You are on page 1of 12

A Class of Single Machine Central Tendency-Dispersion Bicriteria Problems Author(s): John Mittenthal, M.

Raghavachari and Arif Rana Reviewed work(s): Source: The Journal of the Operational Research Society, Vol. 47, No. 11 (Nov., 1996), pp. 1355-1365 Published by: Palgrave Macmillan Journals on behalf of the Operational Research Society Stable URL: http://www.jstor.org/stable/3010201 . Accessed: 15/10/2012 06:02
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Palgrave Macmillan Journals and Operational Research Society are collaborating with JSTOR to digitize, preserve and extend access to The Journal of the Operational Research Society.

http://www.jstor.org

Journal of the Operational Research Society (1996) 47, 1355-1365

? 1996 OperationalResearch Society Ltd.All rights reserved 0160-5682196 $12.00

A Class of Single Machine CentralTendencyDispersion BicriteriaProblems


JOHN MITTENTHAL1, M. RAGHAVACHARI2 and ARIF RANA3 Institute, USA and Polytechnic USA, 2Rensselaer at Tuscaloosa, of Alabama 'The University Sciences,Pakistan of Management University 3Lahore bicriteriapenalty functionsof completiontimes on a single In this paper, a numberof non-regular is a measureof centraltendency(CT) of completion machineare studied.In each function,one criterion (DSP). times, for example,mean or median,and the otheris a measureof completiontime dispersion of optimal Five differentDSP measuresare consideredin the paper.We discuss the characterizations for the pure sequences for some new functions and summarizeprevious results when appropriate in the context of of CT and DSP. Further, bicriteriaproblemas well as for the convex combination of variance is madebetweenthe two dispersion measures a comparison measure, minimizinga dispersion of completiontimes andtotal absolutedeviationsof completiontimes (TADC). multi-objective Key words: sequencing,

INTRODUCTION Machine scheduling has been a major area of research for more than four decades. Single machine schedulingforms the basic model from which more complex multi-machinemodels are developed. Most of the early research in single machine scheduling has been in the context of a single criterion, for example, average completion time, total tardiness,numberof tardyjobs, variance of completion times, to deal with a bicriteriaproblem;specifically,the sum of the etc. Smith' was among the firstresearchers job completiontimes and the numberof tardyjobs. In the past decade there has been a growing interest in other single machine bicriteria scheduling problems. A number of review articles on multiple criteria scheduling research summarizethis work. Dileepan and Sen2 address single machine, bicriteriaproblems involving regular, or non-decreasing, problems functionsof completiontime. Fry,Armstrong,and Lewis' cover single machine, multi-criteria that may involve both regularand non-regularfunctions of completion time. A non-regularfunction of completiontimes has the potentialto decreasewhen the completion time increases, for example, mean earliness.All the non-regularfunctionsmentionedin their survey involve job due dates. Fry et al.3 note that even though non-regularpenalty functions are becoming more important with the increasing acceptanceof the JITphilosophy,thereexists a gap in the areaof bicriteriaresearchwhen one criterionis Nagar, Haddock and Heragu4address single and multiple machine, multiple and bicriteria non-regular. schedulingproblems. However,the problems and results surveyed are restrictedto regularfunctions of completion time. In this paper, we consider bicriteria sequencing problems involving a measure of central tendency (CT) ofjob completiontimes and a measureof dispersion(DSP) of job completiontimes. A CT measure is usually a regularfunction while a DSP measure is, in general, a non-regularfunction. CT measures the median of completion times. DSP measures include C, the mean of completion times, and Cmed, include VAR, TADC, ADMean, ADMedian and R. These are defined subsequentlyin the paper. VAR and TADC have merited attention on their own as evidenced by their inclusion in the Baker and Scudder'review paperon early/tardymeasuresinvolvingjob due dates. However,VAR, TADC and all of the other measures considered here are independent of job due dates. Thus, the CT-DSP problems consideredherebegin to fill the gap mentionedby Fry et al.3 and have not yet been consideredfully in the literature.

