Int. J.

Production Economics 141 (2013) 112–126

Contents lists available at SciVerse ScienceDirect

Int. J. Production Economics

journal homepage: www.elsevier.com/locate/ijpe

Robust scheduling for multi-objective flexible job-shop problems with

random machine breakdowns
Jian Xiong a,b,n, Li-ning Xing a, Ying-wu Chen a
a
Department of Management, College of Information System and Management, National University of Defense Technology, Changsha, 410073 Hunan, PR China
b
School of Engineering and Information Technology, University of New South Wales at the Australian Defence Force Academy, ACT 2600, Australia

a r t i c l e i n f o a b s t r a c t

Article history: This study addresses robust scheduling for a flexible job-shop scheduling problem with random
Received 14 October 2011 machine breakdowns. Two objectives – makespan and robustness – are simultaneously considered.
Accepted 28 April 2012 Robustness is indicated by the expected value of the relative difference between the deterministic and
Available online 6 May 2012
actual makespan. Utilizing the available information about machine breakdowns, two surrogate
Keywords: measures for robustness are developed. Specifically, the first suggested surrogate measure considers
Flexible job-shop schedule problem the probability of machine breakdowns, while the second surrogate measure considers the location of
Machine breakdowns float times and machine breakdowns. To address this problem, a multi-objective evolutionary
Robust scheduling algorithm is presented in this paper. The experimental results indicate that, compared with several
Surrogate measure
other existing surrogate measures, the first suggested surrogate measure performs better for small
Multi-objective evolutionary algorithm
cases, while the second surrogate measure performs better for both small and relatively large cases.
& 2012 Elsevier B.V. All rights reserved.

1. Introduction the measure robustness is mainly classified into two categories

The job-shop scheduling problem (JSP) has been well studied
in the manufacturing systems field during the past few decades.
The classic JSP, which is a combinatorial optimization problem, is
strongly NP-hard (Garey et al., 1976). An extension of the JSP, the
flexible job-shop scheduling problem (FJSP) has received consid-
erable attention in the field. The FJSP consists of two sub-
problems, including machine assignment and operation sequence.
Thus, the FJSP is a more complex version of the JSP and is also
strongly NP-hard. Bruker and Schlie (1990) were among the first
to address the FJSP and proposed a polynomial algorithm to the
problems with two jobs. Most existing research addressed this
problem with the assumption that the parameters are known and
deterministic. However, in real-world manufacturing systems,
schedules are often confronted with uncertain or stochastic
factors. For example, resource shortages and machine break-
downs can delay a schedule’s completion time.
In this study, we focus on the FJSP while considering uncer-
tainty, referred to as the stochastic FJSP (S-FJSP). The methodol-
ogy of robust scheduling (de Vonder et al., 2008) is employed to
solve the S-FJSP. In robust scheduling, robustness is usually the
objective for the schedules. According to the existing literature,
n
Corresponding author at: Department of Management, College of Information
System and Management, National University of Defense Technology, Changsha,
410073 Hunan, PR China.
E-mail address: xiongjian1984@hotmail.com (J. Xiong).
(Herroelen and Leus, 2005): quality robustness and solution robust-
ness. The former is often used to indicate the insensitivity of the
schedule performance under uncertainty in terms of the objective
value, such as makespan (Leon et al., 1994; Jensen, 2003; Lin and
Liao, 2007) and tardiness (Wu et al., 1999; Tseng et al., 2009),
while the latter usually refers to the insensitivity of activity start
(or complete) times to the uncertainty. Usually, the measure of
solution robustness can be considered the stability of the schedule
(Goren and Sabuncuolu, 2008). Quality robustness indicates the
ability to preserve a specified level of solution quality in the
presence of uncertainty. Because makespan is the most important
objective in the FJSP, which is also one of the objectives of our
research, for a predesigned schedule, we are interested in the
preservation of makespan when uncertainty is present. Moreover,
in most cases, solution robustness of a schedule is not large enough
to accommodate all possible changes (Goren and Sabuncuolu,
2008), so investigation of the quality robustness of a schedule in
the presence of uncertainty is required. Thus, in this research, we
focus on quality robustness rather than solution robustness. For the
remainder of this paper, the term ‘‘robustness’’ refers to quality
robustness.
With simultaneous consideration of makespan and robustness,
the FJSP holds a multi-objective nature. The multi-objective
approaches for the S-FJPS are seldom reported in the literature.
Usually, the two objectives are combined, and the problem is
transferred to a single-objective problem, such as that described
by Al-Hinai and ElMekkawy (2011). However, providing a wide

