You are on page 1of 14

Int J Adv Manuf Technol (2009) 45:759–771

DOI 10.1007/s00170-009-2015-x

ORIGINAL ARTICLE

New dispatching rules to minimize rejection


and tardiness costs in a dynamic flexible flow shop
K. Kianfar & S. M. T. Fatemi Ghomi & B. Karimi

Received: 10 November 2008 / Accepted: 9 March 2009 / Published online: 1 April 2009
# Springer-Verlag London Limited 2009

Abstract This paper studies a flexible flow shop system collection of tasks. It is an important decision-making
considering dynamic arrival of jobs and the ability of process in the operation level. In a modern manufacturing
acceptance and rejection of new jobs. The problem environment, many scheduling problems occur. Most of the
objective is to determine a schedule that minimizes sum scheduling problems are very significant and hard to solve,
of the tardiness and rejection costs of jobs. A 0–1 mixed because of the complex nature of the problems. In this
integer model of the problem is formulated. Since this paper, we will focus on flexible flow shop scheduling
problem class is NP-hard, four dispatching rules have been systems. In the considered problem, jobs arrive dynamically
developed to solve the problem approximately. Moreover, a over the scheduling period and the system has the ability to
discrete event simulation model of the flexible flow shop accept or reject new arrived jobs. The objective is to find a
system is developed for the purpose of experimentation. schedule that minimizes sum of the tardiness and rejection
Four dispatching rules from the literature and four new costs of jobs. The following gives a brief survey of the
dispatching rules proposed in this paper are incorporated in literature relevant to the above problem.
the simulation model. Simulation experiments have been Hunsucker and Shah [2] examined the performance of
conducted under various experimental conditions charac- six different dispatching rules to minimize two tardiness-
terized by factors such as shop utilization level, due date based criteria in a dynamic flow shop. Lodree et al. [3]
tightness and number of stages in flexible flow shop. The suggested a new rule to minimize the number of tardy jobs,
results indicate that proposed dispatching rules provide while Branke and Mattfeld [4] demonstrated that avoiding
better performance under problem assumptions. early idle times helps in minimizing total tardiness in a
dynamic flow shop. More recently, Swaminathan et al. [5]
Keywords Dynamic flexible flow shop . Tardiness . Accept/ examined minimization of total weighted tardiness in a
reject jobs . Mathematical programming . Simulation . dynamic flow shop where new jobs arrive at every shift
Dispatching rule change, while Alfieri [6] studied the interaction between a
number of dispatching rules and due-date quoting policies
in a simple dynamic flow shop. For the mean flow time
1 Introduction objective, Hunsucker and Shah [7] investigated a con-
strained hybrid flow shop, which allows multiple identical
Production scheduling can be defined as the allocation of machines at each stage, and concluded that the SPT rule
available production resources over time to perform a performed best. Rajendran and Alicke [8] developed
dispatching rules to take into account the presence of
bottleneck machines in a static flow shop. The only
K. Kianfar : S. M. T. Fatemi Ghomi (*) : B. Karimi previous work that deals specifically with the mean flow
Department of Industrial Engineering, time criterion in a dynamic flow shop is that of Rajendran
Amirkabir University of Technology,
and Holthaus [9]. They compared 13 existing and proposed
424 Hafez Avenue,
Tehran, Iran dispatching rules in a ten-machine flow shop and also
e-mail: fatemi@aut.ac.ir concluded that the SPT dispatching rule gives the lowest
760 Int J Adv Manuf Technol (2009) 45:759–771

mean flow times for all the machine utilization levels RR (Raghu and Rajendran [24]), and PT+WINQ (Holthaus
examined [10]. and Rajendran [25]). From a relative evaluation of many
In 2003, Reddy and Narendran [11] studied a five- dispatching rules from the literature survey, it appears that
machine permutation flow shop problem with dynamically there is a scope to develop efficient dispatching rules that
arriving jobs belonging to different families in a stochastic could address a number of performance measures such as
environment where both the process times and the time the minimization of mean, maximum, and variance of flow
between arrivals are assumed to be exponential random time and tardiness of jobs and also minimizing the
variables. Sequence-dependent setup times occur between percentage of tardy jobs.
the jobs of distinct families. Reddy and Narendran As it is obvious from the literature reviewed above that
compared the quality of nine heuristics (a combination of dispatching rules are the most common solving approaches
three dispatching rules and three queue selection rules) with to the dynamic scheduling problems but in the case of static
regard to simulated data. Their objective was to minimize problems, other methods such as genetic algorithms
(1) the average job flow time in the system, (2) the average (Jungwatanakit et al. [26]) and heuristic approaches
job tardiness, and (3) the percentage of tardy jobs. (Pugazhendhi et al. [27]) are used.
Barrett and Kadipasaoglu [12] conduct a simulation The rest of the paper is organized as follows. Section 2
study that investigates several performance measures in the describes the considered problem. The mathematical model
dynamic flow shop. In years 1997 and 1998, Barman [13, for the flexible flow shop problem under consideration is
14] conduct two simulation studies that examine several introduced in Section 3. In Section 4, new dispatching rules
different performance measures for a hybrid flow shop. The for the problem are proposed and the important aspects of
key idea from these studies is the approach of employing the simulation model developed are described in Section 5.
combinations of dispatching rules as opposed to one rule. Section 6 presents the details of the experimentation and the
Rajendran and Holthaus [9] conduct simulation studies for results and analysis of different scenarios. Finally, Section 7
dynamic flow shops and job shops. They also propose three is devoted to conclusions and recommendations for future
dispatching rules to minimize flow-time-related perfor- studies.
mance measures.
Jayamohan and Rajendran [15] provide a set of new
dispatching rules to minimize various performance meas- 2 Problem definition
ures such as mean, maximum, and variance of flow time
and tardiness in dynamic shops. A static rule which This paper considers the problem of scheduling jobs in a
minimizes the number of tardy jobs is also proposed. To flexible flow shop environment. A flexible flow shop is a
evaluate these proposed rules, their relative performance is generalization of the classical flow shop model. There are
analyzed in open job shops and reported in comparison M stages and some stages may have only one machine, but
with the standard benchmark rules such as the SPT and at least one stage must have multiple machines. The jobs
EDD, popular rules like ATC and MOD, and the best have to visit the stages in the same order string from stage
performing rules in literature such as RR, PT+WINQ, PT+ one through stage M. A machine can process at most one
WINQ+SL and AT+RPT. job at a time and a job can be processed by at most one
Survey papers by Day and Hottestein [16], Panwalker machine at a time. Pre-emption of such a processing is not
and Iskander [17], Blackstone et al. [10], Haupt [18], and allowed. The problem consists of assigning jobs to
Ramasesh [19] provide a good insight into various machines at each stage and sequencing the jobs assigned
dispatching rules in dynamic job shop and flow shop to the same machine so that some operational criteria are
problems. minimized.
However, when job arrivals are dynamic, the use of The dynamic flow shop considered is one where jobs
dispatching rules has to be resorted according to Hunsucker arrive continually over time, without advance knowledge of
and Shah [7, 20]. Their studies were limited by the their process time requirements. Hence, the product mix is
selection of dispatching rules such as FIFO, SPT, and in continues state of change as new jobs arrive and
LPT, and efficient rules such as MOD (see Baker and Kanet completed jobs leave the shop.
[21]) and ATC (Vepsalainen and Morton [22]) were not Another assumption in this problem is the ability of
included. Moreover, a comprehensive set of measures accepting or rejecting new orders. According to this, when
involving mean and variance of flow time and mean and a new job enters the shop, we can decide to accept or reject
variance of tardiness of jobs was not considered. In it. This decision is based on production or rejection costs.
comparison to these studies, the study by Rajendran and When a job is accepted to be processed on stages, tardiness
Holthaus [9] appears more complete with respect to many cost may occur and in the other hand, when a job is rejected
dispatching rules such as COVERT (see Russell et al. [23]), to be processed a fix cost have to be paid as rejection cost.
Int J Adv Manuf Technol (2009) 45:759–771 761