Correspondence: J. Mittenthal, Department of Management Science and Statistics, The University of Alabama at Tuscaloosa, Tuscaloosa, AL 35487-0226, USA

1356

Journalof the OperationalResearchSociety Vol. 47, No. II

Generally, the minimization of a DSP measure is more difficult than the minimization of a CT measure. For example, it is known that VAR minimization is NP hard and TADC minimization is polynomial in complexity. See Kubiak6for the formerresult and Bagchi7for the latter.Of course, both CT measures consideredhere can be minimized in O(n log n) time by sequencing the jobs in shortest processing time (SPT) order. In bicriteriaCT-DSP minimizationprpblems,a good sequence is one in which both the CT measure and the DSP measure are low. Thus, a pure bicriteriaminimization method based on the concept of Paretooptimal solutions is a naturalapproachto take. Anothercommon approachto bicriteriaproblems is the minimizationof a convex combinationof the criteria.These are the two modelling and solution methods consideredhere. As noted by Bagchi7,functionsthat simultaneouslyminimize a measureof CT and a measureof DSP need to be consideredwhen it is desiredto minimize averagein-processinventoryas well as the variation in the quality of service providedto differentjobs. The minimizationof averagein-process inventoryis equivalentto minimizing mean completion time, mean waiting time, mean time in system, or mean lateness. It is well known that an SPT sequence minimizes the mean completion time. The second objectiveof variationcontrol,namely ensuringthatwork for all customersbe completedas close to some targetsas possible, has assumed greaterimportancewith the increasingpopularityof concepts like JIT (Just-In-Time) scheduling.Forexample,Mertenand Muller8 note thatin organizingcomputerfile systems with large data sets it is desirableto provide uniform response time to users. The objective is to find a sequencethatminimizes the variationof access times to differentrecordsin the file. Anothersituationin which one may wish to have low variationof completiontimes arises in problemsinvolvingjobs having a common due date d. In such problems, the objective of minimizing EInZ(Ci - d)2 penalizes larger deviations from the common due date more than smaller deviations. This type of penalty structureis useful for two reasons. First,it balances inventorycharges,incurredon jobs completedpriorto their due date, againstchargesfor expeditingdeliveriesof jobs completedaftertheirdue dates. Second,the penalty structure discouragesexcessively early or late job completions.Further, Vani and Raghavachari9 showed that minimizing XIn=(Ci - d)2 with d as a decision variable is equivalentto minimizing VAR. It is thereforedesirableto obtain a sequence that has low mean and low variationof completion times. The convex combination of the CT and DSP measures is a natural function to minimize in the bicriteria problem. The convex combination objective also arises in certain single machine, single criterionstochastic sequencingproblems. Mittenthaland Raghavachari'0 consider the minimizationof a deviationof completiontimes from a common due datewhen the machineis subjectto random quadratic that occur accordingto a Poisson process. They show that the deterministic breakdowns equivalentof the problemreduces to the minimizationof a particularconvex combinationof C and VAR. An example of a CT-DSP function in the literaturearises from Bagchi7who seeks simultaneouslyto minimize the mean (a regularfunction) and variance(a non-regularfunction) of completiontimes. This paper considers a convex combinationof the Total Absolute Differences of Completiontimes (TADC) and the mean. TADC was first suggested by Kanet" as an alternatemeasure of variation within a scheduling context. Anotherexample is containedin De, Ghosh and Wells'2,where they give a pseudopolynomial algorithmto minimize a convex combinationof mean and varianceof completiontimes. The next section considers differentcombinationsof measuresof CT and DSP and gives a summary of the previous work, as well as some new results. In the following section, a detailed comparison between minimizing TADC and minimizing VAR is made.

CT-DSP FUNCTIONS Henceforth,the following notation is used: n Number of jobs.


Pm Processing time of the mth longest job (m> P2 > . > p,n), m = 1, 2, Cm Completion time of the mth longest job, m = 1, 2,...,n.
.

n.

P[m]Processing time of the job in position m of a sequence. C[mj Completion time of the job in position m of a sequence.

