Single Machine Scheduling with time dependent learning eect

Pradipta Patra 13th March 2012

Introduction

Literature on single machine scheduling problem assumes that the time to process a job is xed or constant. But, in many real life scenarios this need not be true. The processing time of a job depends on starting time and position of the job in a sequence. For example: rms and employees perform the same job repeatedly. As they spend more time in performing a job, their eciency increases and time to process a job scheduled later (in sequence) is shorter. This phenomenon is called learning eect in literature. Another suitable example that can be cited is the inspection of steel plates and bars (produced by a foundry) in order to nd surface defects. Here the learning rate of the inspector does not depend on the number of plates or bars he has inspected. It depends on the actual total surface area of plates or total length of bars he has inspected. One class of literature assumes that the actual processing time of a job is function of its starting time. This branch of literature we are not discussing here. We will delve into the other branch where the actual processing time of a job is a function of its position in a sequence.

Literature Review

Biskup(1999) was the rst to study a job independent learning eect model. The processing time of a job j, scheduled in position r is given as follows: pjr = pj ra where j, r=1, ...., n pj is the normal sequence independent processing time of job j. a 0 is the constant learning index and n is the total number of jobs.

Biskup showed that the single machine scheduling problem with learning eect remains polynomially solvable if the objective is to minimze deviation from a common due date or objective is to minimize sum of ow times. But, the problem studied by Biskup had a drawback. In many real life scenarios, learning of some jobs can be better than that of others (all of which are in the scheduling sequence). This situation is called job-dependent learn-

ing eect. Mosheiov and Sidney (2003) came up with a job dependent learning effect model as follows: pr = pj raj where j, r = 1, ...., n j pr : actual processing time of job j if scheduled in position r. j aj : negative job dependent parameter. Kuo and Yang(2006a) used the above notion to study a time based learning eect model in single machine makespan scheduling problem. The model assumptions, notations, denitions are as follows: (a) There are n jobs (J1 , ......., Jn ) to be processed on a single machine. Each of these are available at time zero. (b) pi is normal processing time for job i i.e. sequence independent processing time or processing time if scheduled rst in sequence. (c) p[k] is normal processing time of a job if it is scheduled in kth position of sequence. Processing time of (k + 1)th job onwards is smaller due to learning eect. (d) pir is actual processing time of job Ji , if it is scheduled in position r in sequence. pir is dened as follows: pir = pi if r = 1
r1

pir = (p[1] + p[2] + ....... + p[r 1])a pi = (

p[k] )a pi if r 2

a 0 is constant learning index. pi 1 in order to gurantee that it is a learning eect.

On combining pir can be written as (1 +

p[k] )a pi .

(e) Model discussed in previous section indicates that, the more one prac-

tices better is the learning performance. (f) For a given schedule q, Ci = Ci (q) represents completion time of job Ji . The maximal completion time (completion time of last job that is leaving the system) or makespan is denoted by Cmax . The model is denoted by 1/LEtb /Cmax . Results: In classical single machine scheduling problem, makespan value is sequence independent. For the 1/LEtb /Cmax problem, there exists an optimal schedule that satises the following condition: the sequence of all jobs except the one processed rst is the smallest processing time rst. Kuo and Yang(2006b) extended the above model for the case where total completion time is minimized. The notation for the new model is:
n

1/LEt /
i=1

Ci . Ci is same as dened above.

Results: The classical problem of minimizing total completion time on a single machine is optimized by shortest processing time (SPT) policy. For job independent learning eect, the SPT rule holds. This was shown by Biskup(1999).
n

For time based model 1/LEt /

Ci , there exists an optimal schedule in

which job sequence is determined by shortest processing time (SPT) rule. Kuo and Yang(2006c) extended the above basic scheduling problem to a group scheduling problem with following characteristics: (a) The jobs are classied into groups. (b) There is no setup time between two consecutive jobs in same group.

