illustration of the schedule π 2 constructed by Proc(B, B, ) in Case 2, where Q(B) p 1 and Q(B) > p 1. The machines and do not idle; the machine M 3 may idle between processing the job set B and the job set B and may idle between processing the job set B and the job set. M 3 starts processing the job set at time p 1 + Q(B). 1. The jobs are processed consecutively on the machine since J 1 is the largest job. The completion time of is thus C π2 1 = Q(O) + P (). 2. The jobs are processed consecutively on the machine due to Q(B) p 1 and Q(B) > p 1. The completion time of is thus C π2 2 = Q(O) + P (). 3. The machine M 3 starts processing the job set consecutively at time p 1 + Q(B) due to Q(B) p 1. The completion time of M 3 is C π2 3 = P () + p 1 + Q(B). Note that C3 π2 = P () + p 1 + Q(B) P () + Q(B) + Q(B) = Q(O) + P (), implying max = P () + p 1 + Q(B). Combining Eq. (5) with Lemma 3.4, we have C π2 C π2 max = P () + p 1 + Q(B)

8 8 Liu et al. /v:september 18, 2018 = P () + Q(O) + p 1 Q(B) P () + Q(O) + p 1 (1 ɛ)opt 2 C max + ɛopt 2 < (1 + ɛ)c max, where the last inequality is due to OPT 2 p 1 < C max. This finishes the proof of the lemma. Lemma 3.6 In the algorithm A(ɛ), if p 1 ɛopt 1 < Q(O \ A) < p 1, then C π1 max < (1 + ɛ)c max. Proof. Denote A = O \ A. Note that the job set A computed in Step 2.1 of the algorithm A(ɛ) satisfies p 1 < Q(A) (1 + ɛ)opt 1, and the schedule π 1 is constructed by Proc(A, A, ). By a similar argument as in Case 2 in the proof of Lemma 3.5, replacing the two job sets B, B by the two job sets A, A, we conclude that the makespan of the schedule π 1 is achieved on the machine M 3, Cmax π1 = P () + Q(O) + p 1 Q(A). Combining Eq. (1) with the assumption that p 1 ɛopt 1 < Q(A), we have C π1 max < P () + Q(O) + ɛopt 1 C max + ɛopt 1 < (1 + ɛ)c max, where the last inequality follows from OPT 1 Q(O) C max. This finishes the proof of the lemma. Theorem 3.7 The algorithm A(ɛ) is an O(n min{log n, log (1/ɛ)}+1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)})- time (1 + ɛ)-approximation for the M3 prpt C max problem when p 1 q l+1. Proof. irst of all, the procedure Proc(X, Y, ) on a bipartition {X, Y} of the job set O takes O(n log n) time. Recall that the job set A is computed by a (1 + ɛ)-approximation for the Min-Knapsack problem, in O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)}) time; the other job set B is computed by a (1 ɛ)-approximation for the Max-Knapsack problem, also in O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)}) time. The total running time of the algorithm A(ɛ) is thus polynomial in both n and 1/ɛ too. When Q(O) p 1, or the job set O \ A computed in Step 2.1 of the algorithm A 1 (ɛ) has total processing time not less than p 1, the schedule constructed in the algorithm A(ɛ) is optimal by Lemmas 3.1 and 3.2. When Q(O \ A) < p 1, the smaller makespan between the two schedules π 1 and π 2 constructed by the algorithm A(ɛ) is less than (1 + ɛ) of the optimum by Lemmas 3.5 and 3.6. Therefore, the algorithm A(ɛ) has a worst-case performance ratio of (1 + ɛ). This finishes the proof of the theorem. 4 A 4/3-approximation for the case where p 1 < q l+1 In this section, we present a 4/3-approximation algorithm for the M3 prpt C max problem when p 1 < q l+1, and we show that this ratio of 4/3 is asymptotically tight. Theorem 4.1 When p 1 < q l+1, the M3 prpt C max problem admits an O(n log n)-time 4/3- approximation algorithm.