J.Mittenthal et a/.-Single MachineBicriteriaProblems

1357

C Mean completion time C =nE


CMed
-

n1

n m=1

Cmi].

Completion time of the medianjob.


Cmed
=

C[(n+l)/2]

when n is odd when n is even

| (C[n/2] + C[(n/2)+1])/2

The DSP measures used are: (i) Variance (VAR) VAR =-E
in

n m=1

(Cm

C)

(ii)Total absolute deviation from the mean (ADMean) ADMean =


n E lCm-Cl m=1

(iii)Total absolute deviations of completion times (TADC)


n n

TADC = L

{Cm-Ckl

k=1 m=k

(iv)Total absolute deviation from the median (ADMedian)


n

ADMedian =
m=1

CCm-Cmed

and (v) Range (R) of completion times


R= C[n] C[1]

The CT measures used are mean and median. Throughoutthe paper it is assumed that all jobs are available at time zero. It is also assumed that all processing is to be done on a single machine and preemptionis not allowed. Ten differentfunctionsarise from all possible combinationsof the differentmeasuresof CT and DSP considered.For most of these problems, results for a convex combinationof the two criteria,namely min XCT+(1- {)DSP,

O<

i <

are presented.Note that, in a bicriteria scheduling problem, the objective is ideally to find the set of sequencesthat are betterthan all possible sequences in at least one criterion.Such a set is also called the non-dominated(or efficient) set. The pure bicriteriaobjective function can be expressed as min(CT, DSP) Even though both Bagchi7and De, Ghosh and Wells12 use a convex combinationof the two objectives, De, Ghosh and Wells'3 note that many efficient points are missed when using such a formulation. However,CT-DSPfunctionsby their very naturehave a large numberof non-dominatedsequences that and Bagchi7).Thus it appears grow exponentiallywith the numberof jobs (see Eilon and Chowdhury"4 that the efficient sequencesmissed by a convex combinationobjective may not result in a great loss of information,since there are likely to be other sequences on the convex combinationfrontierthat have criteriavalues close to those of the missing sequences. Nevertheless, for some of the problemsthat are more mathematically tractable,results for the non-dominatedset are also presented. A quick review with some new results follow (for a summarysee Table 1).

1358 Journalof the OperationalResearchSociety Vol. 47, No. II


TABLE

1(a). Summaryof the bicriteriafamily offunctions (convex combinations)


Measuresof centraltendency
C Cmed

Measuresof Dispersion VAR

ADMean TADC ADMedian R


a

Consideredby De et al. 12convex combinationof c2 and VAR proposed by Mittenthal,Raghavachariand Rana15.Also consideredby Mittenthaland Raghavachari1 in a different Qontext where units match. All cases have V-shaped optimal schedules Unknown Consideredby Bagchi7 Multiple optimal V-shaped solutions Polynomial algorithm Not investigated Multiple V-shaped optimal sequencesa

Not investigated

Not investigated V-shaped optimal sequences' V-shaped optimal


sequences'

Multiple V-shaped optimal


sequences'

Results shown in this paper. TABLE 1(b). Summaryof the bicriteriafamily offunctions (pure bicriteria)

Measureof
dispersion C

Measuresof centraltendency
Cmed

TADC ADMedian R
a

Not every non-dominatedsequence is necessarily V-shaped Not investigated n V-shaped, non-dominatedsequencesa

V-shaped, non-dominated
sequences'

V-shaped, non-dominated
sequences'

Ln/21 + 1 V-shaped, nondominatedsequencesa

Results shown in this paper.