0925-5273/$ - see front matter & 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.ijpe.2012.04.015
 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126
range of solutions to decision makers might be more useful. Then,
decision makers can determine trade-offs between makespan and
robustness for their schedules. Thus, in this work, we study the
S-FJSP in a multi-objective optimization framework.
Because the uncertainty is ‘‘unknown’’, it is difficult to evalu-
ate the effects of uncertainty of the baseline schedule. One
approach for evaluating uncertainty is to simulate the uncertainty
scenarios. However, exhausted simulations may have large time
requirements (Liu et al., 2007). To tackle this problem, certain
surrogate measures (SMs) are usually employed to approximate
the robustness of a schedule. Slack-based SMs are often employed
in approaches for robust scheduling (Leon et al., 1994; Al-Fawzan
and Haouari, 2005). Another assumption for robust scheduling is
that the uncertainty is bounded, suggesting that some uncertainty
information is available in advance and could be utilized when
constructing the baseline schedule. However, most slack-based
SMs neglect this information. To utilize the information about
uncertainty, we propose two surrogate measures and investigate
their performances within a multi-objective framework. In the
first suggested surrogate measure, the workload information for
each machine is utilized, and the robustness is measured as the
weighted sum of the float time for each operation. For the second
suggested surrogate measure, the machine breakdown locations
are incorporated in the calculation of robustness. Computational
results suggest that the proposed surrogate measures are effective
for the S-FJSP.
The rest of the paper is organized as follows. In Section 2, we
briefly review the literature on robust scheduling. The problem
formulation and multi-objective optimization for FJSP are pre-
sented in Sections 3 and 4, respectively. In Section 5, we examine
several robust measures for S-FJSP and suggest two new surrogate
measures. The details of proposed multi-objective evolutionary
approach are presented in Section 6. Experimental results are
reported in Section 7. Finally, some conclusion and future work
are given in Section 8.
2. Robust scheduling
According to the existing literature, in job-shop scheduling
problems, uncertainty is attributed to two main aspects: the job
processing times, as described by Wu et al. (1999), Xia et al.
(2008), Tseng et al. (2009), and Lei (2010) and the machine
breakdowns or maintenance (Liao and Chen, 2003; Liu et al.,
2007; Goren and Sabuncuolu, 2008; Lei, 2011; Dong and Jang, in
press). Regarding the generation of robust schedules, the
approaches can also be divided into two categories: buffering
and non-buffering approaches.
Several studies (Mehta and Uzsoy, 1998, 1999; O’Donovan
et al., 1999) have shown that advanced insertion of appropriate
idle times is effective for handling disruptions. However, there are
two main difficulties for buffering strategies (Al-Hinai and
ElMekkawy, 2011): (1) the strategies must decide the locations
for inserting idle times and (2) they must decide the number of
idle times. In non-buffering approaches, a robustness measure
based on the baseline schedule is usually an objective for
optimization. Although extensive methods have been reported
in the literature, the concept and definition of robustness of a
schedule is still an open issue. Leon et al. (1994) used the average
float time as the surrogate model of robustness. Wu et al. (1999)
used expected weighted total tardiness as the robustness mea-
sure to study the job-shop scheduling problem with process time
variability. In Liao and Chen (2003), minimizing the maximum
tardiness was taken as the optimization objective. Tseng et al.
(2009) also focused on total tardiness in the single machine
problem with controllable process times. Gu et al. (2010)
113
proposed a competitive co-evolutionary quantum genetic algo-
rithm for the stochastic job shop scheduling problem while
minimizing the expected value of makespan. A similar research
was presented by Lei (2011), where minimizing the stochastic
makespan was taken as the objective for the job-shop scheduling
problem subject to random breakdowns. Dong and Jang (in press)
considered the performance measures of mean tardiness and
mean flow time under machine breakdowns for production
scheduling at a job shop.
Compared to the classic JSP, few studies have addressed the
FJSP under uncertainty. Mahdavi et al. (2010) described a simula-
tion-based decision support system for production control of a
stochastic flexible job shop manufacturing system. Wang and Yu
(2010) proposed a filtered beam search based heuristic algorithm
to solve the FJSP with machine availability constraint considera-
tions. Lei (2010) presented an FJSP with fuzzy processing time and
proposed a decomposition-integration genetic algorithm (DIGA)
to minimize the maximum fuzzy completion time. Moradi et al.
(2011) investigated an integrated FJSP with preventive mainte-
nance activities under the multi-objective optimization
approaches. Al-Hinai and ElMekkawy (2011) studied the FJSP
with random machine breakdowns using a two-stage hybrid
genetic algorithm. In their paper, robustness and stability of a
schedule were investigated with a bi-objective approach, and
three stability measures were suggested.
Because of its practical importance, robust scheduling is
widely used in other scheduling problems, such as production
planning (Daniels and Kouvelis, 1995; Aytug et al., 2005) and
project scheduling (Herroelen and Leus, 2005). Most literatures
regarding robust project scheduling are focused on resource-
constrained project scheduling. Al-Fawzan and Haouari (2005)
proposed a robustness measure of a given schedule that is based
on the total amount of free slack for all activities. Kobylański and
Kuchta (2007) proved the deficiencies of the robustness measure
in Al-Fawzan and Haouari (2005) and modified this measure to be
the minimum of all the free slack or the minimum of the ratios of
free slack/duration. Chtourou and Haouari (2008) followed the
approach of Al-Fawzan and Haouari (2005) but assigned weights
to the free slack of an activity, considering the number of its
successors and/or the sum of its required resources. In their work,
12 surrogate robustness measures based on free slack were
proposed and assessed. Lambrechts et al. (2008) measured the
robustness with the weighted deviation between the planned and
actual activity starting times during project execution and
employed a free slack-based objective function in the tabu search
procedure. Hazir et al. (2010) introduced surrogate measures to
estimate the schedule robustness for a discrete time/cost trade-
off problem. The surrogate measures were also based on slack,
including weighted slack, slack utility function and dispersion
of slack.
3. Problem formulation
3.1. Deterministic FJSP
FJSP is a generalization of the classic job-shop scheduling
problem (JSP). Similar to the JSP, FJSP takes into account the
assignment of each operation to a machine and sets its starting
and ending times. However, the task is more challenging than the
classic one because it requires a proper selection of a machine
from a set of machines to process each operation. Then, the FJSP
can be described as follows:
 There are n independent jobs that are indexed by i.
 There are m machines indexed by k.
114 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126
 Each job i consists of a sequence of ni operations Oi1 ,
Oi2 , . . . ,Oini .
 The execution of operation Oij ð1 r j rni ) requires one
machine out of a set of given machines represented as
M ij A M.
 The processing time of an operation Oij on machine k is
predefined and given by pijk.
 The objective is to find a schedule that has the lowest value of
makespan.
 For a schedule, the workload on machine k is represented as
Wk and the total workload of the schedule is denoted as Wtot.
There are some assumptions and constraints in FJSP:
 Jobs are independent from each other.
 Machines are independent from each other.
 Setup time of machines is negligible.
 Move time between operations is negligible.
 At a given time, a machine can execute at most one operation.
 No more than one operation of the same job can be executed
at a time.
 There are no precedence constraints among the operations of
different jobs.
3.2. Machine breakdown formulation
In practical manufacturing, the execution of a schedule is
usually confronted with disruptions and unforeseen events.
Although in reality, a shop floor might suffer from various
uncertain factors, we focus on the situation that the shop floor
may experience random machine breakdowns during schedule
execution. Similar to other literature on S-FJSP, we assume that
some information about the uncertainty of machine breakdowns
is available in advance, and can be quantified by some
distributions.
Machine breakdown is usually described in terms of two
parameters (Holthaus, 1999; Gholami et al., 2009; Wang and
Choi, in press): the first one is the mean time required to repair
the machine after its breakdown, and the second one is the
machine breakdown level, which represents the percentage of
time that a machine has failures. In this research, the machine
breakdown generation is comparable to the assumption and
procedure presented in Al-Hinai and ElMekkawy (2011). First of
all, it is assumed that all possible breakdowns are aggregated as
one breakdown. When a machine breakdown occurs, the opera-
tion being processed is resumed after the repair is finished. The
related parameters are approximated based on machine’s busy
time, represented as Tbusy. For each machine Mk, the aggregated
repair time after its failure, denoted as RTk, is assumed to follow
uniform distribution and represented as follows:
RT k ¼ ½b1 T kbusy , b2 T kbusy 
where b1 and b2 are the coefficients between the interval [0, 1],
T kbusy is the busy time of machine Mk. The machine breakdown
level is characterized by two parameters: probability of machine
breakdown (pk) and machine breakdown time (BTk). The prob-
ability of machine Mk to fail is approximated by the following
empirical relation (Al-Hinai and ElMekkawy, 2011):
T kbusy
pk ¼
W tot
where Wtot is the total workload of all machines. The machine
breakdown time is generated using the following uniform
distribution:
BT k ¼ ½a1 MS, a2 MS
where MS is the makespan of a schedule, a1 and a2 are the
coefficients between interval [0, 1].
4. Multi-objective optimization for S-FJSP
One of the most important issues for deterministic FJSP is
minimizing makespan of the schedules, which is also the objec-
tive addressed in most (if not all) literature of single-objective
FJSP. However, schedules that are developed based on minimum
makespan are not just very short, but also very dense and
compact (Al-Hinai and ElMekkawy, 2011). In other words, mini-
mum makespan schedules are prone to be sensitive to uncertain-
ties. In the presence of uncertainty, e.g., machine breakdown,
robust schedules are expected to perform well under stochastic
environment. Since makespan is the main objective for FJSP, the
solution quality of a schedule in the presence of machine break-
down could be measured by the delay of the schedule. Leon et al.
(1994) defined the schedule delay as the difference between the
deterministic makespan before disruption, and the actual make-
span after execution. It is calculated as follows:
dðSÞ ¼ MðSÞM0 ðSÞ ð4Þ
where M0(S) is the deterministic makespan of schedule S assum-
ing no disruption, while M(S) is a random variable which denotes
the actual makespan of S.
In this research, we use relative difference between determi-
nistic and actual makespans to indicate the schedule delay. Thus,
the schedule delay can be rewritten as follows:
MðSÞM0 ðSÞ
d0 ðSÞ ¼ ð5Þ
M 0 ðSÞ
Since M0 ðSÞ is constant once the schedule is determined, the
0
expected value of delay d ðSÞ can be written as follows:
0 E½MðSÞM 0 ðSÞ
E½d ðSÞ ¼ ð6Þ
M0 ðSÞ
0
It is clear that small expected delay E½d ðSÞ indicates the
schedule S is less sensitive to the disruption. Thus, a robust
predesigned schedule is expected to be with least delay con-
0
fronted with uncertainty. However, E½d ðSÞ can be reduced by
increasing the baseline makespan. In Leon et al. (1994), expected
makespan and expected delay are linearly combined to indicate
the robustness of a schedule for FJSP.
From the discussion above, it might be safe to consider that
the makespan, both deterministic and actual, and the schedule
delay are in conflict to a certain extent. In other words, there exist
a set of schedules which cannot be differentiated without any
preference from decision makers. The schedules with different
deterministic makespans perform differently and have various
levels on schedule delay. Since different decision makers could
focus on different characteristics of a schedule, it is worth
ð1Þ
providing the whole set of schedules to the decision makers
who can make a trade-off among the schedules.
When simultaneously considering the deterministic makespan
and the schedule delay, the S-FJPS can be modeled as a multi-
objective optimization problem. It can be formulated as follows:
obj: : ð1Þ min f 1 ¼ M0 ðSÞ
0
ð2Þ min f 2 ¼ E½d ðSÞ ð7Þ
ð2Þ
5. Robustness measure for S-FJSP
An analytical closed form of the expected value of the
objective function f2 in Eq. (7) is usually not available (Jin and
ð3Þ Branke, 2005). To easily compute the robustness measure, two
 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126 115

main approaches are employed in the literature: scenario-based
approaches and surrogate measures. In this paper, several robust-
ness measures are investigated in the multi-objective optimiza-
tion framework. The performance of different robustness
measures within the optimization problem given in Eq. (7) is
addressed in Section 7.
5.1. Scenario-based robustness measure
The scenario modeling approach is often employed to represent
the uncertainty of an optimization problem. The uncertainty can
then be modeled using discrete or continuous scenario sets with
intervals with known and specific terminals (Leung and Yue, 2004).
A scenario-based approach was recently used to measure the
robustness of a solution for scheduling problems by Abbass et al.
(2008), Chaari et al. (2011), and Xiong et al. (2011). For this type of
approach, the Monte Carlo method is usually employed to approx-
imate the expected value of the objective measure in the presence of
uncertainty. For example, the objective measure f2 in the studied
problem, given as in Eq. (7), can be approximated as follows:
1 PN
i ¼ 1 M i ðSÞM 0 ðSÞ
RM1 ¼ N
M 0 ðSÞ
where N is the sample size and Mi(S) is the actual makespan of a
schedule with a specific uncertainty. From Eq. (8), for each iteration
of fitness evaluation, the schedule must be evaluated N times in the
uncertainty sample space. This simulation-based fitness evalua-
tion is clearly time and resource intensive. Thus, in certain studies,
the robustness of a schedule was measured by surrogate measures
(SMs).
5.2. Slack-based SM for robustness measure
The slack-based method is usually used in SM applications to
generate robust schedules (Goren and Sabuncuolu, 2008). In
general, two types of slack are widely used in the scheduling
literature: total slack and free slack (Hazir et al., 2010). Total slack
is the difference between the earliest start time and latest start
time of an activity, while free slack is the amount of time that an
refer to the ‘‘total slack’’. Leon et al. (1994) used the average float
time of all operations to measure the robustness of schedules. The
experimental results reported by Leon et al. (1994) indicate that
there is a high correlation between the robustness and the
average float time of schedules. The robustness surrogate mea-
sure presented by Leon et al. (1994), denoted as RM2 in the
present paper, can be described as follows:
PO
ft
RM2 ¼ i ¼ 1 i ð9Þ
O
where O represents the number of operations, fti indicate the float
time of the ith operation.
In Al-Fawzan and Haouari (2005), the total sum of the free
slacks is proposed to measure the robustness of the schedules of
resource-constrained project scheduling. Since the schedules of
FJSP bear some similarity with project scheduling, this SM could
also be used to measure the robustness of the schedules of FJSP.
We denote this type of SM as RM3, which is calculated as the total
sum of free slacks of all operations, and is given as follows:
X
O
RM3 ¼ fsi ð10Þ
i¼1
ð8Þ
Similarly, O represents the number of operations and fsi indicate
the free slack of the ith operation.
5.3. Surrogate measures of robustness for the S-FJSP
Although slack-based (both sum and average slack) SMs are
widely used in the fields of project scheduling and job shop
scheduling, we state that the sum and the average of slack are not
suitable to measure the robustness of the schedules for the FJSP.
First, we consider a free slack-based example to illustrate this
statement. Suppose we have an FJSP with three machines and
three jobs and the process times for each operation on the
different machines are shown in Table 1. Two baseline schedules
Table 1
Process time table for an FJSP problem with size 3  3.
Oij M1 M2 M3