9 Approximating M3 prpt C max 9 Proof. Consider first the case where there are at least two open-shop jobs. Construct a permutation schedule π in which the job processing order for is J l+3,...,,, J l+1, J l+2, where the jobs of are processed in the LPT order; the job processing order for is J l+2, J l+3,...,,, J l+1 ; the job processing order for M 3 is J l+1, J l+2, J l+3,...,,. See igure 4.1 for an illustration, where the start time for J l+3 on is q l+1, and the start time for J l+3 on M 3 is 2q l+1. One can check that the schedule π is feasible when p 1 < q l+1, and it can be constructed in O(n log n) time. J l+3,..., J l+2 J l+3,..., M 3 J l+1 J l+2 J l+3,..., 0 q l+1 2q l+1 C π J l+1 J l+2 Jl+1 igure 4.1: A feasible schedule π for the M3 prpt C max problem when there are at least two open-shop jobs and p 1 < q l+1. The makespan of the schedule π is C π max = P () + Q(O) + q l+1 q l+2. Combining this with Eq. (1), we have max C π max P () + Q(O) + q l C max. When there is only one open-shop job (that is, = {J 1, J 2,..., 1 } and O = { }), construct a permutation schedule π in which the job processing order for is,, where the jobs of are processed in the LPT order; the job processing order for is, ; the job processing order for M 3 is, (see for an illustration in igure 4.2). If P () q n, which is shown in igure 4.2, then π has makespan 3q n and thus is optimal. If P () > q n, then π has M 3 0 q n 2q n 3q n igure 4.2: A feasible schedule π for the M3 prpt C max problem when there is only one open-shop job and p 1 < q n ; the configuration shown here corresponds to P () q n and thus π is optimal. makespan C π max 2q n + P () 4 3 C max by Eq. (1). This finishes the proof of the theorem. Remark 4.2 Construct an instance in which p i = 1 l 1 for all i = 1, 2,..., l, q l+1 = 1 and q i = 1 n l 2 for all i = l + 2, l + 3,..., n. Then for this instance, the schedule π constructed in the proof of Theorem 4.1 has makespan Cmax π = l 1 ; an optimal schedule has makespan C max = l n l 2 (see for an illustration in igure 4.3). This suggests that the approximation ratio of 4/3 is asymptotically tight for the algorithm presented in the proof of Theorem 4.1.

10 10 Liu et al. /v:september 18, 2018 O \ J l+1 O \ J l+1 J l+1 O \ J l+1 J l+1 M 3 J l Cmax = l n l 2 igure 4.3: An optimal schedule for the constructed instance of the M3 prpt C max problem, in which p i = 1 l 1 for all i = 1, 2,..., n, q l+1 = 1 and q i = 1 n l 2 for all i = l + 2, l + 3,..., n. 5 NP-hardness for M3 prpt, (n 1, 1) C max when p 1 < q n Recall that we use M3 prpt, (n 1, 1) C max to denote the special of M3 prpt C max where there is only one open-shop job, i.e., = {J 1, J 2,..., 1 } and O = { }. In this section, we show that this special case M3 prpt, (n 1, 1) C max is already NP-hard if the unique open-shop job is larger than any flow-shop job, i.e., p 1 < q n. We prove the NP-hardness through a reduction from the Partition problem [3], which is a well-known NP-complete problem. Theorem 5.1 The M3 prpt, (n 1, 1) C max problem is NP-hard if the unique open-shop job is larger than any flow-shop job. Proof. An instance of the Partition problem consists of a set S = {a 1, a 2, a 3,..., a m } where each a i is a positive integer and a 1 + a a m = 2B, and the query is whether or not S can be partitioned into two parts such that each part sums to exactly B. Let x > B, and we assume that a 1 a 2... a m. We construct an instance of the M3 prpt C max problem as follows: there are in total m + 2 flow-shop jobs, and their processing times are p 1 = x, p 2 = x, and p i+2 = a i for i = 1, 2,..., m; there is only one open-shop job with processing time q m+3 = B + 2x. Note that the total number of jobs is n = m + 3, and one sees that the open-shop job is larger than any flow-shop job. If the set S can be partitioned into two parts S 1 and S 2 such that each part sums to exactly B, then we let J 1 = J 1 {J i a i S 1 } and J 2 = J 2 {J i a i S 2 }. We construct a permutation schedule π in which the job processing order for is J 1, J 2, J m+3, where the jobs of J 1 and the jobs of J 2 are processed in the LPT order, respectively; the job processing order for is J 1, J m+3, J 2 ; the job processing order for M 3 is J m+3, J 1, J 2. See igure 5.1 