C and VAR Here the objective is to minimize


n i
2_nC2

+( (-A)E

n E

This convex combinationmodel was consideredby De, Ghosh and Wells12. They prove that the optimal sequence is V-shaped and present an algorithmwith pseudo-polynomialtime complexity. A V-shaped sequencehas jobs in non-increasingorderof processingtimes before the job with the shortestprocessing time, and in non-decreasingorderof processingtimes afterit. VAR is measuredin squaredunits and C is measuredin originalunits. Mittenthal,Raghavachari and Rana"5 point out thatthe units of the two criteria do not match,possibly leading to differentoptimalsequenceswhen the processingtimes are measuredin either seconds or minutes. Even more is true. Suppose thatPi ,P2 .... ,Pn are the processing times of the jobs. Consideran equivalentproblemwhere the processing times are changed by a factor-, namely the new processing times are TP1,TP2, .... ,-uPn. For these processing times, the equivalent objective is to
minimize

n
A
C12nC2 )(1-)Z

n
Cl

(1)

It follows from Theorem 2 of Mittenthaland Raghavachari'0 that for all sufficiently small -, an SPT In the minimizes bicriteria sequence Equation (1). contrast, pure problem remains unchanged by the in the times. change processing To remedy this problem, Mittenthal, Raghavachariand Rana"5 propose a convex combination of squareof the mean completiontime and VAR as an alternatefunction.They show that the new function also has V-shapedsequences and can be solved using a slight modificationof De, Ghosh and Wells12 pseudo-polynomialalgorithm.
C and ADMean

We do not have any complete results to reportfor this CT-DSP problem.

et al.-Single J.Mittenthal

MachineBicriteriaProblems

1359

C and TADC The convex combinationmodel of this function was consideredby Bagchi7.It has multipleV-shaped optimal sequences. Bagchi7 also gives a polynomial algorithmto solve the problem. However,the pure bicriteria version of the problemmay have non-dominatedsequencesthat are not V-shaped.For example, it can be verified that the sequence (given in terms of processing times) 6 9 5 11

is non-dominated(C = 18 and TADC = 80) even though it is not V-shaped.

Cmed and ADMedian

Lemma 1 Minimizing ADMedian is equivalentto minimizing TADC. Proof It can be verified that ADMedian can be rewrittenas n s3 min{n-m + 1, m-1 }p[m]

m=l
n

while TADCcan be rewritten as m=1


The

E (n- m+1)(m- I)p[m]

positional weights of the two functions are (0, 1, 2, .. ., 2, 1) and (0, n - 1, 2(n - 2), .. ., (n - 2)2, n - 1), respectively.It is well known that if the positionalweights of an of the processingtimes then that functioncan be minimized objectivefunctionare independent by the jobs suchthatthe longestjob is in the positionwith the smallestweight,the second sequencing and longestjob is in the positionwith the next smallestweight,and so on (see Hardy,Littlewood and andTADC haveexactlythe sameranking, Polya16).Sincethepositional weightsof bothADMedian these the same set of are independent of the processing functions are minimized times, by sequences. An important of Lemma1 is thatanysequence ADMedian will minimize thatminimizes implication for instance to conclude thatthe purebicriteria CT-DSP TADCandvice versa.However, problem (C, musthave the same ADMedian) has the sameParetooptimalsolutionsas (C, TADC),all sequences relativerankings and TADC.This is not the case in general.For with respectto both ADMedian the times 4, 6, 9, and 11. For (C, ADMedian), example,considera 4-job problemwith processing while the sequence 11-4-6-9 is sequence4-9-6-11, given by processingtimes, is non-dominated Incontrast, dominated. for(C, TADC)4-9-6-11is dominated and11-4-6-9is non-dominated. Hence,the sets of Paretooptimalsolutions for thesetwo problems cannotbe provento be equal. If we allowd to be an unconstrained decision the common duedateof alljobs, variable representing thend = Cmed minimizes the function
n

f(Cl, C2,...,@cn, d) =E

iCm- di m=1

(2)

andBagchi,Sullivan andChang'9). Hall'8 thisfunction is alsoequivalent (see Kanet'7, Thus,minimizing to minimizingADMedian. Therefore the algorithm a set of optimal used by Hall'8for generating for (2) canbe usedforgenerating forTADCandADMedian sequences optimal sequences (fromLemma thatone of the optimal 1). Note further sequences generated by Hall'salgorithm,
P1 ,P3,Ps,
* * .* ,4P

was suggested by Eilon and Chowdhury'4as a heuristicfor minimizing VAR. We next prove that the pure bicriteriaproblem min(Cmed, ADMedian) has V-shapednon-dominated sequences. See the Appendix for details.

