Published on

02-Jul-2016View

212Download

0

Transcript

ELSEVIER Operations Research Letters 21 (1997) 69-76

Single machine scheduling with discretely controllable processing times

Zhi -Long Chen a'*, Q ing Lu b, Guochun Tang c

aDepartment of Civil Engineering & Operations Research, Princeton University, Princeton, NJ 08544, USA b Department of Decision Sciences, National University of Singapore, 10 Kent Ridge Crescent, Singapore 0511, Singapore

c Department of Management, Shanghai Second Polytechnic University, Shanghai 200041, People's Republic of China

Received 1 April 1995; revised 1 December 1996

Abstract

In the field of machine scheduling problems with controllable processing times, it is often assumed that the possible processing time of a job can be continuously controlled, i.e. it can be any number in a given interval. In this paper, we introduce a discrete model in which job processing times are discretely controllable, i.e. the possible processing time of a job can only be controlled to be some specified lengths. Under this discrete model, we consider a class of single machine problems with the objective of minimizing the sum of the total processing cost and the cost measured by a standard criterion. We investigate most common criteria, e.g. makespan, maximum tardiness, total completion time, weighted number of tardy jobs, and total earliness-tardiness penalty. The computational complexity of each problem is analyzed, and pseudo-polynomial dynamic programming algorithms are proposed for hard problems. 1997 Elsevier Science B.V.

Keywords: Single machine scheduling; Discretely controllable processing times; NP-hardness; Dynamic programming

I. Introduction

Scheduling problems with controllable process- ing times have received increasing attention during the last decade. Work on this area was initiated by Vickson [18, 19], and Van Wassenhove and Baker [17]. For a survey of this area till 1990, the reader is referred to [13]. Recent results include [20, 14, 2,16, 4-7].

It was assumed in most of the above papers that the processing time of a job is continuously con- trollable from its normal length to its minimum

* Corresponding author.

possible length, i.e. the actual processing time of a job can be controlled to be any number in a speci- fic interval. This assumption is valid for some prac- tical situations. For example, in a job shop environ- ment where the processing of jobs is controlled by the machine speed which depends on the amount of a divisible resource (e.g. electricity) allocated and hence can vary continuously from its normal speed to its maximum possible speed, the possible pro- cessing time of a job can be controlled to be any number in the interval corresponding to the speed interval of the machine. While in another job shop where the machine consists of a main facility and several auxiliary facilities (e.g. support devices) and its speed is determined by the pattern of the

0167-6377/97/$17.00 1997 Elsevier Science B.V. All rights reserved PII S0 1 67-63 77(97)000 1 0-2

70 Z-L. Chen et al. / Operations Research Letters 21 (1997) 69-76

auxiliary facilities used, the machine can only have a finite number of speeds because there are a lim- ited number of possible patterns of the auxiliary facilities. Under such a situation, the possible pro- cessing time of a job can only be one of a few specified lengths, for instance, two possible lengths with respect to the case where the auxiliary facilities have two patterns. For another example, in a manufacturing system where job processing times depend on the units of an indivisible resource (e.g. labor) allocated, job processing times can only have finitely many (up to the maximum units of the resource) possibilities. The continuous model used in the above-cited literature does not fit these situ- ations where processing times are only discretely controllable. The aim of this paper is to introduce a discrete model to handle these situations.

The discrete model can be formally described as follows. We are given a set of jobs N = { 1, 2 . . . . . n} to be processed on a machine. Each job j ~ N has k (k ~> 1 and given) possible processing times: aj l , aj2, ... ,ajk with aj l 2> aj2 > ... > ajk, where ajl is called the normal processing time and ajk the minimum possible processing time of job j. A pro- cessing cost cji is associated with each possible processing time aji (i = 1, 2,.. . ,k). It is assumed that 0 ~< c j l < cj2 < ... < Cjk. This assumption is reasonable since more resource (e.g. a faster speed of the machine, or more units of labor) and hence higher costs are needed to achieve smaller process- ing times. Let xj denote the actual processing time of job j in a schedule; then x~ must be one of the possible processing times associated with job j, i.e. xj ~ {a~, a j2 . . . . , a~k}. A processing cost c~i is incur- red if the actual processing time of job j is xj = aj~.

To the best of our knowledge, there is only one published paper [7] dealing with the discrete model. Daniels and Mazzola [7] considered an NP-hard flow shop scheduling problem where job processing times are discretely controllable through the allocation of a limited amount of re- source and the objective is to minimize the make- span. They gave an optimal algorithm and a heuristic for solving the problem.

In this paper, we investigate a class of single machine scheduling problems with controllable processing times under the discrete model. The objective of each problem is to minimize the sum of

the total processing cost and the cost measured by a standard criterion. We examine each of the com- mon criteria: makespan, maximum tardiness, total completion time, weighted number of tardy jobs, and total earliness-tardiness penalty. We clarify the complexity of each problem.