for an illustration, in which J 1 starts at time 0 on, starts at time x on, and starts at time B + 2x on M 3 ; J 2 starts at time B + x on, starts at time 2B + 4x on, and starts at time 2B + 5x on M 3 ; J m+3 starts at time 0 on M 3, starts at time B + 2x on, and starts at time 2B + 4x on. The feasibility is trivial and its makespan is Cmax π = 3q m+3 = 3B + 6x, suggesting the optimality. Conversely, if the optimal makespan for the constructed instance is Cmax = 3B + 6x = 3q m+3, then we will show next that S admits a partition into two equal parts. irstly, we see that the second machine processing the open-shop job J m+3 cannot be, since otherwise has to process all the jobs of before J m+3, leading to a makespan larger than 3B + 6x; the second machine processing the open-shop job J m+3 cannot be M 3 either, since otherwise M 3 has no room to process any job of before J m+3, leading to a makespan larger than 3B + 6x too. Therefore, the second machine processing the open-shop job J m+3 has to be, see igure 5.2 for an illustration. Denote the job subsets processed before and after the job J m+3 on as 1 and 2, respectively. Since x > B, neither of 1 and 2 may contain both J 1 and J 2, which have processing times x. It

11 Approximating M3 prpt C max 11 J 1 J 1 J 1 J 1 J 2 J 2 J 2 J 2 J m+3 J m+3 M 3 J m+3 J 1 J 1 J 2 J 2 0 B + 2x 2B + 4x 3B + 6x igure 5.1: A feasible schedule π for the constructed instance of the M3 prpt C max problem, when the set S can be partitioned into two equal parts S 1 and S 2. The partition of the flow-shop jobs {J 1, J 2 } is correspondingly constructed. In the schedule, the jobs of J 1 and the jobs of J 2 are processed in the LPT order, respectively. J m+3 J m M 3 J m+3 0 B + 2x 2B + 4x 3B + 6x igure 5.2: An illustration of an optimal schedule for the constructed instance of the M3 prpt, (n 1, 1) C max problem with O = {J m+3 } and q m+3 = B + 2x. Its makespan is C max = 3B + 6x = 3q m+3. follows that 1 and 2 each contains exactly one of J 1 and J 2, and subsequently P ( 1 ) = P ( 2 ) = B + x. Therefore, the jobs of 1 \ {J 1, J 2 } have a total processing time of exactly B, suggesting a subset of S sums to exactly B. This finishes the proof of the theorem. 6 An PTAS for M3 prpt, (n 1, 1) C max Recall that the PTAS designed in Theorem 3.7 designed for the general M3 prpt, (n 1, 1) C max problem when p 1 q n also works for the special case M3 prpt, (n 1, 1) C max when p 1 q n. In this section, we design another PTAS for the M3 prpt, (n 1, 1) C max problem when p 1 < q n, which is denoted as M3 prpt, (n 1, 1), p 1 < q n C max for simplicity. That is, we design a (1 + ɛ)-approximation algorithm C(ɛ) for this case, for any given ɛ > 0, and its running time is polynomial in both n and 1/ɛ. 6.1 A polynomial-time solvable case The following lemma states that if the total processing time of all the flow-shop jobs is no greater than q n, then we can easily construct an optimal schedule in linear time. Lemma 6.1 If P () q n, then the M3 prpt, (n 1, 1), p 1 < q n C max problem is solvable in linear time and C max = 3q n. Proof. In this case, we construct a permutation schedule π in which the job processing order for is,, where the jobs of are processed in the given (arbitrary, no need to be sorted)

12 12 Liu et al. /v:september 18, 2018 order; the job processing order for is, ; the job processing order for M 3 is,. As depicted in igure 4.2, the jobs of are processed consecutively on each machine, starting at time 0 on, starting at time q n on, and starting at time 2q n on M 3 ; the unique open-shop job starts at time 0 on M 3, starts at time q n on, and starts at time 2q n on. The schedule is feasible due to P () q n and its makespan is 3q n, and thus by Eq. (1) it is an optimal schedule. 