1360 Journalof the OperationalResearchSociety Vol.47, No. II

TheoremI For the bicriteriapenalty function, min(Cmed, ADMedian) the non-dominated(Pareto optimal) sequences are V-shaped. (3)

Remarks 1. Note that for any E [0, 1], one of these non-dominated,V-shaped sequences is optimal for the convex combinationversion of the (Cmed, ADMedian) problem. 2. The proof of Theorem 1 also applies to the pure bicriteriaproblem min(Cmed, TADC) since, as pointed out in Lemma 1, the positional weights of both ADMedian and TADC have exactly the same ranking. Hence, we conclude that min(Cmed, TADC) has V-shaped, non-dominated sequences,while the convex combinationversion of the problemhas V-shapedoptimal sequences. An algorithmfor generatingthe set of non-dominatedsequences for Equation(3) is given below. In the algorithm,B is used to representthe set of jobs whose completiontime occurs before Cmed and A is used to representthe set of jobs whose completion times occur after Cmed. Step 1. Let SI PI, P3, P5, -- P6, P4, P2- Among all the optimal sequences obtained from Hall's algorithm,SI has the largestCmed since all pairs ofjobs with the same positionalweight have the job with the shorterprocessing time in the position before Cmed. Step 2. Let Si+, be the sequence obtainedfrom Si by interchanging a subset ofjobs in B with a subset of jobs in A such that
E

meA

pm< E pm and (
meB

meB

PM -E

meA

Pm)

is minimum.Then rearrange the jobs in B and A in LPTand SPT orders,respectively.Note that and largerADMedian)thanSi. Also, the effortrequired to complete Si+, will have a smallerCmed of sets B andA, which in turntend to increaseas Step 2 grows exponentiallywith the cardinality n increases. However, this is unavoidablesince the number of non-dominatedsequences also grows exponentiallywith n. Step 3. Repeat Step 2 until the shortest (n + 1)/2 (or n/2 if n is even) jobs are in B and the rest in A. Step 4. The Si's are the set of non-dominatedsequences. For example suppose in a four job problem, the processing times are 5, 6, 9 and 11. The algorithm generates the following set of sequences.
Cmed

S1-11,6,5,9 S2_ 11,5,6,9 S3 9,6,5,11 S4 9,5,6,11 S5 6,5,9,11

19.5 19.0 17.5 17.0 15.5

ADMedian 25 26 27 28 34

Jobs in B 11, 6 11, 5 9,6 9,5 6,5

Jobs in A 5, 9 6, 9 5,11 6,11 9,11

C and R If the firstjob starts processing at time zero, then given that no inserted idle time is allowed, the completiontime of the last job is a constant(sum of all the processing times). Thus minimizing R would be the same as maximizing the processing time of the firstjob. Thereforeany sequence with the largest job in the firstposition is optimal.R can only take n values, with each value corresponding to one of the n jobs in the firstposition (ifp[l] = Pi, then R = Lm Pm- Pi). The minimumC value, for a given firstjob, is the mean of the sequence where the rest of the jobs are in non-decreasingorder of processing times, namely these sequences given by processing times are (pi,Pn,Pn-1, *p? n, and * Pi+-si' *.**Pi) if i 7& (Pns Pn-l . .. 'Pi) if i = n. Also note that an SPT sequence will have the smallest mean but the largestR
..

J.Mittenthal eta-Single

MachineBicriteriaProblems

1361

(shortestjob in the firstposition), and as the job in the firstposition is interchanged with jobs with larger processingtime, R will decreasebut C will increase.Thus therewill be n non-dominated points with the n different jobs in the firstposition and the rest of thejobs in SPTorder.If the convex combinationof C and R is being considered,the optimal sequence will be one of the non-dominatedpoints dependingon the value of A.