(c) Each group needs group setup time for setting up tools, jigs and xtures on machines. (d) Kuo and Yang(2006c) introduces time dependent learning eect on two single machine group scheduling problems: Firstly, a makespan minimization problem. Secondly, a total completion time minimization problem. (e) The two models that we deal with here are denoted by: 1/G, S, LEt /Cmax and 1/G, S, LEt / Cij . G denotes group scheduling problem. S denotes existence of sequence independent group setup time. Remaining notations are same as already discussed above. Notations used are as follows: m : the number of groups (m 2) Gi : the ith group, i = 1, 2, ..., m ni : the number of jobs in group Gi , i = 1, 2, ..., m n : the total number of jobs (i.e. n1 + n2 + + nm = n) Jij : the jth job in group Gi , j = 1, 2, ..., ni si : the group setup time of group Gi pij : the normal processing time of Jij in the original sequence pr : the actual processing time of Jij which is scheduled in the rth poij sition in a sequence in group Gi pi[k] : the normal processing time of Ji[k] which is scheduled in the kth position in a sequence in group Gi pk : the actual processing time of Ji[k] which is scheduled in the kth posii[k] tion in a sequence in group Gi Cij : the completion time of Jij

Ci[k] : the completion time of Ji[k] which is scheduled in the kth position in a sequence in group Gi Cmax : the makespan of all jobs Cij : total completion time of all jobs Results: For 1/G, S, LEt /Cmax model the optimal schedule satises: (a) the job sequence in each group is SPT rst (b) the groups can be sequenced in any order. For 1/G, S, LEt / Cij model the optimal schedule satises: (a) the job sequence in each group is SPT rst (b) the groups are sequenced in nonni

decreasing order of (si +

pj )/ni . i[j]

Latest Developments

Wang et al.(2008) extended the above single machine scheduling models with time dependent learning eect to cases where each job j has a weight wj and due-date dj . They also use three dierent measure of performance: weighted sum of completion times, maximum lateness, number of tardy jobs of a given permutation. Weighted Sum of completion times: wj C j

Maximum lateness(Lmax ): max{Cj dj |j = 1, 2, ...., n} Number of tardy jobs: Uj

Cj represents completion time for job j. The three models that are dealt here are: 1/LEt / wj Cj ; 1/LEt /Lmax ; 1/LEt / Uj .

3.1

Weighted Sum of Completion Times(1/LEt /

wj Cj ):

Weighted Shortest Processing Time (WSPT)Rule: This is a variation of the SPT rule. Let pi and wi denote the processing time and the weight associated with the ith job to be done in the sequence ordered by the WSPT rule. The WSPT rule says that jobs should be processed in the following order: p1 p2 p .................. wn w1 w2 n In the single machine environment with ready time set at 0 for all jobs, the WSPT minimizes the weighted mean ow time. But, when we take into account learning eect, WSPT does not always yield optimality. We provide an example below. Example: Say, n = 2; p1 = 1; p2 = 2; w1 = 10; w2 = 21; a = 0.5. The schedule according to the WSPT rule is 2 1, yielding the value wj Cj = 67.77. But, the optimal sequence is 1 2, yielding the optimal value wj Cj = 60.70.

Wang et al. uses the WSPT rule as a heuristic for the problem 1/LEt / wj C j . We derive a worst case error bound as a measure of performance for the heuristic.

Result1: Let S be an optimal schedule and S be a WSPT schedule for the wj Cj (S) 1 problem 1/LEt / wj Cj . Then 1 = n wj Cj (S ) (1+ pj pmin )a
j=1

where pmin = min{pj |j = 1, 2, ...., n} and the bound is tight. Proof:

wj Cj (S) = w1 p1,1 +w2 (p1,1 +p2,2 )+...............+wn (p1,1 +............+pn,n ) = w1 p1 + w2 (p1 + p2 (1 + p1 )a ) + ............. + wn (p1 + p2 (1 + p1 )a + ........ + pn (1 + p1 + ....... + pn1 )a )
n j

w1 p1 + w2 (p1 + p2 ) + ...... + wn (p1 + ....... + pn ) =

wj
k=1

pk 1.

The above inequality follows from the fact that, (1 + p1 + ........ + pn )a Now the weighted completion time for optimal schedule,

wj Cj (S ) = (w1 p1,1 +w2 (p1,1 +p2,2 )+...............+wn (p1,1 +............+pn,n )) = w[1] p[1] + w[2] (p[1] + p[2] (1 + p[1] )a ) + ............. + w[n] (p[1] + p[2] (1 + p[1] )a + ........ + p[n] (1 + p[1] + ....... + p[n1] )a )
n j n1

j=1 n

w[j]
k=1 j

p[k] (1 +
j=1 n

p[j] )a pj pmin )a

j=1

wj
k=1

pk (1 +
j=1

The rst of the above inequalities follow from the fact that, (1 + p[1] + ........ + p[n] )a 1.