6.2 Structural properties of optimal schedules We assume in the rest of the section that P () > q n, from which we conclude that contains at least two jobs. We explore the structural properties of optimal schedules for designing our approximation algorithm. Lemma 6.2 There exists an optimal schedule for the M3 prpt, (n 1, 1), p 1 < q n C max problem in which the open-shop job is processed on M 3 before it is processed on. Proof. We prove the lemma by construction. Suppose π is an optimal schedule in which the open-shop job is processed on M 3 after it is processed on. (That is, the machine order for is,,, M 3,, with exactly one of the s replaced by.) Recall that we use Sj i (Cj i, respectively) to denote the start (finish, respectively) time of the job J j on M i in π. Clearly, Cn 1 Sn. 3 We determine in the following two time points t 1 and t 2. We see that if is first processed on, i.e., Cn 2 Sn, 1 then all the jobs of processed before on can be finished on M 3 by time Cn, 2 due to p 1 < q n. It follows that if is first processed on, then we may assume without loss of generality that in π, there is no job J j such that Sj 3 < C2 n < Cj 3 (that is, the processing periods of J j on M 3 and of on intersect at time Cn). 2 If there is a job J j such that Sj 3 < S1 n < Cj 3 (that is, the processing periods of J j on M 3 and of on intersect at time Sn), 1 then we set t 1 = Sj 3, otherwise we set t 1 = Sn 1 (see igure 6.1 for an illustration). M 3 0 t 1 J j Y... X J k t 2 igure 6.1: Two possible values for t 1 : if the job J j as shown exists, then t 1 = S 3 j, otherwise t 1 = S 1 n; symmetrically, two possible values for t 2 : if the job J k as shown exists, then t 2 = C 1 k, otherwise t 2 = C 3 n. Symmetrically, if is last processed on, i.e., C 3 n S 2 n, then all the jobs of processed after on can be started on by time S 2 n, due to p 1 < q n. It follows that if is last processed on, then we may assume without loss of generality that in π, there is no job J k such that S 1 k < S2 n < C 1 k (that is, the processing periods of J k on and of on intersect at time S 2 n). If there is a job J k such that S 1 k < C3 n < C 1 k (that is, the processing periods of J k on and

13 Approximating M3 prpt C max 13 of on M 3 intersect at time C 3 n), then we set t 2 = C 1 k, otherwise we set t 2 = C 3 n (see igure 6.1 for an illustration). We note that both t 1 and t 2 are well defined. On the machine, the job is either finished before time t 1 (i.e., C 2 n t 1 ), or started after time t 2 (i.e., t 2 S 2 n), or is processed in between the closed time interval [C 1 n, S 3 n]. We thus obtain from π another schedule π by 1) moving the job subset X originally processed on in between the closed time interval [C 1 n, t 2 ] to start exactly q n time units earlier, while moving on to start at time t 2 q n, and 2) moving the job subset Y originally processed on M 3 in between the closed time interval [t 1, S 3 n] to start exactly q n time units later, while moving on M 3 to start at time t 1. See for an illustration in igure 6.2 and how it is obtained from the configuration in igure 6.1. The schedule π is feasible because the processing M 3 0 t 1 X J k... J j Y t 2 igure 6.2: The schedule π obtained from the optimal schedule π by swapping the job subset X processed on in between the closed time interval [C 1 n, t 2 ] with on, and swapping the job subset Y processed on M 3 in between the closed time interval [t 1, S 3 n] with on M 3. times of on all three machines are still non-overlapping and all the other jobs are flow-shop jobs which can only be processed earlier on and/or be processed later on M 3. Since no other job is moved, one clearly sees that π maintains the same makespan as π and thus π is also an optimal schedule, in which is processed on M 3 before it is processed on. This finishes the proof of the lemma. rom Lemma 6.2, we conclude that in the optimal schedules the machine order for is, M 3,,,, with exactly one of the s replaced by. The following two lemmas discuss the optimal schedules constrained to the machine order for. Lemma 6.3 If the machine order for is forced to be, M 3, or M 3,,, then an optimal schedule π can be constructed in O(n log n) time and its makespan is C max = 2q n + P (). Proof. We prove the lemma for the case where the machine order for is, M 3, ; the case where the machine order for is M 3,, can be almost identically argued, by swapping M 3 with. Let π be a schedule in which the machine order for is, M 3,. Since is last processed by while all the other jobs are flow-shop jobs, we may swap with the job subset processed on after, if any. This swapping certainly does not increase the makespan and thus we may assume that, on, all the jobs of are processed before. We prove the lemma by showing that Cmax π 2q n + P () and constructing a concrete schedule with its makespan equal to 2q n + P (). Recall that we use Sj i (Ci j, respectively) to denote the start (finish, respectively) time of the job J j on M i in π. Similar to the place where we determine the time point t 1 in the proof of Lemma 6.2, all the jobs of processed before on, which form the subset X, can be finished on M 3 by time Cn, 2 due to p 1 < q n. It follows that may immediately start on M 3 once it is finished on, i.e.,