Cmed and R

It follows from a standard job interchangeproof that an SPT sequence minimizes Cmed.Therefore, (Cmed, R) is quite similar to (C, R). The only differenceis that there are multiple sequences capable of value correspondingto an existing sequence. Specifically, in an n job problem the achieving a Cmed sequences (i, Pn i Pn-I. Pl) for i = 1, 2, ..., Ln/21+ 1, where Lx]denotes the largest ipi+IqPi-l .* 'P integer less than or equal to x, all have the same Cmed value, but the sequence correspondingto i= Ln/2j + 1 has the smallest R value and so dominates the others. Hence, the discussion in the preceding section also applies here with the exception that Ln/2J+ 1 is the numberof non-dominated sequences.

Cmed and

VAR,Cmed and ADMean, C and ADMedian

These threecombinationsare not investigatedin this paper.This subjectivedecision was based on the observationthat in each of these (CT, DSP) pairs the dispersionmeasure is definedrelative to the other measures of central tendency. TADC: A Possible Surrogatefor VAR Let v1 and v2be two performance measures.Define v1 to be a perfect surrogatemeasurefor v2 if (i) vI is less computationally complex thanv2, and (ii) for any two sequences 7t and a, vl (7) < v1(r) if and only if v2(70) < v2(o-). That is, v1 is a perfect surrogate measurefor v2 if vI is easier to optimize than v2 and all sequences have the same relativerankingof best to worst under each of the measures.Hence, a perfect surrogatemeasure,if it existed, could be used in place of a more complex measure in both single and multiple criteriaproblemsto achieve an equivalent,yet easier to optimize, problem. In this section, the hypothesis that TADC is a perfect surrogatemeasure for VAR is considered. TADC was introducedby Kanet" as an alternatemeasure of variation.He points out that TADC is equivalentin complexity of calculationto variance,and demonstratesthat
n n

TADC
-

E
n

k=I m=k

E ICm-CkI

E (nm-rm2 +2m-n-1)p[m]. m=l

Thus TADC is a linear combinationof the orderedprocessing times. Its advantageover varianceis that optimal solutions to TADC are usually abundantand easy to find. Recall that Bagchi7establishedthat TADC has polynomial complexitywhile Kubiak6showed VAR is NP hard. Hence, property(i) from the perfect surrogatemeasure definitionis met. To investigateproperty(ii) of the definition,two differentsimulationtests were performed.In the first test, 120 problems were randomly generated and the TADC of the minimum VAR sequence was comparedto the minimumTADC. The processing times were integers, randomlygeneratedusing three different distributionswith different standarddeviations. The numberof jobs varied between 4 and 10. The processing times for the three sets of problems were generated using Uniform (1, 100), Normal (50, 25) and Normal (50, 10) distributions,respectively. In each of the 120 problems, the minimum VAR sequence also had the minimumTADC. While this of course does not prove thatthe sequence with minimum VAR also has minimum TADC, it does suggest a high frequency of occurrence.

1362

Journalof the OperationalResearchSociety Vol. 47, No. II

In the second test, scatterplots of standard deviation (SD) againstTADC for 100 randomlygenerated sequences were obtainedfor three differentproblem sets. The three sets were generatedusing a Uniform (1, 999), a Normal (500, 150), and an Exponential (300) distribution,respectively. Note that in an exponential distributionthe values of the mean and median are usually different.For each distribution, either 100, 50, or 20 points were evaluated and displayed on a scatter plot. In all nine scatter plots, TADC and SD were closely related and their relationshipcould be approximated by a straightline. See Figures 1(a)-{d) for a subset of the scatterplots and correspondingdistributions. It is clear, from careful inspection of these figures, that property(ii) is not satisfied. However, these figures do suggest that a sequence having a greater TADC value than another sequence often has a greater SD value. Furthermore,note that a sufficient condition for TADC to be a perfect surrogate measure for VAR is for the coefficient of determination,R2, to equal 1, and that the lowest R2 from Figure 1 is 0.937. Thus, while TADC is not a perfect surrogate measurefor VAR, it appearsto be the next best thing-a surrogatemeasurethat approximates the sequence rankingof VAR nearly all of the time. Finally, Figures 2(a) and 2(b) suggest why TADC may be less computationallycomplex than VAR. The two figures show all possible sequences of a set of jobs on a two dimensional graph representing two criteria.Figure 2(a) has C2 on the horizontalaxis and VAR on the vertical axis. Figure 2(b), on the otherhandhas C on the horizontalaxis and TADC on the vertical axis. The processing times of the jobs are as follows 5 6 9 11