The second of the above inequalities will become more evident if we write out the inequality for n = 2.
n j

So we have, wj Cj (S )

wj Cj (S)
j=1 n j

wj
k=1 n

pk and pj pmin )a .

wj
j=1 k=1

pk (1 +
j=1

From these two inequalities we get the desired upper bound for 1 . When a = 0 the WSPT rule is optimal and bound for 1 is tight. Result2: For the problem 1/LEt / wj Cj , if the jobs have agreeable weights, i.e., pj pk implies wj wk for all the jobs j and k, an optimal schedule can be obtained by sequencing the jobs in nondecreasing order p of wjj , i.e., the WSPT rule is optimal. Proof: We try to prove this by using contradiction. Suppose there exists an optimal schedule which does not follow the WSPT rule. So we can nd atleast two p p adjacent jobs j and k such that wjj > wk . This logic is as per the denition k of WSPT described above. This implies that pj pk . Say in the schedule , job j is scheduled in rth position and job k in (r +1)th position. Now we come up with a new schedule where the positions of j and k are interchanged. The completion times of jobs processed before j and k are not aected by this interchange. Whereas completion times of the ones after j and k are not increased by interchange because pj pk .
r2 r1

Cj ()= p[1] + p[2] (1 + p[1]

)a

+ ...... + p[r1] (1 +
q=1 r2

p[q] ) + pj (1 +
q=1 r1

p[q] )a p[q] )a +

Ck () = p[1] + p[2] (1 + p[1] )a + ...... + p[r1] (1 +

p[q] )a + pj (1 +
q=1

pk (1 +
q=1

p[q] + pj )a

r2

r1

Ck ( )= p[1] + p[2] (1 + p[1] )a + ...... + p[r1] (1 +

p[q] )a + pk (1 +
q=1 r1

p[q] )a p[q] )a +

Cj ( ) = p[1] + p[2] (1 + p[1] )a + ...... + p[r1] (1 +

p[q] )a + pk (1 +
q=1

pj (1 +
q=1

p[q] + pk )a wj Cj () wj Cj ( )

From the above expressions we can compute and after some algebra show that wj Cj ()

wj Cj ( ) > 0. Hence

wj Cj () > wj Cj ( ). So is not an optimal solution. This is a contradiction to our assumption. Hence the optimal schedule follows WSPT rule. The motivation for result 2 is that optimal solutions for the classical versions do not hold in the presence of a time-dependent learning eect. But some special cases of these scheduling problems with a time-dependent learning eect modeled as above can be solved in polynomial time. We provide two more results that are corollaries of result 2. Result 3: For the problem 1/LEt , pj = p/ wj Cj , an optimal schedule can be obtained by sequencing the jobs in non-increasing order of wj .
p p2 Proof: This can be seen easily from result 2. As per result 2, w1 w2 1 p .................. wn is optimal. Since pj = p we have w1 w2 ........... wn n as the optimal rule.

Result 4: For the problem 1/LEt , wj = kpj / wj Cj , an optimal schedule can be obtained by sequencing the jobs in non-decreasing order of pj i.e. the SPT rule is optimal.

3.2

Maximum Lateness(1/LEt /Lmax ):

Earliest Due Date (EDD) Rule: In the single machine environment with ready time set at 0 for all jobs, the earliest due date rule orders the sequence of jobs to be done from the job with the earliest due date to the job with the latest due date. Let di denote the due date of the ith job in the ordered sequence. EDD sequences jobs such that: d1 d2 .......... dn .

10

EDD, in the above setting, nds the optimal schedule when one wants to minimize the maximum lateness, or to minimize the maximum tardiness. The optimality of EDD is not maintained when we take learning eect into consideration. We provide an example of this below. Example: Suppose n = 2 i.e. there are two jobs to be processed, p1 = 1, p2 = 100, d1 = 1, d2 = 0, a = 0.5. As per the EDD rule jobs should be sequenced as 2 1. The Lmax for this is (100 + 1(1 + 100)0.5 ) = 100.099. The other option is to follow the schedule 1 2. The Lmax for this is (1 + 100(1 + 1)0.5 ). So the EDD rule is not optimal when there is learning eect. In this paper, the EDD rule is used as a heuristic to solve the 1/LEt /Lmax problem. Since it is not an optimal solution we provide worst case error bound for the heuristic. Result 5: Let S be an optimal schedule and S be an EDD schedule for the
n Lmax (S)+dmax Lmax (S )+dmax

problem 1/LEt /Lmax . Then 2 = and the bound is tight, where Cmax is the optimal makespan of the problem 1/LEt /Cmax . Result 6: For the problem 1/LEt /Lmax , if the jobs have agreeable conditions, i.e., pi pj implies di dj for all the jobs i and j, an optimal schedule can be obtained by sequencing the jobs in non-decreasing order of dj , i.e., the EDD rule is optimal. Result 7: For the problem 1/LEt , pj = p/Lmax , an optimal schedule can be obtained by sequencing the jobs in non-decreasing order of dj , i.e., the EDD rule is optimal. Result 8: For the problem 1/LEt , dj = kpj /Lmax , an optimal schedule can be obtained by sequencing the jobs in non-decreasing order of dj , i.e., the EDD rule is optimal.

pi i=1 Cmax

3.3

Number of tardy jobs(1/LEt /

Uj ):

The classical problem 1/ Uj is solved by using Moores algorithm. We provide a brief discussion on this algorithm.