The problem of scheduling in dynamic job shops and Cjt Completion time of job j at stage t
flow shops has been extensively studied for many years and Tj Tardiness of job j
t
it attracts the attention of the researchers and practitioners Xi;j;l A Boolean variable; 1 if job j is scheduled
equally. As a result of the scheduling decisions, job j immediately before job l on machine i at stage t and
receives a completion time cj, which is compared to its due 0 otherwise
date, dj. Let Tj represent the tardiness of job j. The objective Yj A Boolean variable; 1 if job j is accepted to be
of this paper is to determine the job schedules that the sum processed and 0 otherwise (this variable is fixed to 1
of rejection and tardiness costs of jobs is minimized. for all of the existing jobs at the start of scheduling
The following are other problem characteristics and period)
assumptions in this paper. rejj Rejection cost of job j
TCj Cost of one unit time tardiness for job j
1 Accepted jobs require processing on each stage.
2 A job once taken up for processing on a machine should
be completed before another job is taken, i.e. job pre- 3.2 Mathematical formulation
emption is not allowed.
3 There is no interruption in the shop floor, e.g. no The problem can be formulated as follows:
machine breakdowns.
4 There are no precedence constrains between jobs. X
n X
n
5 There are unlimited buffers between stages. min ¼ ðTCj : Tj Þþ ð1  Yj Þ: rejj ð1Þ
j¼1 j¼1
Here, the necessary notation is provided for the
scheduling problem discussed in this paper. We adapt the
three-field notation system α/β/γ of Graham et al. [1] to X
mt X
n
describe the scheduling problem. The α field describes the t
Xi;j;l ¼ Yl 8t; l j l 6¼ 0 ð2Þ
shop environment. The β field describes the setup infor- i¼1 j¼0
mation, other shop conditions, and details of the processing
characteristics, which may contain multiple entries. Finally,
the γ field contains the objective to be minimized. X
mt X
nþ1

According to this type of notation, the problem will be


t
Xi;j;l ¼ Yj 8t; j j j 6¼ n þ 1 ð3Þ
P P i¼1 l¼1
noted as FF=rj = Tj þ Rj . In this notation, FF shows
that the problem is the type of flexible flow shop. Symbol
in the second field denotes that all of the jobs are not nP
þ1
available at the beginning of scheduling period and the last t
Xi;0;l ¼1 8t; i ð4Þ
l¼1
field of the notation describes that the objective of the
problem is to minimize sum of the tardiness and rejection
costs.
P
n
t
Xi;j;ðnþ1Þ ¼1 8t; i ð5Þ
j¼0
3 Mathematical model

This section provides a mixed integer linear programming t


Xi;j;j ¼ 0 8t; i; j ð6Þ
formulation for the problem under consideration. At first,
notations used in this model are introduced.
P
n nP
þ1
3.1 Notations t
Xi;j;l ¼ t
Xi;l;j 8t; i; lj l 6¼ 0 ; l 6¼ n þ 1 ð7Þ
j¼0 j¼1

t Stage index, t=1,2,...,k


i Machine index, i=1,2,...mt
j,l Job index, j=1,2,...n 2 f0; 1g 8t; i; j; l
t
Xi;j;l ð8Þ
rj Arrive time of job j
mt Number of parallel machines at stage t
n Number of jobs
dj Due date of job j P
mt
utj  utl þ ðn þ 1Þ : xti;j;l  n 8t; j; l j j 6¼ l ð9Þ
ptj Process time of job j at stage t i¼1
762 Int J Adv Manuf Technol (2009) 45:759–771

 m  
Pt process time of job l, if job l is immediately scheduled after
Clt  Cjt  Otl þ t
Xi;j;l  1 :B 8t; j; lj j 6¼ l job j. The value of B is set to a very big constant. Constraint
i¼1
(11) ensures that the completion time of every job at each
ð10Þ
stage is a non-negative value. Constraint (12) specifies the
conjunctive precedence constraints for the jobs, which says
Cjt  0 8t; j that a job cannot start its processing at stage t+1 before it
ð11Þ
finishes at stage t. Constraint (13) applies only to stage one,
saying that a job cannot start its processing at stage one
before its release date. Constraint sets (14) and (15)
Clt  Clt1  yl : Otl 8t; l j t 6¼ 1 ð12Þ determine the correct value of the tardiness (Tj). Constraint
set (14) determines the correct value of the lateness (Lj) and
Eq. (15) specifies only the positive lateness as the tardiness.
It is noted that an optimal solution can be obtained by
Cj1  rj þ O1j 8j ð13Þ
running commercial mathematical programming software,
Lingo 8.0, with an Intel Pentium 2.4 GHz CPU. We have
found that the mathematical model can be used for solving
Tj  Cjk  dj 8j ð14Þ
problems with up to six jobs and four stages in acceptable
time.