activity can be delayed without delaying the start of the very next
J1
activity (Al-Fawzan and Haouari, 2005). Fig. 1 provides an example O11 1 3 2
of the difference between total slack and free slack. From the O12 3 4 2
figure, only operations O23, O32 and O33 have free slack, indicated O13 3 1 6
J2
by gray blocks. However, the earliest and latest operation times
O21 3 1 3
ðO11 ,O12 ,O13 ,O21 ,O22 ,O23 ,O31 ,O32 ,O33 Þ are ð0; 2,6; 0,2; 6,0; 1,5Þ and O22 3 4 4
ð0; 2,6; 2,3; 7,1; 3,6Þ, respectively. Thus, the total free slack of each O23 4 4 5
operation is ð0; 0,0; 2,1; 1,1; 2,1Þ. J3
Total slack is also referred to as float time in certain studies O31 2 1 1
O32 2 4 4
(Gao et al., 2008). To more clearly differentiate these two
O33 5 6 4
concepts, for the remainder of this paper, we use ‘‘float time’’ to

M3 31 12 23

M2 21 32 13

M1 11 22 33

1 2 3 4 5 6 7 8 9 10

Fig. 1. An example of the difference between total slack and free slack.
116 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126

are considered, as shown in Fig. 2. The two schedules have the
same makespan (10 units). Schedule a (top of the figure) has a
free slack sum of seven units, shown in the figure with gray blocks
with a dashed line, while this sum for schedule b (bottom of the
figure) is only three units. If considered based on the total or
average free slack, schedule a has a better robustness measure
than b. Suppose that machines M1, M2 and M3 will suffer from
breakdowns during the time intervals [4,5], [5,6] and [4,5],
respectively. The realized execution of the two baseline schedules
r
is shown in Fig. 3, denoted as ar and b , where breakdowns are
indicated with dark blocks. Despite having more free slack,
schedule a has one unit makespan delay, while schedule b can
deal with the machine breakdowns without delaying the comple-
tion times for all the jobs.
As shown in the above example, slack-based measures for the
robustness of a schedule focus on the structure of the schedule,
but neglect the uncertainty. In most robust scheduling problems,
it is assumed that some information about uncertainty, such as
the probability density function, is available in advance. Thus, this
information can be utilized for baseline schedule generation,
which is common in real-world applications. For example, when
constructing a schedule for a job-shop, it is reasonable to avoid
assigning certain operations to a machine that often suffers from
breakdowns. For the problem studied in the present work, two

Fig. 2. Two baseline schedules with a makespan of 10.

Fig. 3. Two realized schedules that have suffered from machine breakdowns.
 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126 117

types of information related to uncertainty (machine break- The second suggested RM is denoted as RM5. The degree of
downs) can be utilized. The first type of information is related overlap between float time and breakdown period is indicated
to the probability of breakdown occurrence for each machine with a function of distance between the average central points of
during the execution of a schedule. The second type is the the two periods, as shown in Fig. 5. Thus, RM5 is provided as
location of the breakdown for each machine. follows:
As assumed in Section 3.2, the probabilities of machine break-
X
O
1
downs are related to the busy time of each machine. Because the RM5 ¼ m  ftm ð12Þ
m k
maximum busy time of a machine is equal to its workload, a m ¼ 1 expð9Aveft Avemb 9Þ
machine with a heavy workload is more likely to suffer a break-
where km is the index of the machine which processes the mth
down after hard work. In this case, certain float times for the
operation, Avem ft is the average central point of the float time
machine might reduce the makespan delay for the whole sche- m
interval of mth operation (if exists) and Avekmb is the average
dule that results from machine breakdowns. Thus, we consider
central point of the machine breakdown interval on machine km.
that the float time on a machine with a heavier workload is more
Here we simply use the exponential function to indicate the non-
important than that on the machine with a lighter workload.
linear relation between the overlap degree and the schedule’s
Then, we modify the slack based robustness measure as follows:
performance in the presence of breakdowns.
Robustness of a schedule for the FJSP can be measured with the
We have five SMs for robustness measure, three SMs (RM1, RM2
weighted sum of float times for each operation. The first suggested
and RM3) are from the literature, and two SMs (RM4 and RM5) are
robustness measure is denoted as RM4 and is formally described
suggested for this S-FJSP. One of the main focuses of this study is the
as follows:
comparison of the performances of different SMs within a multi-
XO objective optimization framework (as shown in Eq. (7)). This
Wm
RM4 ¼ ft ð11Þ comparison will be implemented with the following steps:
W tot m
m¼1 Step 1: Replace the robustness objective f2 in Eq. (7) with each
of the five SMs. Then, the five non-dominated sets can be obtained
where O is the number of operations, ftm denotes the float time of
with the multi-objective optimization algorithm.
the mth operation, W m is the workload of the machine on which
Step 2: Investigate the performances of the solutions in each
the mth operation is processed, W tot is the total workload.
non-dominated set. In this step, a simulation-based method is
For the FJSP, the performance of a schedule in the presence of
employed to investigate the obtained schedule in the presence of
uncertainty is likely related to the nature of the uncertainty
breakdowns. The robustness objective in Eq. (7) is approximated
affecting the schedule, e.g., the location period of the uncertainty.
with the expected value from sampling. Then, an updated non-
We also provide schedule b, shown in Fig. 2, as an example.
dominated set can be obtained for each SM.
Suppose the breakdown of machine M2 will not occur during the
Step 3: Compare the updated non-dominated sets using a
time interval [5, 6], but during the interval [4, 5]. Then, the
multi-objective measure. In this study, we employ set cover to
schedule will be delayed by one unit, as shown in Fig. 4.
indicate the performance of a non-dominated set, which will be
Thus, it is intuitive that it will be useful to incorporate the
explained in detail in the experimental section.
location information of machine breakdowns into the baseline
schedule generation. We propose another surrogate measure
based on this idea, denoted as RM5. RM5 involves scheduling
operations on a machine in such a way that provides as much
float time as possible during the periods where the machine
breakdowns will likely occur. Thus, the float time within the
machine breakdown period is considered more important than
the float time outside the breakdown period. Fig. 5 provides an
example of this measure. Suppose there are three float times on a
machine that each are the same duration: float time1, float time2
and float time3. Because the float time is expected to absorb
uncertainty (machine breakdowns), the degree of overlap
between the float time period and breakdown period can be used
to indicate the importance of the float time. In the example
shown in Fig. 5, float time2 is considered more important than
float time1 and float time3. Next, we define the second surrogate
measure of robustness as a weighed sum of float times with
consideration of the location information of machine breakdowns. Fig. 5. An example of different float times on a machine.

Fig. 4. The realized schedule b suffered from different machine breakdowns.