14 14 Liu et al. /v:september 18, 2018 S 3 n = C 2 n, while all the other jobs of not processed by time C 2 n on M 3, which form the subset Y, can be moved to be processed after (in their original processing order on M 3 in π). See for an illustration in igure 6.3. Since X Y =, the completion time for the machine M 3 is at least P (X ) + q n + q n + P (Y) 2q n + P (), and consequently C π max 2q n + P (). M 3 0 X Y igure 6.3: The schedule π in which the machine order for is, M 3,. The job starts on M 3 immediately after it is finished on. The job subset X is processed on before and the job subset Y is processed on M 3 after. Clearly, if X = in the schedule π, and the jobs of are processed on all the three machines in the same LPT order, then the completion times for,, M 3 are max{p (), 2q n }+q n, q n +P (), and 2q n + P (), respectively. rom the fact that P () > q n, we conclude that the makespan of such a schedule is exactly 2q n + P (). Note that the schedule can be constructed in O(n log n) time. This finishes the proof of the lemma. Lemma 6.4 If the machine order for is forced to be M 3,,, and 1, 2 are the subsets of jobs forced to be processed before and after on the machine, respectively, then an optimal schedule π can be constructed in O(n log n) time and its makespan is C max = max{p i1 + P ( 1 ), q n } + q n + max{p i2 + P ( 2 ), q n }, where J i1 is the largest job of 1 and J i2 is the largest job of 2. Proof. Consider a feasible π in which the machine order for is M 3,,, and 1, 2 are the subsets of jobs processed before and after on, respectively. Similarly as in the proof of Lemma 6.3, since is last processed by while all the other jobs are flow-shop jobs, we may swap with the job subset processed on after, if any, in the schedule π. This swapping certainly does not increase the makespan and thus we may assume that, on, all the jobs of are processed before. Symmetrically, we may also assume that, on M 3, all the jobs of are processed after. Note that 1, 2 are the subsets of jobs processed before and after in the schedule π, respectively. We can also assume that 1, as otherwise we may swap to process first on and then to process on M 3 to convert π into a feasible schedule for the case where the machine order for is, M 3,. It follows that the optimal schedule π constructed in O(n log n) time in Lemma 6.3 serves, with its makespan Cmax = 2q n + p(). or the same reason, we can assume that 2. Since all the jobs of 1 have to be finished by time Sn 2 on the two machines and, we have Sn 2 p i1 + P ( 1 ). rom Cn 3 Sn, 2 we conclude that Sn 2 max{p i1 + P ( 1 ), q n }. Similarly, since all the jobs of 2 have to be processed on the two machines and M 3, and the earliest starting time is Cn, 2 the completion time for the machine M 3 is at least Cn 2 + p i2 + P ( 2 ). Note that the earliest starting time for processing on is Cn 2 too. It follows that the makespan of

15 Approximating M3 prpt C max 15 the schedule π is C π max max{p i1 + P ( 1 ), q n } + q n + max{p i2 + P ( 2 ), q n }. See for an illustration in igure 6.4. M J i1 1 J i1 J i2 2 J i2 2 J i1 1 J i2 2 igure 6.4: The schedule π in which the machine order for is M 3,,. The job subsets 1 and 2 are processed on before and after, respectively. The jobs J i1 and J i2 are the largest job of 1 and 2, respectively. On the other hand, if the jobs of 1 are processed in the same LPT order on and in the schedule π, then they are finished by time Sn 2 on and. urthermore, due to p i1 p 1 < q n, all the jobs of 1 can be processed in the same LPT order on M 3 and finished by time Cn. 2 Also, if the jobs of 2 are processed in the LPT order on following the jobs of 1, then by time Cn, 2 the job J i2 will be finished on and thus can start on at time Cn. 2 It follows that if the jobs of 2 are processed in the same LPT order on and M 3 in the schedule π, then they are finished by time Cn 2 + p i2 + P ( 2 ) on and M 3. This way, the makespan of such a schedule π is Cmax π = max{p i1 + P ( 1 ), q n } + q n + max{p i2 + P ( 2 ), q n }. Note that the schedule can be constructed in O(n log n) time. This finishes the proof of the lemma. 