Tj  0 8j ð15Þ
4 Proposed dispatching rules

In this section, four new dispatching rules are proposed for


yj 2 f0; 1g 8j ð16Þ the considered problem. Details of these rules are described
in the following four sub-sections. It is notable that all these
Equation (1) describes the objective function as sum of rules use a simple heuristic method to accept or reject the
the tardiness and rejection costs. We have Xi;0;l t
¼ 1 if the new jobs which arrive during the scheduling period. The
job l is sequenced as the first job on machine i at stage t, heuristics works in such a way that when a new job (job i)
t
and Xi;j;ðnþ1Þ ¼ 1 if job j is sequenced as the last job on enters the system, the tardiness costs of jobs are calculated
machine i at stage t. Constraints (2) to (8) ensure that the for two times; one time, including the job i in the
partial schedule on each machine at each stage is feasible. scheduling program and the other time without considering
Constraint sets (2) and (3) ensure that only one of the job i. If the difference between these two costs is higher
accepted jobs is assigned to each sequence position at each than the rejection cost of job i, the job will be accepted and
stage, and when a job is rejected, it does not enter the enters the scheduling; otherwise we will reject the new job
schedule. Constraint sets (4) and (5) ensure that only one and pay the rejection cost.
job will be assigned to the first and last positions,
respectively, on each machine at each stage. Constraint (6)
4.1 Proposed dispatching rule 1
assures that after the job has been finished at any stage, it
cannot be reprocessed at the same stage. Constraint (7)
This rule is developed base on Cooperative Dispatching
forces to construct a consistent sequence at every stage.
t (CD) method, which was introduced by El-Bouri et al. [28].
Constraint (8) specifies the decision variables Xi;j;l as binary
CD was designed to minimize mean flow time in a dynamic
variables. It is emphasized that job 0 and job n+1 are not
flow shop and they asserted that this rule gives improve-
real jobs and they are merely added for formulation of the
ments ranging between 1.5% and 11.4% over other
model.
dispatching rules considered in that study. Proposed
Constraint (9) prevents sub-tours in the sequence of jobs
dispatching rule 1 (PDR1) uses the main idea of CD rule
on machines. Existence of sub-tours makes the problem to
and adjusts it for the flexible flow shop environment and
be infeasible, similar to the TSP problem. Constraints (10)
tardiness performance measure.
to (13) find the completion time of every job. Constraint
The notations used in this proposed dispatching rule
(10) is a set of disjunctive constraints. It states that if jobs j
(PDR1) are given as follows:
and l are scheduled on the same machine at a particular
stage with job j scheduled before job l, then job j must M Total number of stages in the flexible flow shop
complete the processing before job l can begin. This s Stage where the current dispatching decision is
constraint set forces job l to follow job j by at least the pending (current stage)
Int J Adv Manuf Technol (2009) 45:759–771 763

m Index for stages Element zi,q represents the local tardiness cost at stage
k Index for machines m=s +q−1 in the event that the ith job of the set Ω is
y Index for jobs dispatched next. Hence, matrix Z can be shown as
ns Number of dispatching candidate jobs queuing at follows:
stage s
2 3
Pj,m Process time for job j on stage m z1;s ::: z1;m ::: z1;M
Ω Ordered set of dispatching candidate jobs ready 6 : : : 7
6 7
for processing on stage s Z¼6
6 zi;s ::: zi;m ::: zi;M 7
7
i Index for job which heads set Ω(lead job) 4 : : : 5
Rm Ready time for stage m zns ;s ::: zns ;m ::: zns ;M
rj Release time for job j
β Ordered set of dispatching candidate jobs that To calculate each of the matrix Z entities, the total
excludes current candidate under consideration, tardiness related to the optimal sequence of set Ω (headed
β=Ω−{i} by job i) on a stage, m (fi,m) is multiplied by a weighting
Ym Set of jobs queued at stage m’s buffer factor (Wm) which indicates current importance of each
Φm Set of jobs being processed on stage m stage, m. Next, the final cost of scheduling jobs (headed by
i
Cy;m ðbÞ Completion time for job y on a stage m according job i) on stage is obtained using Eqs. (17) and (18) given as
to the sequence of jobs contained in β leaded by follows:
job i  
Yi,m Earliest possible completion time for job i (lead uq ¼ min zi;q q ¼ 1; 2; :::; M  s þ 1 ð17Þ
i¼1;2;:::;ns
job) on stage m
1 i,m Slack time for job i on stage m
liy;m Slack time for job y on stage m subject to job i be MP
sþ1
the lead job ri ¼ ðzi;q  uq Þ i ¼ 1; 2; :::; ns ð18Þ
q¼1
Ti,s Tardiness for job y on stage m subject to job i be
the lead job
i
Ty;m Tardiness related to the earliest possible The job i*, dispatched at stage t, is the one that
completion time for job i on stage m minimizes Eq. (18) or:
Fi,m Total local tardiness of jobs on stage m, in the i ¼ ij minfri g ð19Þ
event job i is dispatched at stage m
zi,m Tardiness to stage m, resulting from selecting job So, PDR1 dispatches the i*th job in the queue at stage s.
i at the dispatching stage s
υm Minimum tardiness cost for stage m 4.1.1 Determining a local tardiness cost, zi,m
ρi Total penalty cost for all stages, s through m,
resulting from the selection of job i at stage s Each zi,m is obtained by solving a parallel machine
Wm Weighting factor for stage m scheduling problem to find an optimal order of the current
jobs at set Ω on stage m, subject to job i takes the first
position in that sequence. As stated before, the local
PDR1 is designed such a way that in each step selects tardiness cost, zi,m, is the total tardiness of the optimal
one job from a set Ω of queued candidates to process next sequence for Ω (headed by job i) at stage m, multiplied by a
on current stage (s). The downstream of the dispatching weighting factor which indicates stage's importance in
stage is consulted in a pooling process that seeks from each scheduling the flexible flow shop. Thus, the next step of
stage a priority ranking of the candidate jobs. According to the algorithm is to determine the local tardiness costs, zi,m,
this, PDR1 is classified as a state-dependent dispatching and the importance factor of stages, wm.
rule. A consulted stage endeavors to receive the jobs Determining fi,m values needs knowledge about the
contained in Ω according to a sequence that allows it to minimum sum of job's tardiness costs in sequence Ω on
minimize its total local tardiness. The lead job in a stage stage m, headed by job i. For this purpose, we should first
desired sequence for Ω is the job it favors to be loaded next find the completion time of jobs. The completion time of
at the dispatching stage s. lead job i is given by
At the first step of PDR1, total tardiness matrix (matrix
Z) should be derived from the local tardiness costs. The  
rows of the matrix z represent, in order, the jobs contained g i;m ¼ max Rm ; g i;m1 þ pi;m
ð20Þ
in Ω, and the columns are related to stages s through m. g i;0 ¼ ri
764 Int J Adv Manuf Technol (2009) 45:759–771