To facilitate the presentation, we define the fol- lowing notation: r~ ready time of job j dj due date of job j Cj completion time of job j in a given schedule Ej max(0,d~ - Ci), earliness of job j in a given

schedule Tj max(0, C~ - dj), tardiness of job j in a given

schedule U~ 1 if Cj > dj and 0 otherwise c~ unit cost of earliness /3 unit cost of tardiness wj cost if job j is tardy, i.e. Cj > d~ l j i (x j ) 1 if xj = a~i and 0 otherwise (i = 1, 2 ... . , k) TPC Y~j~N~=lcjilji(x~), total processing cost in

a given schedule C~,ax maxiEN{Cj}, makespan of a given schedule T~,ax max~N {T~}, maximum tardiness of a given

schedule Suppose that all the parameters are nonnegative

integers and any preemption is not allowed when processing. For ease of problem formulation, we use the three-field problem classification "1"[" (in- troduced by Graham et al. I-9]) to define a machine scheduling problem. The first field represents the machine environment, the second field describes the scheduling constraints, and the third field gives the objective function. We use the following nota- tion in the second field. "rj" indicates that job ready times are not all equal, "d~ = d" indicates that jobs have a common due date, and "dj = d > D" indi- cates that jobs have a common and unrestrictively large due date (e.g. d r - d >1 ~ j~N aj l ) where D rep- resents a large enough number and has no other meaning. See 1-3] for a discussion of "restrictive" and "unrestrictive" versions of common due date scheduling problems. Finally, "cm" and "dm" de- note the models where processing times are, respec- tively, continuously controllable and discretely controllable.

Now, we can state the seven problems we con- sider in this paper as follows, where P4' is the

Z-L. Chen et al. / Operations Research Letters 21 (1997) 69-76 71

special case of P4 with dj = d, and P5' is the special case of P5 with dj = d and wj = w. (P1) lldmly~Cj + TPC. (P2) 1 I dj = d > D, dmla ~ Ej + fl y~ Tj + TPC. (P3) 1 I rj, dml Cm,x + TPC. (P4) 1 I dml Tmax + TPC. (P4') l id j - d, dmlTm~x + TPC. (P5) l ldmlywiU j + TPC. (P5') l[dj - d, dmlwE U s + TPC. Note that, through a suitable scaling of the process- ing costs, the objective functions in the above prob- lems are capable of capturing the trade-off between traditional scheduling performance measures and expenditure on resources. In Section 2, we show that P1 and P2 are polynomially solvable. In Section 3, we prove that P3,P4' and P5' are all NP-hard. We then give pseudopolynomial dynamic programming algorithms in Section 4 for solving the hard problems P3, P4 and P5. This implies that P3,P4,P4' ,P5 and P5' are all NP-hard in the ordinary sense. Note that in the following sections, when we refer to problems with controllable processing times under the continuous model, as often assumed in the literature, we assume that the total processing cost TPC = ~j~N fj(xj) with all f~(. ) being nonincreasing linear functions.

2. Polynomial solvability of P1 and P2

The problems l l l~wjCj and l[di==-d>D] ~ Ej + fl~ Tj (with uncontrollable processing

times), and the problems 1 ]cml 5~ Cj and 1 I dj = d > D, cmlct~Ej + fly. T i (with continuously con- trollable processing times) are all polynomially solvable. It is well-known that the weighted- smallest-processing-time-first (WSPT) sequence is optimal for 1 I[~wiC j. For the other three prob- lems, see 1-15, 18, 14], respectively. The complexity of the problem: l lcmlEwiCj + TPC remains open 1-13] despite that its uncontrollable version: l ll~wjCj is extremely easy. When the given common due date is restrictive, the problem l l d j -d lc t~E j+f l~T j is NP-hard [10], and hence the problems 1 ]dj = d, cm ]ot ~, Ej + f l~T j+ TPC and l ld j=d, dmlctZEj+ f lZT j+ TPC are also NP-hard.

In this section, we show that both problem PI: ltdmp~ Cj + TPC, and problem P2: lrdj - d > D, dml~ K Ej + fl 5" Tj + TPC are polynomially solv- able by formulating them as n n assignment prob- lems.

2.1. Assignment problem formulation for P1

Consider problem PI: 1 IdmlY, Cj + TPC. There are n positions in the schedule, each of which must contain exactly one job. If job l is scheduled at position m in a schedule, then the contribution made by job l to the total objective function is

k

fi,,(xt) = (n - m + 1)xt + ~ ctilli(xt), i=1

where the first term is the contribution to the total completion time and the second term is that to the total processing cost. To minimize f/,.(xt), one should set the actual processing time of job l, xt, to be all for some 1 ~< i ~< k such that f tm(au) = Vt,n, where

vt,, = min {ftm(Xl) lXt e {a, . . . . . alk } }

= min {(n - -m+l )a t j+cz j} . (1) j~{1 .. . . . k}

It follows that the cost incurred by assigning job l at position m, and then selecting its actual pro- cessing time optimally, is exactly v~,, given by (1), independent of the other jobs. This permits us to formulate problem P1 as the following n x n assign- ment problem:

/=1 m=l

~ytm= 1, m=1,2 . . . . . n, (3) /=1

Ylr,=l l-----1,2, ... ,n, (4) m=l

where

Y~,, = {~ if job 1 is scheduled at position m't5~tJ otherwise

and vt,,, are given by (1).