TADC = - 4226.7 + 1446.7 SD RA2 = 0.994 480000 460000 440000 g4200001 420000 400000
(-.4

TADC = - 1750.9 + 1415.2 SD RA2= 0.982 260000 240000

g~~~~~~ ~~~240000l, ~ 220000


180000 160000
I

~~~~~~~~~~~200000

380000
360000 340000 240

140000 320 340

(a)

260

280

300

(b)

100

120

140

160

180

200

SD

TADC = - 1.9687e+4 + 5819.3 SD RA2 = 0.997 1.80e+6 -1 1 1.70e+6 vP

TADC 90000

- 2530.6 + 221.44 SD RA2= 0.937

~~~~~~~~~~~~800001
70000 60000

1.60e+6 Sw 1.50e+6

50000

~~~~~ ~~~40000 ~
20000 100

l 50e+6 1

,{

~~~~~~~~~~30000
200 300 400

1.40e+6

()

240

260

2SD

30

320

(d)

SD

FIG.1(a) TADC against SD, n = 50, U(l, 999), (b) TADC against SD, n = 50, N (500, 150), (c) TADC against SD, n =100, U(1, 999), (d) TADC against SD, n = 20, EXP(300).

J.Mittenthal et al.-Single MachineBicriteriaProblems 110

1363

100
90

00

a
a

a
V

a a
a

so80 70 *

00 0

60-

60 00
5

00
C0

50 1 200

300

400

500

(a)

Mean Squared

90
0i

0
0

0
0

80

a
00

70~

~~ _0

(b)

60 16

17

18

19

20

21

22

23

Mean
FIG.2(a) VAR against C2, (b) TADC against C.

Note that in comparisonto Figure 2(a), the points seem to align themselves horizontallyin Figure 2(b). This alignmentreduces the numberof non-dominatedsequences from 8 to 7. The main reason behind this alignment is that TADC has more sequences with the same objective value.

CONCLUSIONS We define and provide motivation for the class of CT-DSP problems. A set of new results are developed which togetherwith existing results help to complete the picture with regardto this class of problems. A summary of known results are presented in a tabular form. A comparison of the two dispersion measures,VAR and TADC, shows that TADC is a good surrogatefor VAR.

1364

Journalof the OperationalResearchSociety Vol. 47, No. II

APPENDIX Proof of Theorem1 Let B = {m: Cm< Cmed} and A = {m: Cm> Cmed}. Let ISI denote the cardinalityof the set S. Suppose n is odd. Note that since IBI= (n + 1)/2 and Cmed= C[(n+1)/2],the value of the median of the completiontime only dependson the processingtimes of the jobs in B. Thus, for any permutation jobs in B, Cmed is constant.Also note that ADMedian can be rewrittenas n
E min{n-m + 1, m-1 lp[m], m=1

consequentlyits positional weights are (0, 1, 2, .. ., 2, 1) and are independentof the processing times. Hence, for any given sets A and B, ADMedian will be minimized when the jobs in B are sequenced in non-increasingorder of processing times and the jobs in A are sequenced in non-decreasingorder of processing times (see Hardy, Littlewood and Polya'6). Therefore,given a non-V-shapedsequence, its ADMedian can be decreasedby reshedulingthe jobs in B in non-increasingorder and the jobs in A in non-decreasingorderwithout changing the value of Cmed.Thus only V-shapedsequences can be nondominated. Suppose n is even. Then
Cmed = (C[n/2] + C[(n/2)+l])/2 = C[n/2] +P[(n/2)+1]/2,