118 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126
6. Evolutionary algorithm for S-FJSP
It is clear that we need a multi-objective optimization
approach to solve the problem. To perform this task, one of the
classic MOEAs, named NSGA-II (Deb et al., 2002), is employed as
the multi-objective optimization algorithm. Since the evolution
speed of genetic algorithms (GAs) is relatively slow, in the
proposed approach, useful information contained in the solutions
is extracted and utilized during the evolution process.
6.1. Framework of the algorithm
In our previous research (Xing et al., 2010), a knowledge-based
heuristic searching architecture (KBSHA) was proposed to utilize
the heuristic information during the search process. The KBSHA
consists of two modules, namely, heuristic searching module and
knowledge module. The former takes charge of searching through
the vast solution space, while the latter is used to learn some
available knowledge from the optimization, and then apply the
existing knowledge to guide the current heuristic searching. In
this research, we propose an instantiation of KBSHA to solve
S-FJSP. The framework of the proposed algorithm is shown in
Fig. 6.
Fig. 6. Framework of the proposed multi-objective evolutionary algorithm.
Since FJSP consists of two sub-tasks: machine assignment and
operation sequencing, in the knowledge model, two types of
knowledge are developed (Xing et al., 2010): operation assign-
ment machine knowledge (OAMK) and operation assignment
priority knowledge (OAPK). The OAMK is accumulative knowl-
edge of assigning the given operation to a more appropriate
machine, while the OAPK is accumulative knowledge of the more
appropriate processing priority for the given operation. In each
iteration of evolution process, these two types of knowledge are
extracted and updated. Based on the available knowledge, the
feasible schedules can be generated. For the details about knowl-
edge representation, extraction, update and solution generation
based on knowledge, reader are refers to Xing et al. (2010). In the
proposed algorithm, the knowledge is utilized by the following
procedure: every dth generation, the last x% individuals in the
population are replaced by the new generated solutions based on
the accumulative knowledge OAMK and OAPK.
6.2. Chromosome representation and population initialization
6.2.1. Chromosome representation
We combine the two-vector representation presented by Gao
et al. (2008) and the task sequencing list presentation proposed by
Kacem et al. (2002a) to form the chromosome.
A chromosome consists of two parts: (1) machine assignment
vector (called v1 ) and (2) operation sequence vector (called v2 ). A
permutation-based representation is used for operation sequence
vector representation. The length of each chromosome is equal to
the total number of operations, denoted as O. For each position
oð1 ro r OÞ in the chromosome, the machine assignment vector
v1 ðoÞ represents the machine selected for the operation on
position o, the element in operation sequence vector v2 ðoÞ
represents the index of operation in the oth position. The opera-
tions are scheduled by sequence as their positions in operation
sequence operator, e.g., the operation in the first operation is
scheduled first. Fig. 7 shows an example of chromosome repre-
sentation of an FJSP with four jobs, 12 operations and four
machines.
6.2.2. Population initialization
The algorithm starts with population initialization following
the approach by localization of Kacem et al. (2002a). The improved
assignment procedures presented in Pezzella et al. (2008) are
employed:
 Assignment Rule 1: search for the global minimum in the
processing time table.
 Assignment Rule 2: randomly permute jobs and machines in
the table.
In the experiments of this research, we employ the same assign-
ment initialization proportion as in Pezzella et al. (2008), i.e., 10%
of initial population will be generated by Assignment Rule 1 and
90% of it will be generated by Assignment Rule 2.
Once the assignments are settled up, the sequence of opera-
tions on the machine need to be determined. In the proposed
approach, mix of the following four commonly used rules is used
to select an operation from an eligible operation set: Longest
Processing Time (LPT), Most Work Remaining (MWR), Most
Operation Remaining (MOR), and Randomly Selection (RS).
6.3. Decoding
Since no-idle time insertion method is employed in this
research and the makespan is taken as one of the objectives,
chromosomes decoding follows an active decoding procedure
 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126
Fig. 7. An example of chromosome representation.
(Gao et al., 2007). In other words, no operation can be started
earlier without delaying any other operations or violating the
precedence constraints. With the consideration of stochastic
environment, the effects resulted from the uncertain factors
(machine breakdowns) are taken into account in decoding pro-
cedure. The right-shift policy is employed in schedule construc-
tion when there is any machine breakdown.
6.4. Genetic operators
6.4.1. Machine assignment genetic operators
The crossover operator for machine assignment is taken from
Kacem et al. (2002a). Two mutation operators are designed in the
proposed algorithm. The first operator, denoted as Assign-
ment_MO1, is based on Machine Based Mutation (MBM)
(Al-Hinai and ElMekkawy, 2011), where a random number of
operations are selected and reassigned to another machine.
However, the makespan of a schedule is invariant when moving
operations that are not on critical paths to other machines (Ho
and Tay, 2008). Thus, in Assignment_MO1, MBM is modified to
reassign the critical operation(s) in a random selected interval to
other machines. The second mutation operator for machine
assignment, denoted as Assignment_MO2, is based on the immi-
gration mutation presented in Gao et al. (2008). We modify it to
randomly generate new machine assignment for individuals
according to the accumulative knowledge OAMK.
6.4.2. Operation sequence genetic operators
We apply a two-point crossover introduced in Hartmann
(1998) for operation sequence. For each crossover, two crossing
points will be generated. Two children are constructed by inter-
changeably inheriting different parts from parents. In order to
preserve the precedence constraints, when inheriting the second
and third parts, all operations that are already in the previous
parts should be eliminated. The mutation operator for operation
sequence is adopted from Shadrokh and Kianfar (2007) and
modified as randomly permutating the position of the first critical
operation in a random selected interval from the operation
sequence.
7. Computational experiments
7.1. Benchmark problems
Although there are no standard benchmark problems for the
stochastic FJSP, as in Al-Hinai and ElMekkawy (2011), we can use
a number of deterministic FJSP benchmarks reported in existing
literature and modify them into the stochastic versions by
introducing perturbations on machine availabilities. The bench-
mark instances studied in this research include total flexibility
FJSSP (T-FJSP) and partial flexibility FJSSP (P-FJSP). We divided the
test instances into two sets: test set I and test set II. The test set I,
which was taken from Kacem et al. (2002a,b) and Li et al. (2011),
including 10 T-FJSP instances with different sizes varying
between 4  5 and 15  10. The first five instances of test set I
119
had no release date, while the release dates were considered in
the last five instances. The instances in test set I are represented as
case1–case10. The release dates of case6–case10 are given as
follows:
case 6 (4  5): r1 ¼3, r2 ¼5, r3 ¼1, r4 ¼ 6;
case 7 (8  8): r1 ¼2, r2 ¼ 5, r3 ¼8, r4 ¼8, r5 ¼1, r6 ¼9, r7 ¼4,
r8 ¼ 0;
case 8 (10  7): r1 ¼2, r2 ¼4, r3 ¼9, r4 ¼6, r5 ¼7, r6 ¼5, r7 ¼7,
r8 ¼ 4, r9 ¼1, r10 ¼0;
case 9 (10  10): r1 ¼2, r2 ¼4, r3 ¼9, r4 ¼6, r5 ¼7, r6 ¼5, r7 ¼7,
r8 ¼ 4, r9 ¼1, r10 ¼0;
case 10 (15  10): r1 ¼5, r2 ¼3, r3 ¼6, r4 ¼4, r5 ¼9, r6 ¼6, r7 ¼7,
r8 ¼ 2, r9 ¼9, r10 ¼0, r11 ¼14, r12 ¼13, r13 ¼11, r14 ¼ 12, r15 ¼5.
The second set of test instances, test set II, consists of 10
BRdata instances (Brandimarte, 1993) with job numbers ranging
from 10 to 20, machine numbers from 4 to 15 and the number of
operations for each job ranging from 5 to 15. The 10 instances in
test set II are denoted from MK01 to MK10.
7.2. Parameter settings
The proposed algorithm was coded and implemented in C# on
a Core(TM)2 Duo CPU 2.66 GHz. For the search model (multi-
objective genetic algorithm), the fine tuned parameters are listed
as follows: the population size ranged from 40 to 400 due to the
complexity of test instances, generation number 100, the percen-
tages of sequence generation rules: Long Processing Time (LPT)
30%, Most Work Remaining (MWR) 30%, Most Operation Remain-
ing (MOR) 30% and Randomly Selection (RS) 10%, machine assign-
ment crossover rate 0.75, machine assignment mutation rates
Assignment_MO10.2 and Assignment_MO20.2, operation sequence
crossover rate 0.9, operation sequence mutation rate 0.1. For the
knowledge model in the algorithm, every fifth generation, the last
20% solutions in the population were replaced. Furthermore, in
order to alleviate the effect resulted from stochastic nature of the
algorithm, each experiment was run repeatedly for 30 times with
different random seeds.
We generated two scenarios for machine breakdowns, scenario
I and scenario II. Similar with Al-Hinai and ElMekkawy (2011), in
the first scenario, machine breakdowns happened during the first
half of the scheduling horizon, while in the second scenario,
breakdowns happened in the second half. Table 2 shows the
parameters of the two scenarios.
7.3. Analysis of different robustness measures
In this section, performance of robustness measures within a
multi-objective framework is analyzed. By implementing the
algorithm with different robustness measures, five non-domi-
nated sets can be obtained. To investigate performance for
solutions in the non-dominated sets under machine breakdowns,
a simulation-based approach was employed to approximate the
realized non-dominated schedules. Robustness was indicated by
120 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126

the average value of 1000 samples, as follows: We employed the measure titled set cover (Deb, 2001),
P1000 denoted as SC, to perform a comparison among non-dominated
1
1000 i¼1M i ðSÞM 0 ðSÞ sets obtained with different robustness measures. Let A and B be
f2 ¼ ð13Þ
M 0 ðSÞ the two obtained non-dominated sets, such that SC(A,B) is defined
as the ratio of solutions in B, which are dominated by the
Then, each non-dominated set was updated by replacing the robust-
solutions in A. It is clear that a higher value of SC(A,B) indicates
ness objective with the simulation value as described in Eq. (13).
a better performance of set A. The value of SC(A,B)¼1 means that
all solutions in B are dominated by the solutions in A. For
Table 2
Machine breakdown scenarios. simplicity, we use SCðRM A ,RM B Þ to indicate the set cover value
of the two non-dominated sets obtained with robustness mea-
Breakdown scenarios b1 b2 a1 a2 sures RM A and RM B . Mean and variance of the set cover for two
scenarios of test set I are reported in Tables 3 and 4.
Scenario I 0.15 0.25 0 0.5
Scenario II 0.15 0.25 0.5 1
As shown in the tables, the simulation-based robustness
measure RM1 performed best among the five RMs. For all 10

Table 3
The values (mean, variance) of SC for the 10 instances of test set I in scenario I. (The mean value of SCðRMA ,RMB Þ is emphasized in boldface when the nondominated set
obtained with RMA is better than the one obtained with RMB .)