11

Moores algorithm: Step 1: Order the jobs in nondecreasing order of their due dates (EDD). Step 2: If no jobs in the sequence are late, stop. The schedule is optimal. Step 3: Find the rst late job in the schedule. Denote this job by . Step 4: Find a job with p = max1;2;...; pi . Remove job from the schedule and process it after the completion of all the jobs that were processed. Go to Step 2. As discussed for the previous two cases, Moores algorithm is not optimal for 1/LEt / Uj . So we nd worst case error bound. Result 9: Let S be an optimal schedule and S be the schedule obtained by Moores Algorithm for the problem 1/LEt / Uj . Then 3 = Uj (S) Uj (S ) (n 1)

Result 10: For the problem 1/LEt / Uj , if the jobs have agreeable conditions, i.e., pi pj implies di dj for all the jobs i and j, an optimal schedule can be obtained by Moores Algorithm. Result 11: For the problem 1/LEt , pj = p/ Uj , an optimal schedule can be obtained by sequencing the jobs in non-decreasing order of dj , i.e., the EDD rule is optimal. Result 12: For the problem 1/LEt , dj = d/ Uj , an optimal schedule can be obtained by sequencing the jobs in non-decreasing order of pj , i.e., the SPT rule is optimal. Result 13: For the problem 1/LEt , dj = kpj / ule can be obtained by Moores algorithm. Uj , an optimal sched-

12

3.4

Computational Experiments

Computational experiments were conducted to evaluate the eectiveness of the heuristics of WSPT, EDD and Moores Algorithm. The heuristic algorithms were coded in VC++ 6.0 and the computational experiments were run on a Pentium 4 personal computer. In order to test the performance of the heuristics of WSPT, EDD and Moores Algorithm relative to the optimal solutions, an enumerative algorithm is developed to nd the optimal value of each test problem. Optimization of the running time of the enumerative algorithm was not tried, since the main goal was to evaluate the performance of the heuristics by comparing the heuristic solutions with the optimal solutions. Performance of the heuristic algorithms is good, especially for larger values of a. Parallely, Wang(2009) showed that SPT rule is optimal for the problem with maximum completion as objective function, provided the learning effect a 1. For 1 a < 0, largest processing time (LPT) rule is optimal.

3.5

Exponential Learning Functions

Wang and Wang (2011) used a new learning eect model called the exponential learning function as follows:
r1

pA [i] pA = pj (at + ) where t = jr

pi
i=1 0

r, j = 1, 2, ......, n; 0; 0; a 1; + = 1;
i=1

pA = 0. [i]

The parameters , , a are obtained empirically. pj is normal processing time of job Jj . pA is actual processing time of a job when scheduled in [k] kth position of sequence. Wang and Wang showed that for maximum complete time as objective func13

pi tion and pj holds

, SPT rule provides optimal schedule. The same result

when sum of th power of completion times is used as the objective function. For weighted sum of completion time as the objective function, WSPT rule gives the optimal schedule. For maximum lateness as objective function, earliest due date (EDD) rule gives the optimal schedule.

3.6

Single Machine Group Scheduling with resource allocation and learning eect

Zhu, Sun, Chu and Liu(2011) borrowed the learning eect model discussed above and extended it to the case where actual job processing time is a function of the amount of resource allocated. The job scheduler in this case has to make two decisions simultaneously: job sequencing and resource allocation. The resource allocation function usually takes two forms: linear and convex. Linear: pA (uj ) = pj bj uj where 0 j pj : nominal processing time of job j bj : compression rate of job j uj : amount of resource allocated to job j uj : maximal amount of resource that can be allocated to job j pA (uj ): actual processing time of job j j Convex: pA (uj ) = ( uj )K where K is a positive constant. j j The problem characteristics are as follows: n jobs are available for processing at time zero. These n jobs are classied into m groups where number of jobs in groups are, n1 , n2 , ........, nm . 14
p