In the above equation Rm is the ready time for each stage sum of the remaining job's local process time), related to
m and is computed by Eq. (21) where Rk,m is the ready time the bottleneck stage, which has the most remaining work.
of the kth machine on a stage m. The second factor also implies that the farther downstream
  a stage is, the less influence it has on the dispatching
Rm ¼ min Rk;m 8k 2 m ð21Þ decision. This is because the validity of the assumptions on
The completion time of jobs contained in set β (set of which the sequence costs are calculated is strongest for the
dispatching candidate jobs that excludes job i) is computed stage immediately downstream, and weakens with increas-
by simulating a EDD flow of parts through the shop, from ing distance from the dispatching stage.
the current state until all in-process jobs ahead of job i in According to above definitions, the importance factor for
the system are flushed. After sorting jobs by use of EDD stage m is given by
dispatching policy, the completion times of jobs in β are 0P P 1
pj0 ;m þ pj0 0 ;m
n o B j0 2Φm j0 0 2Ym C m1
i
Cy;m ðbÞ ¼ max Rm ; Cy;m1
i
þ Py;m Wm ¼ @ A:a ð28Þ
ð22Þ Wmax
Cy;0 ¼ ry
i

where pj′,m and pj′′,m represent the process times for jobs
Here, the calculated completion times are used to
queued at stage m and jobs being processed on stage m,
determine the tardiness of jobs in set Ω. In order to do
respectively. Moreover we have
this, first slack times of the related jobs are computed and
( )
then with respect to these slack times, the tardiness of each P P
job is obtained. Equations (23) to (26) show how to Wmax ¼ max pj0 ;m þ pj0 0 ;m m ¼ s; :::; M
m
j0 2Φm j0 0 2Ψ m
calculate slack times and tardiness of the lead job i and all
jobs in set β. ð29Þ

P
M The term αm−1 adjusts the weights of stages based on
li;m ¼ di  pi;m  g i;m
0 8m ¼ s; s þ 1; :::; M their distance from the dispatching stage. The recommen-
m0 ¼mþ1
ded value for α is 0.6, according to limited experimental
ð23Þ trials using different α values. Although this empirical
study cannot be considered conclusively, this value appears
P
M
liy;m ¼ dy  py;m0  Cy;m1
i
ðbÞ 8m ¼ s; :::; M ; to work fairly well and will be used henceforth.
0
m ¼mþ1
8y 2 b 4.2 Proposed dispatching rule 2
ð24Þ
This rule uses the process times and due dates to minimize
  the total tardiness of jobs in a dynamic flexible flow shop.
Ti;m ¼ max li;m ; 0 8m ¼ s; :::; M ð25Þ The main idea of this rule and the next two rules is accepted
from the Wilkerson–Irwin theory. According to this theory,
n o when jobs i and j are ready to be dispatched in a single
i
Ty;m ¼ max liy;m ; 0 8m ¼ s; :::; M ; 8y 2 b ð26Þ machine problem, in the case that the relation t þ
maxfPi ; Pj g > maxfDi ; Dj gis satisfied, the job with shorter
The above procedure must be applied to all stages process time is selected and otherwise the job with earlier
downstream of the dispatching and then with use of Eq. due date will be selected. In the above relation, t is the time
(27), the total tardiness costs of jobs included in Ω (headed at which the dispatching decision is made.
by job i) is obtained. Notations used in these three algorithms (PDR2 to
PDR4) are given as follows:
P
fi;m ¼ Ti;m þ i
Ty;m 8m ¼ s; s þ 1; ::: ; M ð27Þ Pi,m Process time for job j on stage m
y2b
Di Actual due date of job i
As stated before, the second term included in zi,m is the t Time at which the dispatching decision
importance factor of stages. The effect of importance factor is made
is to give candidate jobs preferred by the more important s Stage where the current dispatching
stages a greater claim in the dispatching process. A stage decision is pending (current stage)
importance is influenced by two factors: its remaining work M Number of stages in the flexible flow shop
load and its sequential distance downstream. The first factor k Index for machine with earliest release
impresses the ratio of a stage's outstanding workload (the date on current stage, s
Int J Adv Manuf Technol (2009) 45:759–771 765

e Index for the last job processed on 4.3 Proposed dispatching rule 3
dispatching machine of current stage s
RPTi,s Remaining process time of job i from This rule is almost similar to PDR2, except the method of
stage s to stage M calculating adjusted job due dates. PDR3 uses slack
Ψm Set of jobs queued at stage m’s buffer splitting method in order to assess adjusted due dates as
Φm Set of jobs being processed on stage m described through Eqs. (32) to (34).
αm Number of jobs in set Ym P
βm Number of jobs in set Φm ðDi  M m¼1 pi;m  ri Þ
si ¼ ð32Þ
di,s Adjusted due date for job i on stage s M