Instances RM RM1 RM2 RM3 RM4 RM5

Case1 RM1 \ (0.6727, 0.0773) (0.7658, 0.1018) (0.6467, 0.0722) (0.6741, 0.1018)
(4  5) RM2 (0.1563, 0.0265) \ (0.6978, 0.0896) (0.6139, 0.0704) (0.5013, 0.0711)
RM3 (0.0516, 0.0053) (0.1421, 0.0429) \ (0.0551, 0.0141) (0.1758, 0.0547)
RM4 (0.1968, 0.0276) (0.3481, 0.0590) (0.7848, 0.1043) \ (0.5532, 0.0633)
RM5 (0.1472, 0.0261) (0.4019, 0.0753) (0.6942, 0.1210) (0.4026, 0.0621) \

Case2 RM1 \ (0.9102, 0.0139) (0.9750, 0.0032) (0.9138, 0.0263) (0.7925, 0.0346)
(8  8) RM2 (0.0419, 0.0075) \ (0.8203, 0.0312) (0.2443, 0.0532) (0.3646, 0.0702)
RM3 (0.0000, 0.0000) (0.0965, 0.0135) \ (0.0276, 0.0063) (0.0582, 0.0121)
RM4 (0.0733, 0.0162) (0.6732, 0.0722) (0.9163, 0.0191) \ (0.5841, 0.0787)
RM5 (0.0827, 0.0104) (0.6225, 0.0845) (0.9232, 0.0147) (0.2919, 0.0539) \

Case3 RM1 \ (0.9527, 0.0065) (0.9933, 0.0013) (0.9578, 0.0123) (0.9075, 0.0123)
(10  7) RM2 (0.0312, 0.0027) \ (0.8055, 0.0334) (0.4265, 0.0691) (0.3774, 0.0678)
RM3 (0.0067, 0.0013) (0.1428, 0.0348) \ (0.0949, 0.0142) (0.0840, 0.0219)
RM4 (0.0236, 0.0043) (0.4400, 0.0609) (0.8166, 0.0468) \ (0.3208, 0.0631)
RM5 (0.0768, 0.0081) (0.5247, 0.0637) (0.8605, 0.0343) (0.4972, 0.0636) \

Case4 RM1 \ (0.9689, 0.0067) (0.9644, 0.0009) (0.9806, 0.0045) (0.9705, 0.0042)
(10  10) RM2 (0.0122, 0.0031) \ (0.8653, 0.0304) (0.2205, 0.0638) (0.3068, 0.0721)
RM3 (0.0067, 0.0013) (0.0543, 0.0253) \ (0.0272, 0.0076) (0.0348, 0.0120)
RM4 (0.0244, 0.0039) (0.5860, 0.0798) (0.9450, 0.0110) \ (0.4939, 0.0742)
RM5 (0.0171, 0.0027) (0.6063, 0.0922) (0.9183, 0.0227) (0.3721, 0.0847) \

Case5 RM1 \ (0.8035, 0.0442) (0.9958, 0.0005) (0.8242, 0.0303) (0.7535, 0.0491)
(15  10) RM2 (0.1243, 0.0152) \ (0.9791, 0.0009) (0.5432, 0.1061) (0.3821, 0.0900)
RM3 (0.0000, 0.0000) (0.0036, 0.0004) \ (0.0103, 0.0024) (0.0042, 0.0005)
RM4 (0.0808, 0.0098) (0.4177, 0.1135) (0.9792, 0.0012) \ (0.4106, 0.0912)
RM5 (0.1552, 0.0250) (0.5535, 0.0781) (0.9884, 0.0012) (0.5060, 0.0823) \

Case6 RM1 \ (0.5502, 0.2382) (1.0000, 0.0000) (0.6471, 0.1609) (0.5833, 0.2347)
(4  5) RM2 (0.1733, 0.0485) \ (1.0000, 0.0000) (0.4228, 0.1520) (0.6000, 0.2233)
RM3 (0.0000, 0.0000) (0.0000, 0.0000) \ (0.0483, 0.0196) (0.0000, 0.0000)
RM4 (0.1286, 0.0375) (0.4147, 0.2347) (0.8952, 0.0730) \ (0.4500, 0.2392)
RM5 (0.1300, 0.0260) (0.3343, 0.2156) (0.9952, 0.0007) (0.3700, 0.1256) \

Case7 RM1 \ (0.7016, 0.0871) (0.9402, 0.0064) (0.6593, 0.1041) (0.6587, 0.0874)
(4  5) RM2 (0.1335, 0.0271) \ (0.9608, 0.0071) (0.3947, 0.0868) (0.4133, 0.0820)
RM3 (0.0000, 0.0000) (0.0111, 0.0036) \ (0.0143, 0.0059) (0.0167, 0.0081)
RM4 (0.1652, 0.0236) (0.4212, 0.0916) (0.9198, 0.0193) \ (0.4717, 0.1053)
RM5 (0.1735, 0.0243) (0.4179, 0.0843) (0.9356, 0.0105) (0.3320, 0.0835) \

Case8 RM1 \ (0.8210, 0.0365) (1.0000, 0.0000) (0.8402, 0.0591) (0.6741, 0.0702)
(4  5) RM2 (0.0799, 0.0149) \ (0.9618, 0.0112) (0.4218, 0.0982) (0.2594, 0.0848)
RM3 (0.0000, 0.0000) (0.0108, 0.0017) \ (0.0083, 0.0020) (0.0000, 0.0000)
RM4 (0.0726, 0.0143) (0.4176, 0.1143) (0.9878, 0.0021) \ (0.2935, 0.0790)
RM5 (0.1419, 0.0177) (0.5453, 0.1173) (0.9778, 0.0143) (0.4910, 0.0592) \

Case9 RM1 \ (0.8205, 0.0317) (1.0000, 0.0000) (0.6348, 0.0843) (0.8371, 0.0456)
(4  5) RM2 (0.0693, 0.0161) \ (0.9683, 0.0039) (0.1290, 0.0370) (0.4320, 0.0988)
RM3 (0.0000, 0.0000) (0.0222, 0.0046) \ (0.0067, 0.0013) (0.0422, 0.0068)
RM4 (0.3251, 0.0813) (0.8104, 0.0823) (0.9867, 0.0052) \ (0.8239, 0.0636)
RM5 (0.0830, 0.0408) (0.5144, 0.1250) (0.9383, 0.0221) (0.1437, 0.0515) \

Case10 RM1 \ (0.8783, 0.0281) (1.0000, 0.0000) (0.7643, 0.0775) (0.9141, 0.0314)
(4  5) RM2 (0.0713, 0.0126) \ (0.9736, 0.0036) (0.2728, 0.0709) (0.4504, 0.1144)
RM3 (0.0000, 0.0000) (0.0179, 0.0056) \ (0.0000, 0.0000) (0.0113, 0.0018)
RM4 (0.1512, 0.0371) (0.6098, 0.1034) (1.0000, 0.0000) \ (0.6871, 0.1122)
RM5 (0.0728, 0.0167) (0.4467, 0.1045) (0.9745, 0.0041) (0.2462, 0.0832) \
 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126 121

Table 4
The values (mean, variance) of SC for the 10 instances of test set I in scenario II. (The mean value of SCðRM A ,RMB Þ is emphasized in boldface when the nondominated set
obtained with RMA is better than the one obtained with RMB .)

Instances RM RM1 RM2 RM3 RM4 RM5

Case1 RM1 \ (0.9543, 0.0080) (0.9383, 0.0152) (0.8308, 0.0418) (0.9356, 0.0198)
(4  5) RM2 (0.0378, 0.0058) \ (0.4944, 0.0907) (0.2232, 0.0543) (0.4348, 0.0586)
RM3 (0.0470, 0.0076) (0.3026, 0.0686) \ (0.1720, 0.0523) (0.3052, 0.0849)
RM4 (0.0778, 0.0179) (0.6918, 0.0531) (0.7206, 0.0604) \ (0.6492, 0.0628)
RM5 (0.0392, 0.0065) (0.3797, 0.0535) (0.4944, 0.1244) (0.2474, 0.0513) \

Case2 RM1 \ (0.8614, 0.0261) (0.9728, 0.0049) (0.8018, 0.0392) (0.9124, 0.0103)
(8  8) RM2 (0.0654, 0.0101) \ (0.9700, 0.0064) (0.2267, 0.0400) (0.4207, 0.0756)
RM3 (0.0000, 0.0000) (0.0048, 0.0007) \ (0.0114, 0.0019) (0.0212, 0.0038)
RM4 (0.1238, 0.0202) (0.6419, 0.0504) (0.9433, 0.0194) \ (0.6029, 0.0888)
RM5 (0.0545, 0.0103) (0.4331, 0.0613) (0.9344, 0.0231) (0.2873, 0.0645) \

Case3 RM1 \ (0.7802, 0.0364) (1.0000, 0.0000) (0.8463, 0.0219) (0.8214, 0.0242)
(10  7) RM2 (0.2844, 0.0514) \ (0.9431, 0.0059) (0.4679, 0.0792) (0.5152, 0.0495)
RM3 (0.0000, 0.0000) (0.0133, 0.0025) \ (0.0198, 0.0067) (0.0302, 0.0069)
RM4 (0.1939, 0.0306) (0.4137, 0.0714) (0.9581, 0.0082) \ (0.5212, 0.0723)
RM5 (0.2456, 0.0523) (0.4165, 0.0909) (0.9271, 0.0081) (0.4046, 0.0796) \

Case4 RM1 \ (0.9250, 0.0219) (1.0000, 0.0000) (0.9086, 0.0505) (0.9207, 0.0233)
(10  10) RM2 (0.0217, 0.0043) \ (0.9627, 0.0092) (0.2998, 0.0588) (0.2616, 0.0477)
RM3 (0.0000, 0.0000) (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0083, 0.0020)
RM4 (0.0364, 0.0066) (0.4591, 0.0764) (0.9469, 0.0024) \ (0.3754, 0.1097)
RM5 (0.0493, 0.0091) (0.4364, 0.1041) (0.9721, 0.0042) (0.3821, 0.0918) \