6.3 The (1 + ɛ)-approximation algorithm C(ɛ) rom each job J i with i > 1 and a bipartition {C, C} of the job subset {J i+1, J i+2,..., 1 }, we obtain a bipartition { 1, 2 } for where 1 = {J i } C and 2 = {J 1, J 2,..., J i 1 } C. Note that J i and J 1 are the largest job of 1 and 2, respectively. We may then use Lemma 6.4 to construct in O(n log n) time an optimal schedule π in which the machine order for is forced to be M 3,,, and 1, 2 are the subsets of jobs forced to be processed before and after on the machine, respectively. The makespan of π is C max = max{p i + P ( 1 ), q n } + q n + max{p 1 + P ( 2 ), q n } (see for an illustration in igure 6.4). Apparently the schedule π relies on the bipartition { 1, 2 } of, and it eventually relies on the job J i and the bipartition {C, C} of {J i+1, J i+2,..., 1 }. We therefore re-denote this schedule as π i (C, C). Let π i denote the best schedule among all π i (C, C) s, over all possible bipartitions {C, C} of {J i+1, J i+2,..., 1 }. Correspondingly, its makespan is denoted as C i max. When 2p i q n, we have 2p 1 q n too; consequently for any bipartition {C, C}, the makespan of the schedule π i (C, C) is p i + P ( 1 ) + q n + p 1 + P ( 2 ) = p i + p 1 + q n + P (). We thus have proved the following lemma. Lemma 6.5 When 2p i q n, the best schedule π i can be constructed in O(n log n) time and its makespan is C i max = p i + p 1 + q n + P (). When 2p i < q n, the best schedule π i or the best bipartition {C, C} of {J i+1, J i+2,..., 1 } is hard to locate (see Theorem 5.1). Nonetheless, we are able to design a fully polynomial time

16 16 Liu et al. /v:september 18, 2018 algorithm B(ɛ) that constructs a feasible schedule π i,ɛ with its makespan C i,ɛ max (1 + ɛ)c i max, for any ɛ > 0. The algorithm B(ɛ) first constructs an instance of the Knapsack problem, in which an item J j (j = i + 1, i + 2,..., n 1) corresponds to the job J j, has a profit p j and has a size p j, and the capacity of the knapsack is set to q n 2p i. It then calls an O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)})-time (1 + ɛ)-approximation for the Min-Knapsack problem to obtain a job subset C; and calls an O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)})- time (1 ɛ)-approximation for the Max-Knapsack problem to obtain another job subset D. In another O(n log n) time, the algorithm constructs two schedules π i (C, C) and π i (D, D), and returns the better one. A high-level description of the algorithm B(ɛ) is provided in igure 6.5. Algorithm B(ɛ): 1. Construct an instance of Knapsack, where an item J j corresponds to the job J j {J i+1, J i+2,..., 1 }; J j has a profit p j and a size p j ; the capacity of the knapsack is q n 2p i Run a (1 + ɛ)-approximation for Min-Knapsack to obtain a job subset C Use Lemma 6.4 to construct a schedule π i (C, C) Run a (1 ɛ)-approximation for Max-Knapsack to obtain a job subset D Use Lemma 6.4 to construct a schedule π i (D, D). 3. Output the better schedule between π i (C, C) and π i (D, D). igure 6.5: A high-level description of the algorithm B(ɛ). Lemma 6.6 When 2p i < q n, the algorithm B(ɛ) constructs a feasible schedule π i,ɛ in O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)}) time, with its makespan Cmax i,ɛ (1 + ɛ)cmax, i for any ɛ > 0. Proof. The running time of the algorithm B(ɛ) is dominated by the two calls to the approximation algorithms for the Knapsack problem, which are in O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)}). Let OPT 3 (OPT 4, respectively) denote the optimal solution to the Min-Knapsack (Max- Knapsack, respectively) problem and also abuse it to denote the total profit of the items in the solution. We therefore have the following (in-)equalities inside the algorithm B(ɛ): OPT 3 = min{p (X ) X {J i+1, J i+2,..., 1 }, P (X ) > q n 2p i }; (8) q n 2p i < P (C) (1 + ɛ)opt 3 ; (9) OPT 4 = max{p (Y) Y {J i+1, J i+2,..., 1 }, P (Y) q n 2p i }; (10) q n 2p i P (D) (1 ɛ)opt 4. (11) Let C 1 max (C 2 max, respectively) denote the makespan of the schedule π i (C, C) (π i (D, D), respec-