di;M ¼ Di ð33Þ

PDR2 assumes that individual stages in a flexible flow


di;m ¼ di;mþ1  Pi;mþ1  s i 8m 2 fðM  1Þ; ðM  2Þ; :::; 1g
shop are self-interested parallel machine problems that seek
to optimize their own local objective functions. Therefore, ð34Þ
it is necessary to adjust due dates of different jobs to be
used in each stage. Start
Adjusted due dates of a job i in each stage are
computed by subtracting the total remaining process times
from the actual due date of job i, as shown in Eqs. (30) and Setting shop parameters
(31). 1- Utilization level
2- Number of stages
X
M
3- Number of machines in each stage
RPTi;s ¼ Pi;m ð30Þ
m¼sþ1
4- Tardiness and rejection costs

di;s ¼ Di  RPTi;s ð31Þ Generation of jobs


1- Inter arrival time: Exponential
Assuming that we want to choose one of the jobs
2- Arrival and process times of jobs
queuing at stages; PDR2 steps are as follows:
3- Due date setting: TWK method
Step 1. Sort the dispatching jobs in non-decreasing order of
due dates.
Step 2. If there is only one job in the list of dispatching
jobs, assign it to the current stage (s); otherwise, go to step Simulation of the dynamic flexible flow shop
3.
1- Queue discipline: based on dispatching rule under study
Step 3. Take two first jobs from the list as jobs i and j, 2- Data collection
respectively. Jobs numbered on arrival in the shop
Step 4. Test the following inequality for jobs i and j Statistics on 1000 completed jobs (250-1250)
3- Experimental design
Number of replications: 10

t þ maxfPi;s ; Pj;s g > maxfdi;s ; dj;s g

Performance related statistics


Step 5. If the relation is true, go to step 6, else select job i 1- Performance measures with respect to tardiness and rejection costs
(job with earlier due date) and go to step 7. 2- Statistical analysis of results using ANOVA and LSD tests
Step 6. Select the job with shorter process time between
jobs i and j.
Step 7. Assign index i to the selected job and index j to the
next job from the list.
Step 8. Repeat the above steps until all jobs in the list are Stop
checked.
Step 9. Dispatch final selected job on the current stage. Fig. 1 Flow chart for flexible flow shop simulation experiment
766 Int J Adv Manuf Technol (2009) 45:759–771

Table 1 Experimental settings for the scenarios

Scenario Experimental setting Purpose of investigation

Utilization level (%) Due date tightness factor Number of stages

1 80 4 10 Base case-analyze the performance of scheduling rules


2 80, 90 4 10 Analyze the effect of changing utilization level
3 80 3, 4, 5 10 Analyze the effect of due date tightness factor
4 80 4 5, 10, 15 Analyze the effect of changing number of stages

where ri is the release time of job i and σi is the average passed. Therefore, the proposed value for the adjusted due
operation slack for job i. date, di,m,t, of job i on stage m at time t is given by
All steps of the rule are just like PDR2, except the
X
M X
M
inequality in step 4 which will change to t þ max di;m;t ¼ Di  t  qi;m0  Pi;m0 ð35Þ
fPi;s ; Pj;s g > maxfdi;s ; dj;s g. m0 ¼mþ1 m0 ¼mþ1

4.4 Proposed dispatching rule 4 where qi,m is the amount of time job i waits on the queue in
front of stage m.
This rule is also similar to rules PDR2 and PDR3. The only Since qi,m depends on the scheduling of jobs down-
difference between this rule and two previous rules is the stream in the flow shop, the sequence of all downstream
way local due dates are presented. In PDR4, the concept of jobs must be known to compute the waiting times
adjusted due dates (ADD) proposed by Lodree et al. [3] is accurately. So, the following equation is an alternative to
applied to compute local due dates. calculate di,s,t using the estimate of the completion time of
According to ADD, a job's due date on the current stage job i if at time t, job i is to be scheduled on current stage s
can be interpreted as the date that a job has to be completed ðbci;s;t Þ.
on the current stage so that its processing on the last stage di;s;t ¼ Di  bci;s;t þ Pi;s ð36Þ
can be completed on or before its actual due date. This can
be achieved by subtracting the job's remaining process where bci;s;t can be determined as
times and remaining waiting times from its actual due date. X
M X
M
The due date should also be adjusted to accommodate the bci;s;t ¼ t þ Pi;s þ qi;m þ Pi;m ð37Þ
current time t. If a scheduling decision is made when t>0, m¼sþ1 m¼sþ1
which is characteristic of dynamic scheduling, then the
Most of the effort in computing di,s,t involves computing
expression Di −t represents the amount of time job i has
bci;s;t , which is determined by an iterative process. At each
available to meet its due date since t time units have already
iteration, a different job i queued in front of stage s is
assigned to this stage and then scheduled, assuming
temporarily that subsequent queues for m 2 fs þ 1; mgobey
220 the FIFO dispatching rule. The completion time of job i on
the last stage is assumed to be the value of bci;s;t .

210
Performance Measure

5 Experimental design for the simulation study


200
In the real world, we usually deal with dynamic environ-
ments which include stochastic events such as random
190

180 Table 2 ANOVA results for one-way analysis for scenario 1

Source of Sum of df Mean F P


170
variation squares square ratio value
EDD RR PT_PW COVERT PDR1 PDR2 PDR3 PDR4
Dispatching Rule D: dispatching rule 167,786.986 7 23,969.569 7.774a 0.000
a
Fig. 2 Mean performance measure (U=80%, c=4, M=10) The mean difference is significant at the 0.05 level
Int J Adv Manuf Technol (2009) 45:759–771 767

Table 3 Results of LSD test 230


utilization level
Scheduling rule Mean Std. deviation High
220 Low

PDR1 180.33a 53.289

Performance Measure
PDR4 183.1a 54.333 210
PDR3 186.49a,b 54.238
PDR2 186.93a,b 53.647 200
COVERT 188.79a,b 54.005
RR 190.92a,b 55.759
190
PT_PW_ODD 204.05b,c 59.832
EDD 214.46c 58.734
180