Case5 RM1 \ (0.7486, 0.0472) (1.0000, 0.0000) (0.7614, 0.0543) (0.7645, 0.0402)
(15  10) RM2 (0.1780, 0.0306) \ (0.9943, 0.0013) (0.3262, 0.0536) (0.3914, 0.0747)
RM3 (0.0000, 0.0000) (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.1576, 0.0323) (0.5494, 0.0922) (1.0000, 0.0000) \ (0.4816, 0.0951)
RM5 (0.2364, 0.0282) (0.4678, 0.0718) (1.0000, 0.0000) (0.3525, 0.0715) \

Case6 RM1 \ (0.9194, 0.0299) (0.9833, 0.0081) (0.6789, 0.1242) (0.8946, 0.0301)
(4  5) RM2 (0.0517, 0.0116) \ (0.9306, 0.0385) (0.1867, 0.0596) (0.4265, 0.0626)
RM3 (0.0067, 0.0013) (0.0278, 0.0073) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.1883, 0.0511) (0.7461, 0.0959) (1.0000, 0.0000) \ (0.7265, 0.0752)
RM5 (0.0656, 0.0130) (0.3961, 0.0628) (0.9606, 0.0052) (0.2211, 0.0474) \

Case7 RM1 \ (0.9023, 0.0132) (1.0000, 0.0000) (0.8763, 0.0345) (0.8532, 0.0423)
(4  5) RM2 (0.0669, 0.0082) \ (0.9033, 0.0216) (0.5262, 0.0739) (0.2984, 0.0788)
RM3 (0.0000, 0.0000) (0.0056, 0.0009) \ (0.0206, 0.0039) (0.0000, 0.0000)
RM4 (0.0584, 0.0114) (0.4289, 0.0619) (0.9063, 0.0641) \ (0.3659, 0.0530)
RM5 (0.0936, 0.0210) (0.5096, 0.1009) (0.9647, 0.0052) (0.5178, 0.0704) \

Case8 RM1 \ (0.9397, 0.0143) (0.9758, 0.0005) (0.8317, 0.0470) (0.9551, 0.0087)
(4  5) RM2 (0.0400, 0.0117) \ (0.9645, 0.0119) (0.2022, 0.0783) (0.3736, 0.0731)
RM3 (0.0000, 0.0000) (0.0067, 0.0013) \ (0.0056, 0.0009) (0.0133, 0.0025)
RM4 (0.1542, 0.0417) (0.6513, 0.1045) (0.9692, 0.0031) \ (0.7543, 0.0456)
RM5 (0.0312, 0.0078) (0.5323, 0.0720) (0.9314, 0.0074) (0.1614, 0.0492) \

Case9 RM1 \ (0.9844, 0.0023) (1.0000, 0.0000) (0.7777, 0.0769) (0.9711, 0.0098)
(4  5) RM2 (0.0000, 0.0000) \ (0.9878, 0.0021) (0.0422, 0.0197) (0.3992, 0.0783)
RM3 (0.0000, 0.0000) (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0048, 0.0007)
RM4 (0.1534, 0.0392) (0.9256, 0.0459) (1.0000, 0.0000) \ (0.9321, 0.0153)
RM5 (0.0114, 0.0021) (0.5344, 0.0963) (0.9814, 0.0031) (0.0400, 0.0108) \

Case10 RM1 \ (0.9591, 0.0215) (1.0000, 0.0000) (0.7813, 0.0801) (0.9550, 0.0244)
(4  5) RM2 (0.0250, 0.0061) \ (1.0000, 0.0000) (0.0984, 0.0402) (0.4125, 0.0758)
RM3 (0.0000, 0.0000) (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.1591, 0.0502) (0.8268, 0.0568) (1.0000, 0.0000) \ (0.8157, 0.0783)
RM5 (0.0321, 0.0107) (0.4764, 0.0933) (1.0000, 0.0000) (0.1463, 0.0585) \

instances, the obtained values SCðRM 1 ,RM x Þ, with x ¼ 2; 3,4; 5, are
significantly larger than the reverse figures SCðRM x ,RM1 Þ. For
instance, in case5, the value SCðRM 1 ,RM 2 Þ is 0.8035, while the
reverse figure SCðRM 2 ,RM 1 Þ is only 0.1243. By looking at the
figures related to RM3, it can be found that RM3 had the worst
performance for all test instances. In our study, the deterministic
makespan was employed as one of the optimization objectives
and remained constant once the baseline schedule was generated.
Thus, a higher value of set cover SC indicates a lower level of
makespan delay in the presence of machine breakdowns, suggest-
ing a better capability to cope with the uncertainty. Hence, it is
likely that for the FJSP, to optimize the robustness defined by
makespan or delay in a multi-objective framework, the simula-
tion-based method is the most appropriate, while the sum of free
slack method is unsuitable. However, as indicated earlier, the
simulation-based method is time intensive. To investigate com-
putational efforts for each robustness measure, we recorded the
average computational time for all five RMs. Table 5 shows the
results in terms of CPU time in seconds over 30 runs for the 10
instances. As shown in the table, RM1 consumed the most time,
while the other four RMs were comparable and time effective. In
our research, the sample size for RM1 was 30. In more complex
problems or uncertain scenarios, a larger sample size may be
needed that will result in a computational time which is
unacceptable.
With regards to RM2, RM4 and RM5, the two suggested RMs in
our research outperformed RM2 in both scenarios. In scenario I (see
Table 3), RM4 performed better than RM2 in four instances: case2,
122 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126

Table 5
Average CPU time (in seconds) for each RM.

Instances RM1 RM2 RM3 RM4 RM5

Case 1 3.82 0.76 0.78 0.80 0.79

Case 2 19.18 6.24 6.46 6.29 6.51
Case 3 18.72 6.30 6.61 6.28 6.29
Case 4 22.74 7.73 7.94 7.96 7.89
Case 5 111.02 35.53 39.04 36.23 35.24
Case 6 3.84 0.77 0.76 0.78 0.77
Case 7 19.17 6.01 6.40 6.32 6.13
Case 8 21.24 7.27 7.42 7.41 7.42
Case 9 53.27 14.91 15.49 13.61 13.78
Case 10 92.59 28.78 29.86 28.21 27.40

case4, case9 and case10, with values of SCðRM 4 ,RM 2 Þ being 0.6732,
0.5860, 0.8104 and 0.6098, respectively. RM2 outperformed RM4 in
only two instances: case1 and case, with values of SCðRM 2 ,RM 4 Þ
being 0.6139 and 0.5432, respectively. In the other four instances,
case3, case6, case7 and case8, the results obtained with RM2 and
RM4 were comparable. As to RM5 in scenario I, RM5 outperformed
RM2 in six instances, while RM2 outperformed RM5 in only two
instances. The results obtained in scenario II (see Table 4) were
similar to those in scenario I. In scenario II, RM4 and RM5 out-
performed RM2 in eight and six instances, respectively, while RM2
outperformed RM4 and RM5 both in only two instances. Comparing
the two suggested RMs, RM4 outperformed RM5 in seven and eight
instances with respect to scenario I and scenario II, respectively,
whereas RM5 outperformed RM4 in three and one instances with
respect to scenario I and scenario II. Hence, RM4 seems more
appropriate than RM5 for test set I.
Case5 in scenario I is an example illustrating the performance
of RM2 and RM4. Fig. 8 shows the obtained non-dominated
solutions of the problem (Eq. (7)) with respect to the surrogate
measures, RM2 and RM4. From the figure, seven non-dominated
solutions were obtained with RM2 (triangles) and five non-
dominated solutions with RM4 (dots). These solutions were
executed in the scenario with a simulation method. The objective
f2 of the problem was obtained with Eq. (13), which is indicated
by the vertical axis in Fig. 8. From the two non-dominated sets, for
each non-dominated solution obtained with RM2, it is dominated
by one or more solutions obtained with RM4. This finding
indicates that the non-dominated solutions obtained with RM4
had a smaller delay percentage in the presence of machine
breakdowns.
A more specific examination was performed on two solutions,
which had a minimum makespan in each non-dominated set. The
first solution was obtained with RM2 (denoted as SRM2 ) and had a
makespan of 11 and delay percentage of 0.3037, while the second
solution was obtained with RM4 (denoted as SRM4 ) and had a
makespan of 11 and delay percentage of 0.2795. The Ga tt charts
for SRM2 and SRM4 are provided in Figs. 9 and 10, respectively. As
shown in Fig. 8, SRM2 was dominated by SRM4 . However, from the
Ga tt charts of the two solutions, each solution had the same
makespan and total float time, with values of 11 and 27,
respectively. Thus, if the robustness was indicated by RM2
(average float time), SRM2 and SRM4 could have the same robust-
ness measures.
An examination of workload and float time distribution for
each machine may reveal the difference between the perfor-
mances of the two solutions in the presence of machine break-
downs. Table 6 shows the workload and float time for each
machine with solution SRM2 and SRM4 . Because it was assumed
that the machine breakdown probability was proportional to the
machine’s busy time which was also related to the workload of
Fig. 8. The obtained non-dominated solutions with RM2 (triangles) and RM4
(dots) for case5 in scenario I.
the machine, it was expected that the float time of an operation
that was assigned to a machine with a heavy workload was more
important than the time assigned to a machine with a light
workload. This concept was also applied for the surrogate mea-
sure RM4. From Table 6, for solution SRM2 , the largest three float
times were on machines M4, M8 and M1, with workloads of 9,
8 and 9, respectively. For the machines with larger workloads M2,
M5 and M7, the float times were only 2, 1, and 2, respectively. By
contrast, for solution SRM4 , the machines with heavy workloads,
such as M2, M9 and M10, had larger float times of 6, 5 and 5,
respectively. Thus, for a relatively low level of machine break-
downs, SRM4 performed better for absorbing the breakdowns than
SRM2 . However, when the level of machine breakdowns are high
enough, RM2 and RM4 may not work effectively. Moreover, a
non-buffering strategy was employed in our approach so the float
time on the machine(s) with the largest workload(s) is(are) often
equal to 0, e.g., M6 and M10 in SRM2 and M1 in SRM4 . For this
reason, we did not employ the max–min of slack that was
suggested as by Kobylański and Kuchta (2007) for the S-FJSP.
To further investigate the performance of different RMs, we
ran the algorithm for test set II. As stated above, RM1 had the best
performance with the largest computational cost. In experiments
for test set II, we only compared the performances among the four
surrogate measures (RM2, RM3, RM4 and RM5). Tables 7 and 8
include results obtained with respect to scenario I and scenario II.
Similar to test set I, RM3 performed worst on test set II, while the
performance of RM4 significantly decreased on test set II. In
scenario I, RM4 was outperformed by RM2 in six instances,
MK02, MK03, MK04, MK08, MK09 and MK10, with the mean
value of SCðRM 2 ,RM 4 Þ being 0.5408, 0.5612, 0.4809, 0.5417,
0.5629 and 0.5129, respectively. The number of instances in
which RM4 performed better that RM2 decreased to 2, MK05
and MK06, with mean value of SCðRM 4 ,RM 5 Þ being 0.4203 and
0.5718. The performance of RM4 for test set II also decreased in
scenario II, where RM2 outperformed RM4 in four instances,
MK03, MK04, MK06 and MK10; RM4 also outperformed RM2 in
four instances, MK01, MK05, MK07 and MK08. This degradation
of performance for RM4 may be attributed to the increase in
instance size and uncertainty (machine breakdowns) level. In the
presence of a relatively large number of machine breakdowns, the
float time on a machine with high breakdown probability was not
capable of absorbing the perturbations. Thus, the consideration of
 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126 123