17 Approximating M3 prpt C max 17 tively). That is, Cmax 1 = 2p i + P (C) + q n + max p i p j + P (C), q n ; (12) j=1 Cmax 2 = 2q n + max p i p j + P (D), q n. (13) j=1 We distinguish three cases: In the first case where p 1 + i 1 j=1 p j + P (C) q n, Eq. (12) becomes i 1 Cmax 1 = 2p i + P (C) + q n + p 1 + p j + P (C) = p i + p 1 + P () + q n Cmax, i j=1 where the last inequality holds by Lemma 6.4, suggesting that the schedule π i (C, C) is optimal. In the second case where p 1 + i 1 j=1 p j + P (D) q n, Eq. (13) becomes C 2 max = 3q n C i max, where the last inequality holds by Lemma 6.4 again, suggesting that the schedule π i (D, D) is optimal. In the last case, we consider p 1 + i 1 j=1 p j + P (C) < q n and p 1 + i 1 j=1 p j + P (D) > q n. Assume in the best schedule π i the bipartition of {J i+1, J i+2,..., 1 } is {C, C }. If P (C ) > q n 2p i, i.e., C is a feasible solution to the constructed Min-Knapsack instance, then we have OPT 3 P (C ). (14) Using Eqs. (9, 14), Eq. (12) becomes C 1 max = 2q n + 2p i + P (C) 2q n + 2p i + (1 + ɛ)opt 3 2q n + 2p i + (1 + ɛ)p (C ) (1 + ɛ)(2q n + 2p i + P (C )) (1 + ɛ)c i max, (15) where the last inequality holds by Lemma 6.4 again. Otherwise we have P (C ) q n 2p i, i.e., C is a feasible solution to the constructed Max- Knapsack instance; we have OPT 4 P (C ). (16) Using Eqs. (11, 16), Eq. (13) becomes C 2 max = 2q n + p 1 + i 1 j=1 p j + P (D) = 2q n + p 1 + i 1 j=1 p j + P (C ) + P (D) P (C ) = 2q n + p 1 + i 1 j=1 p j + P (C ) + P (C ) P (D) 2q n + p 1 + i 1 j=1 p j + P (C ) + ɛp (C ) Cmax i + ɛp (C ) (1 + ɛ)cmax, i (17)

18 18 Liu et al. /v:september 18, 2018 where the second last inequality holds by Lemma 6.4 and the last inequality holds due to the trivial fact that P (C ) C i max. Therefore, in the last case, combining Eqs. (15, 17) we have C i,ɛ max = min{c 1 max, C 2 max} (1 + ɛ)c i max. This finishes the proof of the lemma. The (1 + ɛ)-approximation algorithm C(ɛ) for the M3 prpt, (n 1, 1) C max problem takes advantage of the (1 + ɛ)-approximation algorithm A(ɛ) for the M3 prpt C max problem when p 1 q l+1 (presented in Section 3, see Theorem 3.7), the optimal schedule constructed in Lemma 6.1 for the M3 prpt, (n 1, 1), p 1 < q n C max problem when P () q n, the two optimal schedules constructed in Lemma 6.3 for the M3 prpt, (n 1, 1), p 1 < q n C max problem when the machine order for is forced to be, M 3, or M 3,,, respectively, denoted as π 2,3,1 and π 3,1,2, respectively, and the n 2 schedules π i, i = 2, 3,..., n 1, constructed either in Lemma 6.5 or by the algorithm B(ɛ) (see Lemma 6.6). A high-level description of C(ɛ) is depicted in igure 6.6. Algorithm C(ɛ): 1. If p 1 q n, run the algorithm A(ɛ) and return the produced schedule. 2. If P () q n, use Lemma 6.1 to construct an optimal schedule π 0 and return π Use Lemma 6.3 to construct two optimal schedules π 2,3,1 and π 3,1,2. 4. or each i = 2, 3,..., n 1, 4.1. if 2p i > q n, use Lemma 6.5 to construct the schedule π i ; 4.2. otherwise run the algorithm B(ɛ) to construction a schedule π i,ɛ. Let π 3,2,1 denote the best schedule among these n 2 schedules. 5. Return the best schedule among π 2,3,1, π 3,1,2 and π 3,2,1. igure 6.6: A high-level description of the algorithm C(ɛ). Theorem 6.7 The algorithm C(ɛ) is an O(n 2 min{log n, log (1/ɛ)}+n/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)})- time (1 + ɛ)-approximation for the M3 prpt, (n 1, 1) C max problem. Proof. Recall from Theorem 3.7 that the running time of the algorithm A(ɛ) is in O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)}); Lemma 6.1 construct the optimal schedule π 0 in O(n) time; Lemma 6.3 constructs each of the two optimal schedules π 2,3,1 and π 3,1,2 in O(n log n) time; Lemma 6.5 construct the schedule π i also in O(n log n) time; and the running time of the algorithm B(ɛ) is in O(n min{log n, log (1/ɛ)} + 1/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)}). Since A(ɛ) is called only once, while B(ɛ) could be called for O(n) times, the overall time complexity of C(ɛ) is an order higher, that is, O(n 2 min{log n, log (1/ɛ)} + n/ɛ 2 log (1/ɛ) min{n, 1/ɛ log (1/ɛ)}). The worst-case performance ratio of 1 + ɛ is implied by Theorem 3.7 and Lemmas In more details, if p 1 q n, then the ratio is guaranteed by Theorem 3.7; if P () q n, then the