170
arrival time, machine breakdown, changing due date, order
EDD PT_PW PDR1 PDR3
cancellation, emergency orders, etc. Simulation is among RR COVERT PDR2 PDR4
the ways to analyze dynamic situations.
Dispatching Rule
In this paper, a discrete event simulation model is
developed for the operation of the flexible flow shop Fig. 3 Interaction plot for scenario 2—performance measure
production system. The simulation model is developed using
the C++ programming language and run on a PC with exponential distribution. The mean of this exponential
2.4 MHz Pentium IV processor and 512 MB of RAM. Figure 1 distribution is obtained using the following relationship.
shows a flow chart of the problem simulation experiment. mm
l¼ ð39Þ
U : nm
5.1 Parameter setting of the simulation model
where λ is the mean interval time, μm the mean process
In scheduling problems, the total work content (TWK) time, U the shop utilization level and nm the number of
method has been used widely for due date assignment and machines in a stage, m. Two levels of machine utilization
in the studies by Kanet and Hayya [29] and Raghatz and are tested in the experiments, viz 80% and 90%.
Mabert [30], the TWK method is seen to be superior to In this study, the number of stages is considered at three
other due date setting methods. Using TWK method, the levels of M=5, M=10, and M=15 and the process times of
due date of each job is set equal to the sum of the job arrival jobs are randomly sampled from the integer numbers in the
time and a multiple of the total job process time. Thus, the range of 10–40. Also, the number of machines in each stage
due date of a job is determined by the following equation. is randomly sampled from a discrete uniform distribution
" ! # between 1 and 3 in the case that number of stages is less
X than or equal to 10, otherwise it is sampled from discrete
D i ¼ ri þ Pi;m : c ð38Þ uniform distribution between 1 and 5.
m

where ri is the arrival time of job i, c the due date tightness


factor, and Pi,m is the process time of job i on stage m. 6 Experimentation
Simulation experiment has been carried out for three levels
of due date tightness factor (c=2, c=3, c=4) to represent In the present paper, four dispatching rules are used from
tight and loose due date settings. the literature (EDD, RR, PT+PW+ODD, COVERT) and
It has been observed in the literature that in scheduling compared with four new dispatching rules proposed in
problems, the time between arrivals of jobs follows the Section 4.

Table 4 ANOVA results for two-way analysis for scenario 2

Source of variation Sum of squares df Mean square F ratio P value

D: dispatching rule 24,882.100 7 3,554.586 10.191a 0.000


U: utilization level 1,040.400 1 1,040.400 2.983 0.086
Interaction of DU 1,334.100 7 190.586 0.546 0.798
a
The mean difference is significant at the 0.05 level
768 Int J Adv Manuf Technol (2009) 45:759–771

Table 5 ANOVA results for two-way analysis for scenario 3

Source of variation Sum of square df Mean square F ratio P value

D: dispatching rule 34,657.496 7 4,951.071 12.678a 0.000


C: due date tightness 35,269.858 2 17,634.929 45.157a 0.000
Interaction of DC 6,413.342 14 458.096 1.173 0.298
a
The mean difference is significant at the 0.05 level

Now, we discuss about why the above four rules are the system for the completion of 1,000 jobs. Four scenarios
selected from the literature in our experimental study. EDD is are defined based on the parameter values of Section 5.1
the first dispatching rule that is perhaps the most commonly and summarized in Table 1. The purpose of defining
used in industry for easy implementation and practice on the various scenarios is to evaluate the performance of
shop floor. This rule is known to minimize the values of dispatching rules under different conditions.
maximum tardiness and variance of tardiness, and hence is In the scenarios, jobs are numbered with respect to their
used as a benchmark for performance criteria related to due arrival time and jobs in the range of 250–1,250 are used for the
dates in many studies. In 1993, Raghu and Rajendran [24] computation of the performance measure, because the first 250
proposed a new dispatching rule called RR and compared it jobs should be passed to reach the system to steady state period.
with a number of rules such as ATC, MOD, CR+SPT, and S/
RPT+SPT. Finally, they concluded that RR performs quite 6.3 Results and discussion for scenario 1
well with respect to flow time, tardiness, and root mean
square tardiness in dynamic scheduling problems. The PT+ Scenario 1 represents the base case wherein the purpose of
PW+ODD is considered as another rule with which our analysis is to investigate the performance of rules in dynamic
proposed dispatching rules are compared, it is because the flexible flow shop environment. In this scenario, a single factor
study by Jayamohan and Rajendran [15] showed that PT+ is assessed which is dispatching rules and includes eight levels.
PW+ODD is the best performing rule to minimize total This experiment is implemented in ten replications, considering
tardiness in comparison with some other rules such as SPT, utilization level, U=80%; due date tightness factor, c=4; and
EDD, MOD, PT+WINQ, and AT+RPT. Finally, Rajendran the number of stages, M=10. Figure 2 shows the average
and Holtaus [9] considered a total of 13 dispatching rules and values of the performance measure (sum of tardiness and
concluded that COVERT has the best performance for the rejection costs) in ten replications.
objective of minimizing mean tardiness. Hence, this rule was As shown in Fig. 2, four proposed dispatching rules in
taken as the fourth dispatching rule from the literature to be this study provide better performance in comparison with
compared with our proposed rules.
In the following, some experiments are conducted 240
according to the predefined parameters values to evaluate due date tightness

the performance of four dispatching rules from the literature 230 Tight
and four proposed new rules. Medium
220 Loose
Performance Measure

6.1 Identification of steady state


210

The first step in the simulation experimentation is deter-


200
mining the initial transient period. For this purpose, moving
averages of the output performance measure are plotted to 190
identify the start of steady state period. Steady state period
starts at a time since then moving averages reach a level 180
value. It is found that the moving averages for the
performance measure in this study approaches a level value 170
when about 250 jobs are completed.
160
EDD PT_PW PDR1 PDR3
6.2 Identifying different scenarios RR COVERT PDR2 PDR4