uj

uj <

pj bj ;

j = 1, ..., n

All jobs are ready to be processed on a single machine and are non-preemptive. Each group Gi has a normal setup time si . Each group contains ni jobs which are processed consecutively. Jij denotes jth job in group Gi and has normal processing time pij . J[i][j] denote the job occupying the ith group position and jth internal job position. denote group sequence and i denote internal job sequence in group Gi . Actual processing time of job J[i][j] depends on group position i, internal job position j and amount of resource allocated to it concurrently. When we consider convex resource allocation, the actual processing time of job Jij (scheduled in group position r and internal job position l)is: p ra1 la2 pA = ( ij uij )k , uij > 0; uij is amount of resource that can be allocated ij to job Jij ; a1 < 0, a2 < 0 are learning indices of group and job; k > 0 is a constant. pA = pij ra1 la2 bij uij is the actual processing time when resource alloij cation function is linear. The actual setup time of group Gi (when position of group Gi in sequence is r) is sA = si ra1 . i The two cost functions that are optimized are as follows:
m ni

makespan and total resource cost: 1 Cmax + 2

Vij uij
m ni

total completion time and total resource cost: 1

Cij +2
i=1 j=1

Vij uij

1 , 2 are positive parameters decided by the decision makers. Vij is the cost of one unit of resource allocation to job Jij . GT stands for group technology and RALE stands for resource alloca15

tion and learning eect. The current paper shows that the optimal group sequence for the probm ni

lem 1|GT, RALE, conv|1 Cmax + 2

Vij uij can be obtained in O(m3 )

time. Additionally it also provides optimal resource allocation as a function of group sequence and internal job sequence. The optimal job sequence within a group is also studied. The paper also proposes an optimization algorithm to solve the problem.
m ni

For the problem 1|GT, RALE, lin|1 Cmax + 2

Vij uij , the optimal

job sequence within a group follows the usual SPT rule. The optimal group sequence can be obtained by ordering the groups in non-decreasing order of
ni

si +
l=1

pil la1 .
m ni m ni

The same results are investigated for the problem: 1

Cij +2
i=1 j=1

Vij uij

16

Future Research

(a) One possible small area of future research is to determine the computational complexity of the three heuristics used in the paper by Wang et al.(2008). Also we can try for more sophisticated heuristics. In fact this can be a broad area of research because most of the work done after this have suggested algorithms that can be used to obtain optimal scheduling of jobs. Very few papers have dealt with the computational complexity of the problems. (b) Wang and Wang (2011) used a new learning eect model (exponential learning function). The objective functions considered by Wang and Wang are makespan, total weighted completion time, maximum lateness and sum of th power of completion times. In addition to this we can consider the number of tardy jobs as an objective function for future research. (c) Another area of recent research is to use a model with truncation learning eect. The actual job processing time of a job is a function which depends not only on the processing times of the jobs already processed but also on a control parameter. The use of the truncated function is to model the phenomenon that the learning of a human activity is limited. Very few papers compare the eciency of the various learning eect models used. A future area of reaserch will be to explore few more models and draw up comparison between these.

17

References

Biskup, D., 1999. Single-machine scheduling with learning considerations. European Journal of Operational Research 115, 173-178. Kuo, W.-H., Yang, D.-L., 2006a. Minimizing the makespan in a single machine scheduling problem with a time-based learning eect. Information Processing Letters 97, 64-67. Kuo, W.-H., Yang, D.-L., 2006b. Minimizing the total completion time in a single-machine scheduling problem with a time-dependent learning effect. European Journal of Operational Research 174, 1184-1190. Kuo, W.-H., Yang, D.-L., 2006c. Single-machine group scheduling with a time-dependent learning eect. Computers & Operations Research 33, 2099-2112. Mosheiov, G., Sidney, J.B., 2003. Scheduling with general job-dependent learning curves. European Journal of Operational Research 147, 665-670. Wang, J.-B., Ng, C. T. D., Cheng,T. C. E., Liu, L. L., 2008. Single-machine scheduling with a time-dependent learning eect. Int. J. Prod. Econ., vol. 111, no. 2, pp. 802-811. Wang,J.-B., Wang, J.-J, 2011.Single-machine scheduling jobs with exponential learning functions. Computers & Industrial Engineering 60, 755-759. Zhanguo Zhu, Linyan Sun, Feng Chu, Ming Liu, 2011. Single-machine group scheduling with resource allocation and learning eect. Computers & Industrial Engineering 60, 148-157.

18