Fig. 9. The obtained solution for RM2 with a makespan (f1) of 11 and a delay percentage (f2) of 0.3037.

Fig. 10. The obtained solution for RM4 with a makespan (f1) of 11 and a delay percentage (f2) of 0.2795.
124 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126

breakdown probability did not work effectively for large cases ad
it did for small cases.
From figures related to RM5 reported in Tables 7 and 8, RM5
still performed better for test set II. In scenario I (see Table 7), RM5
outperformed RM2 in six instances, MK01, MK03, MK05, MK06,
MK08 and MK10, while RM2 only outperformed RM5 in two
instances, MK04 and MK07. The set of figures in scenario II (see
Table 8) was similar to the set from scenario I, where RM5
outperformed RM2 in eight instances, MK01, MK02, MK03,
MK05, MK06, MK07, MK09 and MK10, while the reverse figure
occurred in only two instances, MK04 and MK08. The comparison
between RM5 and RM4 was consistent with the results
mentioned above. RM5 outperformed RM4 in seven instances in
scenarios I and II, while the reverse figures were 2 and 3,
respectively.
Based on the above discussion, the first suggested surrogate
measure RM4, which considers the machine breakdown prob-
abilities, is suitable for small cases with a relatively low level of
breakdowns, while it is not suitable for relatively large-scale
cases. Considering the locations of both breakdowns and float
times, the second suggested surrogate measure, RM5, performed
better for flexible job-shop scheduling with single breakdowns on
each machine with simultaneous consideration of makespan and
quality robustness.

Table 6
Total float times and workload on each machine. 8. Conclusions and future work

Schedules Machine M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 In real-world applications, the schedule execution of the FJSP is
SRM2 Workload 9 10 7 9 10 11 10 8 10 11
usually confronted with uncertain events. In this study, we analyzed
Float times 5 2 1 6 1 0 2 6 4 0 the robustness of the FJSP with random machine breakdowns. With
simultaneous consideration of makespan and robustness, the pro-
SRM4 Workload 11 10 7 9 10 10 10 8 10 10
Float times 0 6 1 2 3 0 3 2 5 5 blem was modeled as a multi-objective optimization problem.
The concept of quality robustness was employed to measure the

Table 7
The values (mean, variance) of SC for test set II in scenario I (The mean value of SCðRMA ,RMB Þ is emphasized in boldface when the nondominated set obtained with RMA is
better than the one obtained with RMB .)

Instances RM RM2 RM3 RM4 RM5

MK01 RM2 \ (0.7709, 0.0130) (0.4231, 0.0764) (0.3826, 0.0624)

RM3 (0.0780, 0.0115) \ (0.0479, 0.0124) (0.0647, 0.0112)
RM4 (0.4643, 0.0676) (0.8587, 0.0205) \ (0.4801, 0.0561)
RM5 (0.5072, 0.0683) (0.7861, 0.0142) (0.4143, 0.0821) \

MK02 RM2 \ (0.9722, 0.0042) (0.5408, 0.0463) (0.4403, 0.0412)

RM3 (0.0017, 0.0001) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.3360, 0.0532) (0.9764, 0.0026) \ (0.3151, 0.0498)
RM5 (0.4741, 0.0391) (0.9690, 0.0043) (0.5813, 0.0564) \

MK03 RM2 \ (0.9236, 0.0051) (0.5612, 0.0761) (0.4114, 0.0897)

RM3 (0.0044, 0.0006) \ (0.0033, 0.0003) (0.0053, 0.0004)
RM4 (0.3043, 0.0577) (0.8992, 0.0056) \ (0.2812, 0.0543)
RM5 (0.4758, 0.0895) (0.9254, 0.0049) (0.6299, 0.0823) \

MK04 RM2 \ (0.9435, 0.0258) (0.4809, 0.0938) (0.4683, 0.0922)

RM3 (0.0191, 0.0026) \ (0.0342, 0.0071) (0.0132, 0.0013)
RM4 (0.3127, 0.0714) (0.9177, 0.0186) \ (0.2865, 0.0835)
RM5 (0.3878, 0.1123) (0.9544, 0.0104) (0.5665, 0.1075) \

MK05 RM2 \ (0.9703, 0.0023) (0.3724, 0.0654) (0.3790, 0.0525)

RM3 (0.0000, 0.0000) \ (0.0021, 0.0001) (0.0000, 0.0000)
RM4 (0.4203, 0.0708) (0.9464, 0.0042) \ (0.3905, 0.0721)
RM5 (0.4135, 0.0691) (0.978, 0.0016) (0.4175, 0.0892) \

MK06 RM2 \ (0.4669, 0.1071) (0.3097, 0.0728) (0.3177, 0.0849)

RM3 (0.3025, 0.0632) \ (0.2387, 0.0442) (0.2269, 0.0571)
RM4 (0.5718, 0.0769) (0.5207, 0.0876) \ (0.3424, 0.0707)
RM5 (0.5213, 0.0992) (0.5709, 0.1121) (0.4562, 0.0922) \

MK07 RM2 \ (0.8807, 0.0078) (0.4390, 0.0668) (0.4982, 0.0530)

RM3 (0.0253, 0.0014) \ (0.0102, 0.0005) (0.0211, 0.0013)
RM4 (0.4334, 0.0578) (0.9080, 0.0068) \ (0.5137, 0.0588)
RM5 (0.3994, 0.0403) (0.8387, 0.0077) (0.3961, 0.0544) \

MK08 RM2 \ (0.8928, 0.0144) (0.5417, 0.0683) (0.3409, 0.0855)

RM3 (0.0000, 0.0000) \ (0.0042, 0.0002) (0.0000, 0.0000)
RM4 (0.3199, 0.0582) (0.8970, 0.0154) \ (0.2720, 0.0469)
RM5 (0.4629, 0.0839) (0.9183, 0.0093) (0.5729, 0.0608) \

MK09 RM2 \ (0.4134, 0.0923) (0.5629, 0.1024) (0.4292, 0.1153)

RM3 (0.3172, 0.0431) \ (0.4274, 0.0412) (0.3363, 0.0612)
RM4 (0.2523, 0.0753) (0.2528, 0.0527) \ (0.2476, 0.0727)
RM5 (0.4616, 0.1257) (0.4258, 0.0895) (0.5646, 0.1002) \

MK10 RM2 \ (0.9113, 0.0312) (0.5129, 0.1132) (0.3772, 0.0531)

RM3 (0.0244, 0.0081) \ (0.0543, 0.0085) (0.0351, 0.0046)
RM4 (0.3143, 0.0814) (0.7704, 0.0702) \ (0.2741, 0.0861)
RM5 (0.4421, 0.0825) (0.8281, 0.0221) (0.5709, 0.1187) \
 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126 125

Table 8
The values (mean, variance) of SC for test set II in scenario II (The mean value of SCðRM A ,RMB Þ is emphasized in boldface when the nondominated set obtained with RMA is
better than the one obtained with RMB .)