72 Z-L. Chen et al. / Operations Research Letters 21 (1997) 69 76

The above n n assignment problem can be sol- ved in an O(n 3) time (see, e.g., [1]). Hence, we have shown that P1 can be solved in an O(n 3) time. However, we do not known whether its weighted version: l ldml~ w~ Cj + TPC is polynomially solv- able or NP-hard. We conjecture that this problem is NP-hard (even when k = 2).

2.2. Assignment problem formulation for P2

Clearly, in any optimal schedule for problem P2: 1 I d~ = d > D, dm [ ~ Y, Ej + fl ~ Tj + TPC, there is no inserted idle time between jobs. (However, idle time at the beginning of the schedule may exist). The following lemma states an important property about the optimal solution for the problem.

Lemma 1. There exists an optimal schedule for P2 such that the hth job completes at common due date d, where h = F nfl/(7 + fl) ~, and F x qdenotes the least integer no less than x.

The first terms in (6) and (7) are the contributions, respectively, to the total earliness penalty and to the total tardiness penalty. The second terms in (6) and (7) are the contributions to the total processing cost. Define

min j~{l ..... k}

L'lm rain

j~{1 ..... k}

{(m - 1)ctat; + clj} if m ~< h,

{ (n -m+ 1)[3at~+clj} if m>h.

(8)

Then to minimize the contribution of job l (if it is at position m), one should set xt = au for some 1 ~< i ~< k such that f~,,(a,) = V~m. By the same argu- ment as in Section 2.1, we can formulate P2 as the assignment problem defined by (2)-(5) except that V~m is now given by (8). Hence, problem P2 can be solved in an O(n 3) time.

3. NP-hardness of P3, P4' and P5'

Proof. In the case with uncontrollable processing times, Panwalker et al. [15] proved this property for the problem I [d~==-d>DIe~E~+~yTj . Since the value of h is independent of the job processing times, this property is also valid for P2. []

Thus, there exists an optimal schedule for P2 in which exactly h = [- nfl/(~ + fl)~ jobs are either early or completed at the due date and n - h jobs are tardy. As in Section 2.1, we assume that there are exactly n positions on the machine, numbered 1,2 .. . . ,n. Suppose that in a schedule, job l is scheduled at position m. There are two cases. If m ~< h, then job I is early and the contribution made by it is

k

f//,.(xl) = (m - 1)~xl + ~" culli(Xl). (6) i=1

If m > h, then job l is tardy and the contribution made by it is

k

fl,,(x,) = (n - m + 1)fix, + ~ c,illi(x,). (7) i=1

In the case with uncontrollable processing times, it is well known that the smallest-ready-time-first (SRT) schedule, and the earliest-due-date-first (EDD) schedule are optimal, respectively, for the problem l l r j lCma x and for the problem l i lT . . . . and the problem 1 ]l ~ Uj is polynomially solvable by Moore's algorithm [12]. But the weighted ver- sion of the last problem: 1 ]l S wJ Uj is NP-bard even with a common due date [11].

In the case with continuously controllable pro- cessing times, the problems 1 I r j, cml Cmax + TPC, and 1 Icml Tm,x + TPC are both polynomially solv- able (see [13, 19], respectively), while the problem 1 I cml Y. wj Uj + TPC becomes NP-hard even with dj - d and w I = w [-6].

Surprisingly, under the discrete model, problems P3: llrj, dm[Cm,x + TPC and P4: lldmLTm,x + TPC become NP-hard even with dj - d for P4. The NP-hardness proofs for P3 and P4' (i.e. P4 with d r = d) are given in this section. Also, problem P5': lldj = d, dmlwY~ U i + TPC (a special case of P5) is proved to the NP-hard here. Note that when jobs are simultaneously available, the problem l ldmlCmax + TPC can be easily solved by se- quencing jobs in an arbitrary order and letting

Z-L. Chen et al. / Operations Research Letters 21 (1997) 69-76 73

Xj = aji where i is the index such that aji + cji = min{ajt + cjtll = 1, ... ,k}.

For all proofs, we use the NP-complete problem: KNAPSACK (see, e.g., [8]) for a reduction. An instance of KNAPSACK problem can be stated as follows:

KNAPSACK: Given a set S = {1,2 . . . . . s} of s objects, a cost...