so now its value dependson the processing times of the jobs in B as well as the job in the firstposition of A. Therefore,for any given sets A and B, Cmed will be minimizedwhen the shortestjob in A is scheduled in its firstposition. Also note thatfor any given sets A andB, ADMedianis minimizedwhen thejobs in B are sequenced in non-increasingorder of processing times and the jobs in A are sequenced in nonsequence,its ADMedianand Cmed decreasingorderof processingtimes. Therefore,given a non-V-shaped can be decreasedby reshedulingthejobs in B in non-increasingorderand the jobs in A in non-decreasing order.Thus for the objective function (3), only V-shapedsequences can be non-dominated.

REFERENCES
1. W E. SMITH (1956) Variousoptimizersfor single stage production.Naval Res. Logist. Q. 3, 59-66. 2. P. DILEEPAN and T. SEN(1988) Bicriterionstatic scheduling researchfor a single machine. Omega 16, 53-59. 3. T. D. FRY, R. D. ARMSTRONG and H. LEWIS (1989) A frameworkfor single machine multiple objective sequencing research. Omega 17, 595-607. 4. A. NAGAR, J. HADDOCK and S. S. HERAGU (1995) Multipleand bicriteriascheduling:a literature review.Eur. J. Opl Res. 81, 88-104. 5. K. R. BAKER and G. D. SCUDDER (1990) Sequencingwith earlinessand tardinesspenalties:A review.. Opns Res. 38, 22-36. 6. W KUBIAK(1993) Completiontime varianceminimizationon a single machine is difficult. Opns Res. Letters 14, 49-59. 7. U. BAGCHI(1989) Simultaneousminimizationof mean and variationof flowtime and waiting time in single machine systems. Opns Res. 37, 118-125. 8. A. G. MERTEN and M. E. MULLER (1972) Varianceminimizationin single machine sequencingproblems.MgmtSci. 18, 518528. 9. V VANI and R. RAGHAVACHARI (1987) Deterministicand randomsingle machine schedulingwith varianceminimization.Opns Res. 35, 111-120. 10. J. MrrITENHALand M. RAGHAVACHARI (1993) Stochasticsingle machineschedulingwith quadratic early-tardy penalties. Opns Res. 41, 786-796. 11. J. J. KANET (1981) Minimizingvariationof flow times in single machine systems. Mgmt Sci. 27, 1453-1459. 12. P.DE, J. B. GHOSHand C. E. WELLS extension. Opns (1992) On the minimizationof completiontime variancewith a bi-criteria Res. 40, 1148-1155. 13. P. DE, J. B. GHOSHand C. E. WELLS (1992) Heuristicestimationof the efficient frontierfor a bi-criteriaschedulingproblem. Decision Sciences 23, 596-609. 14. E. EILONand I. C. CHOWDHURY (1977) Minimizingthe waiting time variancein the single machine problem.Mgmt Sci. 23, 567-575. M. RAGHAVACHARI 15. J. MITTENTHAL, and A. I. RANA(1993) On modeling schedulingproblemsthat simultaneouslyminimize mean and varianceof completiontime. TechnicalReport 37-93-371, Decision Sciences and EngineeringSystems, Rensselaer PolytechnicInstitute,Troy,NY 12180. 16. G. H. HARDY, J. E. LITTLEWOOD and G. POLYA (1934) Inequalities. CambridgeUniversityPress, London.

J.Mittenthal etal-Single MachineBicriteriaProblems

1365

17. J. J. KANET (1981) Minimizingthe averagedeviationof completiontimes about common due date. Naval Res. Logist. Q. 28, 643-651. 18. N. HALL models for minimizingcompletiontime variance.Naval Res. Logist. Q. 33, 49-54. (1986) Single and multi-processor 19. U. BAGCHI, R. S. SULLIVAN and Y CHANG (1986) Minimizingmean absolute deviationof completiontimes about a common due date. Naval Res. Logist. Q. 33, 227-240.

Received February 1994; accepted March 1996 after two revisions

You might also like