Dispatching Rule
In this study, each run of the simulation model includes ten
replications and each replication simulates the operation of Fig. 4 Interaction plot for scenario 3—performance measure
Int J Adv Manuf Technol (2009) 45:759–771 769

Table 6 ANOVA results for two-way analysis for scenario 4

Source of variation Sum of square df Mean square F ratio P value

D: dispatching rule 22,517.729 7 3,216.818 8.153a 0.000


M: number of stages 589,776.358 2 294,888.179 747.359a 0.000
Interaction of DM 7,400.508 14 528.608 1.340 0.186
a
The mean difference is significant at the 0.05 level

the four dispatching rules from the literature. Also, among the utilization level changes. Simulation results are obtained
all of the rules, PDR1 has the best performance measure for the two-factor experiments wherein the eight dispatch-
under the conditions of this scenario. ing rules form the first factor and the two levels of shop
Table 2 summarizes the analysis of variance results for utilization form the second factor. Table 4 shows the results
scenario 1. F test is used to determine if there is any of two-factor ANOVA. According to this table, dispatching
significant difference among mean values. Since the P rule is the only significant effect in this scenario hence two
value in ANOVA of Table 2 is less than 0.05, the difference other effects are not statistically significant. As evident
among means at 95% confidence level is significant. Using from Fig. 3, PDR1 performs better than the other rules
LSD method, a multiple comparison test among mean when the utilization level is fixed at 80% or 90%.
values is conducted.
Table 3 gives the LSD results. LSD method classifies 6.5 Results and discussion for scenario 3
dispatching rules into three different groups labeled a, b,
and c. For the proposed performance measure (sum of mean In the base case (scenario 1), the due date tightness factor is
tardiness and rejection costs), PDR1 and PDR4 rules set equal to 4 but here, in order to investigate the effect of
belong only to the group a and seems to be the best due date tightness, the due date factor has been set at 3, 4,
performing rules. Though there is no statistically significant and 5 to represent tight to loose due dates. Simulation
difference among PDR3, PDR2, COVERT, and RR rules, experiments are conducted using a two-factor full factorial
PDR3 provides the smallest value. design. The experimental factors are dispatching rules
(eight levels) and due date tightness (three levels). Ten
6.4 Results and discussion for scenario 2 replications are made for each of the 24 (8 dispatching
rules×3 due date tightness level) simulation experiments.
In this scenario, two levels of shop utilization are used to Table 5 gives the results of two-factorial ANOVA. The
investigate how the system performance is affected when main effects of dispatching rules and due date tightness are
significant in this experiment.
300
number of stages It is obvious that there is an increase in the performance
280 Low measure values when the due date tightness is increased.
Medium Based on Fig. 4, PDR1 rule provides best performance
260
High when due dates are tight but under medium and loose due
date tightness conditions, PDR2 and RR rules are the bests,
Performance Measure

240
respectively.
220
Table 7 Final scenario analysis results
200
Problem condition Best performing rule
180

160 Base conditions (scenario 1) PDR1


Utilization level High PDR1, PDR4
140
Low PDR1
120 Due date tightness High PDR1
Medium PDR2
100
EDD PT_PW PDR1 PDR3 Low RR
RR COVERT PDR2 PDR4 Number of stages High PDR4
Dispatching Rule Medium PDR1, PDR4
Low PDR3
Fig. 5 Interaction plot for scenario 4—performance measure
770 Int J Adv Manuf Technol (2009) 45:759–771