19 Approximating M3 prpt C max 19 optimality is guaranteed by Lemma 6.1; otherwise, Lemma 6.2 states that in the optimal schedule, the machine order for the unique open-shop job is one of, M 3,, M 3,,, and M 3,, : in the first two cases, the optimality is guaranteed by Lemma 6.3, while in the last case, the performance ratio is guaranteed by Lemmas 6.5 and 6.6 together. This finishes the proof of the theorem. 7 Concluding remarks In this paper, we studied the three-machine proportionate mixed shop problem M3 prpt C max. We presented first an PTAS for the case where p 1 q l+1 ; and then proposed a 4/3-approximation algorithm for the other case where p 1 < q l+1, for which we also showed that the performance ratio of 4/3 is asymptotically tight. The 3 prpt C max problem is polynomial-time solvable; we showed an interesting hardness result that adding only one open-shop job to the job set makes the problem NP-hard if the open-shop job is larger than any flow-shop job. The special case in which there is only one open-shop job is denoted as M3 prpt, (n 1, 1) C max. Lastly we proposed an PTAS for this special case M3 prpt, (n 1, 1) C max. We believe that when p 1 < q l+1, the M3 prpt C max problem can be better approximated than 4/3, and an PTAS is perhaps possible. Our last PTAS for the special case M3 prpt, (n 1, 1) C max can be considered as the first successful step towards such an PTAS. Acknowledgement LL was supported by the China Scholarship Council Grant and the undamental Research unds for the Central Universities Grant No YC and AZ were supported by the NSC Grants and ; YC was also supported by the China Scholarship Council Grant JD was supported by the NSC Grant and the Zhejiang Provincial Natural Science oundation Grant No. LY18A RG and GL were supported by the NSERC Canada; LL, YC, JD, GN and AZ were all partially supported by the NSERC Canada during their visits to Edmonton. GN was supported by the NSC Grant , the Natural Science oundation of ujian Province Grant 2016J01332 and the Education Department of ujian Province. References [1] P. Brucker. Scheduling algorithms. Berlin: Springer, [2]. Y. Chin and L. L. Tsai. On j-maximal and j-minimal flow shop schedules. Journal of the ACM, 28: , [3] M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-completeness. W. H. reeman and Company, San rancisco, [4] R. L. Graham, E. L. Lawler, J. K. Lenstra, and R. Kan. Optimization and approximation in deterministic sequencing and scheduling: A survey. Annuals of Discrete Mathematics, 5: , [5] H. Kellerer and U. Pferschy. Improved dynamic programming in connection with an PTAS for the knapsack problem. Journal of Combinatorial Optimization, 8:5 11, 2004.

20 20 Liu et al. /v:september 18, 2018 [6] H. Kellerer, U. Pferschy, and D. Pisinger. Knapsack Problems. Springer, Berlin, [7] C. Koulamas and G. J. Kyparisis. The three-machine proportionate open shop and mixed shop minimum makespan problems. European Journal of Operational Research, 243:70 74, [8] C. Y. Liu and R. L. Bulfin. Scheduling ordered open shops. Computers & Operations Research, 14: , [9] T. Masuda, H. Ishii, and T. Nishida. The mixed shop scheduling problem. Discrete Applied Mathematics, 11: , [10] P. S. Ow. ocused scheduling in proportionate flowshops. Management Science, 31: , [11] S.S. Panwalkar, M. L. Smith, and C. Koulamas. Review of the ordered and proportionate flow shop scheduling research. Naval Research Logistics, 60:46 55, [12] N. V. Shakhlevich and Y. N. Sotskov. Scheduling two jobs with fixed and nonfixed routes. Computing, 52:17 30, [13] N. V. Shakhlevich, Y. N. Sotskov, and. Werner. Complexity of mixed shop scheduling problems: A survey. European Journal of Operational Research, 120: , [14] N.V. Shakhlevich, Y. N. Sotskov, and. Werner. Shop-scheduling problems with fixed and non-fixed machine orders of the jobs. Annals of Operations Research, 92: , [15] V. A. Strusevich. Two-machine super-shop scheduling problem. The Journal of the Operational Research Society, 42: , 1991.