6.6 Results and discussion for scenario 4 Further research can be done to use other improving
algorithms such as metaheuristics or the combination of
In this scenario, three different levels for number of stages metaheuristics and dispatching rules. The results presented
are considered to show how the performance of dispatch- in this paper should be interpreted with reference to the
ing rules is influenced when the number of stages in the dynamic flexible flow shop system considered for the
shop changes. Two-factor simulation experiments are experimental conditions described. Hence, there is a need
conducted with the dispatching rules constituting the first for further research to develop new scheduling rules under
factor and the three number of stages (m=5, m=10, m=15) different conditions. For example, the scheduling rules can
forming the second factor. Ten replications are made for be designed considering setup time between jobs or system
each of the 24 simulation experiments arising out of the disruptions such as breakdowns of machines. Moreover,
combination of eight dispatching rules and the three levels Taguchi’s design of experiments method can be used for
for number of stages. Table 6 presents the two-factor simulation experiments. Taguchi’s method helps in reduc-
ANOVA results. ing the number of experiments when experimentation of a
The main effects of dispatching rules and number of system involves many factors and many levels for each
stages are significant but interaction effects are not factor.
significant for the related performance measure. Figure 5
shows the means of the ten replications for the 24
experiments. It is seen that when the level related to the References
number of stages is high (M=15), performance measure
values are also higher, because in that condition, jobs wait 1. Graham RL, Lawler EL, Lenstra JK, Rinnooy Kan AHG (1979)
for processing at more stages and hence the increase in the Optimization and approximation in deterministic sequencing and
performance measure (sum of mean tardiness and rejection scheduling: a survey. Ann Discrete Math 5:287–326. doi:10.1016/
S0167-5060(08)70356-X
costs) is gained. 2. Hunsucker JL, Shah JR (1992) Performance of priority rules in a
due date flow shop. Omega 20(1):73–89. doi:10.1016/0305-0483
(92)90057-E
7 Conclusions and recommendations for future studies 3. Lodree E, Jang W, Klein CM (2004) A new rule for minimizing
the number of tardy jobs in dynamic flow shops. Eur J Oper Res
159:258–263. doi:10.1016/S0377-2217(03)00404-1
This paper addresses a new flexible flow shop scheduling 4. Branke J, Mattfeld DC (2005) Anticipation and flexibility in
problem in a dynamic environment that has never been dynamic scheduling. Int J Prod Res 43(15):3103–3129.
studied before in the literature. A 0–1 mixed integer linear doi:10.1080/00207540500077140
5. Swaminathan R, Pfund ME, Fowler JW, Mason SJ, Keha A
programming formulation for the problem is provided. (2007) Impact of permutation enforcement when minimizing total
Since the flexible flow shop problem is NP-hard, algo- weighted tardiness in dynamic flowshops with uncertain process-
rithms to find an optimal solution in polynomial time are ing times. Comput Oper Res 34(10):3055–3068. doi:10.1016/j.
unlikely to exist. This paper presents four dispatching rules cor.2005.11.014
6. Alfieri A (2007) Due-date quoting and scheduling interaction in
(PDR1 to PDR4) to achieve near optimal solutions. The production lines. Int J Comput Integrated Manuf 20(6):579–587.
first proposed rule (PDR1) is a kind of cooperative doi:10.1080/09511920601079363
dispatching rules and other three rules are developed based 7. Hunsucker JL, Shah JR (1994) Comparative performance analysis
on Wilkerson–Irwin theory. of priority rules in a constrained flow shop with multiple
processors environment. Eur J Oper Res 72(1):102–114.
Moreover, a discrete event simulation experiment has doi:10.1016/0377-2217(94)90333-6
been developed for this production system. The simulation 8. Rajendran C, Alicke K (2007) Dispatching in flowshops with
output has been subjected to steady state analysis to ensure bottleneck machines. Comput Ind Eng 52(1):89–106.
that further investigations are free from initial bias. Four doi:10.1016/j.cie.2006.10.006
9. Rajendran C, Holthaus O (1999) A comparative study of
dispatching rules from the literature and four new proposed dispatching rules in dynamic flowshops and jobshops. Eur J Oper
dispatching rules are included in the simulation study. Res 116(1):156–170. doi:10.1016/S0377-2217(98)00023-X
Different scenarios characterized by variations in shop 10. Blackstone JH, Phillips DT, Hogg GL (1982) A state-of-the-art
utilization level, due date tightness and number of stages survey of dispatching rules for manufacturing job shop operations.
Int J Prod Res 20(1):27–45. doi:10.1080/00207548208947745
are investigated. The results of simulating different scenar- 11. Reddy V, Narendran TT (2003) Heuristics for scheduling
ios indicate that proposed dispatching rules in this study sequence-dependent set-up jobs in flow line cells. Int J Prod Res
have a significant impact on the shop performance and 41:193–206. doi:10.1080/00207540210163973
provide better performance under most of shop conditions. 12. Barrett R, Kadipasaoglu S (1990) Dispatching rules for a dynamic
flow shop. Prod Invent Manage J 31(1):54–58
Table 7 summarizes the results of scenario analysis experi- 13. Barman S (1997) Simple priority rule combinations: an approach
ments and shows the best performing rules among the eight to improve both flow time and tardiness. Int J Prod Res 35:2857–
dispatching rules under different problem conditions. 2870. doi:10.1080/002075497194480
Int J Adv Manuf Technol (2009) 45:759–771 771

14. Barman S (1998) The impact of priority rule combinations on performance parameters. Int J Prod Res 25:1523–1540.
lateness and tardiness. IIE Trans 30:495–504. doi:10.1080/ doi:10.1080/00207548708919930
07408179808966489 24. Raghu TS, Rajendran C (1993) An efficient dispatching rule for
15. Jayamohan MS, Rajendran C (2000) New dispatching rules for scheduling in a job shop. Int J Prod Econ 32:301–313.
shop scheduling: a step forward. Int J Prod Res 38:563–586. doi:10.1016/0925-5273(93)90044-L
doi:10.1080/002075400189301 25. Holthaus O, Rajendran C (1997) Efficient dispatching rules for
16. Day JE, Hottenstein MP (1970) Review of Sequencing Research. scheduling in job shop. Int J Prod Econ 48:87–105. doi:10.1016/
Nav Res Logistics Q 17:11–40. doi:10.1002/nav.3800170103 S0925-5273(96)00068-0
17. Panwalkar SS, Iskander W (1977) A survey of scheduling rules. 26. Jungwatanakit J, Reodecha M, Chaovalitwongse P, Werner F
Oper Res 25(1):45–61. doi:10.1287/opre.25.1.45 (2008) Algorithms for flexible flow shop problems with unrelated
18. Haupt R (1989) A survey of priority rule-based scheduling. OR- parallel machines, setup times, and dual criteria. Int J Adv Manuf
Spectrum 11(1):3–16. doi:10.1007/BF01721162 Technol 37:354–370. doi:10.1007/s00170-007-0977-0
19. Ramasesh R (1990) Dynamic job shop scheduling: a survey of 27. Pugazhendhi S, Thiagarajan S, Rajendran C, Anantharaman N
simulation research. Omega 18(1):43–57. doi:10.1016/0305-0483 (2004) Generating non-permutation schedules in flowline based
(90)90017-4 manufacturing systems with sequence-dependent setup times of
20. Hunsucker JL, Shah JR (1992) Performance of priority rules in a jobs: a heuristic approach. Int J Adv Manuf Technol 23:64–78.
due date flow shop. Omega 20(1):73–89. doi:10.1016/0305-0483 doi:10.1007/s00170-002-1525-6
(92)90057-E 28. El-Bouri A, Balakrishnan S, Popplewell N (2008) Cooperative
21. Baker KR, Kanet JJ (1983) Job shop scheduling with modified dispatching for minimizing mean flowtime in dynamic flowshop.
due dates. J Oper Manage 3:37–42. doi:10.1016/0272-6963(82) Int J Prod Econ 113:819–833. doi:10.1016/j.ijpe.2007.11.005
90020-1 29. Kanet JJ, Hayya JG (1981) Priority dispatching with operation
22. Vepsalainen APJ, Morton TE (1987) Priority rules for job shops due-dates in a job shop. J Oper Manage 2:167–176. doi:10.1016/
with weighted tardiness costs. Manage Sci 33:1035–1047. 0272-6963(82)90004-3
doi:10.1287/mnsc.33.8.1035 30. Ragatz GL, Mabert VA (1984) A simulation analysis of due date
23. Russel RS, Dar-el EM, Taylor BW (1987) A comparative analysis assignment rules. J Oper Manage 5:27–39. doi:10.1016/0272-
of the COVERT job sequencing rule using various shop 6963(84)90005-6

You might also like