Instances RM RM2 RM3 RM4 RM5

MK01 RM2 \ (0.9848, 0.0005) (0.2673, 0.0979) (0.3143, 0.1392)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.5247, 0.1402) (1.0000, 0.0000) \ (0.4881, 0.1548)
RM5 (0.4931, 0.1210) (0.9952, 0.0005) (0.3396, 0.1016) \

MK02 RM2 \ (1.0000, 0.0000) (0.4229, 0.0581) (0.4013, 0.0737)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.4505, 0.0523) (1.0000, 0.0000) \ (0.3763, 0.0396)
RM5 (0.4838, 0.0697) (1.0000, 0.0000) (0.4677, 0.046) \

MK03 RM2 \ (1.0000, 0.0000) (0.5619, 0.1083) (0.3090, 0.0689)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.3507, 0.0885) (1.0000, 0.0000) \ (0.2268, 0.0733)
RM5 (0.5723, 0.0812) (1.0000, 0.0000) (0.7142, 0.1025) \

MK04 RM2 \ (0.9917, 0.0013) (0.5512, 0.0613) (0.4479, 0.0692)

RM3 (0.0000, 0.0000) \ (0.0182, 0.0063) (0.0000, 0.0000)
RM4 (0.3177, 0.0553) (0.9732, 0.0059) \ (0.3046, 0.0690)
RM5 (0.4047, 0.0805) (0.9875, 0.003) (0.5424, 0.1055) \

MK05 RM2 \ (1.0000, 0.0000) (0.3936, 0.0700) (0.3270, 0.0456)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.4415, 0.0664) (1.0000, 0.0000) \ (0.3538, 0.0511)
RM5 (0.4994, 0.0723) (1.0000, 0.0000) (0.5129, 0.0680) \

MK06 RM2 \ (0.9833, 0.0081) (0.4351, 0.0757) (0.3511, 0.1002)

RM3 (0.0095, 0.0026) \ (0.0067, 0.0013) (0.0095, 0.0026)
RM4 (0.3591, 0.0722) (0.9893, 0.0013) \ (0.3129, 0.0689)
RM5 (0.4879, 0.1054) (0.9602, 0.0185) (0.5080, 0.0992) \

MK07 RM2 \ (1.0000, 0.0000) (0.3471, 0.0365) (0.3836, 0.0585)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.5112, 0.0421) (1.0000, 0.0000) \ (0.5067, 0.0651)
RM5 (0.4619, 0.0546) (1.0000, 0.0000) (0.3712, 0.0683) \

MK08 RM2 \ (0.8000, 0.1433) (0.2688, 0.0570) (0.4442, 0.0722)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.5277, 0.0860) (0.9289, 0.0622) \ (0.5816, 0.1140)
RM5 (0.3942, 0.0833) (0.8125, 0.1404) (0.2778, 0.0902) \

MK09 RM2 \ (1.0000, 0.0000) (0.3862, 0.0725) (0.2847, 0.0271)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.4104, 0.0803) (1.0000, 0.0000) \ (0.3113, 0.0689)
RM5 (0.5418, 0.0477) (1.0000, 0.0000) (0.5043, 0.0702) \

MK10 RM2 \ (1.0000, 0.0000) (0.4875, 0.0881) (0.3713, 0.0740)

RM3 (0.0000, 0.0000) \ (0.0000, 0.0000) (0.0000, 0.0000)
RM4 (0.3309, 0.0756) (1.0000, 0.0000) \ (0.2702, 0.0613)
RM5 (0.4471, 0.0832) (1.0000, 0.0000) (0.5611, 0.0725) \

schedules’ ability to cope with breakdowns and was indicated by the developing more efficient meta-heuristics for large-scale pro-
relative difference between deterministic and actual makespans. To blems. The comparison of our proposed algorithm for the S-FJSP
utilize the available information about machine breakdowns in the with other meta-heuristics, especially nonevolutionary algo-
generation of baseline schedules, we developed two surrogate rithms, will be addressed in our future work.
measures for robustness, RM4 and RM5. Both suggested surrogate
measures were based on float times of the operations. Specifically,
the first suggested surrogate measure, RM4, considered the break-
Acknowledgments
down probability of each machine, while the second suggested
surrogate measure, RM5, considered the locations of breakdowns
This research was supported in part by the China Scholarship
and float times. The experimental results suggested that both RM4
Council and National Natural Science Foundation of China under
and RM5 outperformed other existing surrogate measures to a
the Contract numbers 70971131, 71101150, 70901074 and
certain extent. Furthermore, a multi-objective evolutionary algo-
71001104.
rithm based on our previous research was proposed to address the
multi-objective optimization problem.
Several aspects are of interest for future work. First, the
References
uncertainty scenario can be extended to include other uncertain
factors, such as perturbations of processing times and arrivals/
Abbass, H.A., Bender, A., Dam, H.H., Baker, S., Whitacre, J., Sarker, R., 2008.
cancellations of operations. Second, the measure of schedule Computational scenario-based capability planning. In: Proceedings of the
stability can be investigated along with robustness. In this case, 10th Annual Conference on Genetic and Evolutionary Computation. New York,
appropriate surrogate measures need to be designed to develop NY, USA, pp. 1437–1444.
Al-Fawzan, M.A., Haouari, M., 2005. A bi-objective model for robust resource-
schedules with strong levels for both robustness and stability constrained project scheduling. International Journal of Production Economics
measures. Another interesting direction for future work involves 96, 175–187.
126 J. Xiong et al. / Int. J. Production Economics 141 (2013) 112–126
Al-Hinai, N., ElMekkawy, T.Y., 2011. Robust and stable flexible job shop scheduling
with random machine breakdowns using hybrid genetic algorithm. Interna-
tional Journal of Production Economics 132, 279–291.
Aytug, H., Lawley, M.A., McKay, K., Moha, S., Uzsoy, R., 2005. Executing production
schedules in the face of uncertainties: a review and some future directions.
European Journal of Operational Research 161, 86–110.
Brandimarte, P., 1993. Routing and scheduling in a flexible job shop by tabu
search. Annals of Operations Research 41, 157–183.
Bruker, P., Schlie, R., 1990. Job shop scheduling with multi-purpose machines.
Computing 45, 369–375.
Chtourou, H., Haouari, M., 2008. A two-stage-priority-rule-based algorithm for
robust resource-constrained project scheduling. Computers Industrial Engi-
neering 55 (1), 183–194.
Chaari, T., Chaabane, S., Loukil, T., Trentesaux, D., 2011. A genetic algorithm for
robust hybrid flow shop scheduling. International Journal of Computer
Integrated Manufacturing 24 (9), 821–833.
Daniels, R.L., Kouvelis, P., 1995. Robust scheduling to hedge against processing
time uncertainty in single-stage production. Management Science 41 (2),
363–376.
de Vonder, S.V., Demeulemeester, E., Herroelen, W., 2008. Proactive heuristic
procedures for robust project scheduling: an experimental analysis. European
Journal of Operational Research 189 (3), 723–733.
Deb, K., 2001. Multi-objective Optimization using Evolutionary Algorithms. John
Wiley & Sons, New York.
Deb, K., Pratap, A., Agarwal, S., Meyarivan, T., 2002. A fast and elitist multiobjective
genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation 6
(2), 182–197.
Dong, Y.-H., Jang, J. Production rescheduling for machine breakdown at a job shop.
International Journal Production Research, in press, http://dx.doi.org/10.1080/
00207543.2001.579637.
Gao, J., Sun, L., Gen, M., 2008. A hybrid genetic and variable neighborhood descent
algorithm for flexible job shop scheduling problems. Computers & Operations
Research 35, 2892–2907.
Gao, J., Sun, L., Gen, M., Zhao, X., 2007. A hybrid of genetic algorithm and
bottleneck shifting for multiobjective flexible job shop scheduling problems.
Computers & Industrial Engineering 53, 149–162.
Garey, M.R., Johnson, D.S., Sethi, R., 1976. The complexity of flowshop and jobshop
scheduling. Mathematics of Operations Research 1 (2), 117–129.
Gholami, M., Zandieh, M., Alem-Tabriz, A., 2009. Scheduling hybrid flow shop with
sequence-dependent setup times and machines with random breakdowns.
The International Journal of Advanced Manufacturing Technology 42 (1),
189–201.
Goren, S., Sabuncuolu, I., 2008. Robustness and stability measures for scheduling:
single-machine environment. IIE Transactions 40 (1), 66–83.
Gu, J., Gu, M., Cao, C., Gu, X., 2010. A novel competitive co-evolutionary quantum
genetic algorithm for stochastic job shop scheduling problem. Computers &
Operations Research 37, 927–937.
Hartmann, S., 1998. A competitive genetic algorithm for resource-constrained
project scheduling. Naval Research Logistics 45, 733–750.
Hazir, O., Haouari, M., Erel, E., 2010. Robust scheduling and robustness measures
for the discrete time/cost trade-off problem. European Journal of Operational
Research 207 (2), 633–643.
Herroelen, W., Leus, R., 2005. Project scheduling under uncertainty: survey and
research potentials. European Journal of Operational Research 165, 289–306.
Ho, N.B., Tay, J.C., 2008. Solving multi-objective flexible job shop problems by
evolution and local search. IEEE Transactions on Systems Man and Cybernetics
Part C – Applications and Reviews 38 (5), 674–685.
Holthaus, O., 1999. Scheduling in job shops with machine breakdowns: an
experimental study. Computers & Industrial Engineering 36 (1), 137–162.
Jensen, M.T., 2003. Generating robust and flexible job shop schedules using genetic
algorithms. IEEE Transactions on Evolutionary Computation 7 (3), 275–288.
Jin, Y., Branke, J., 2005. Evolutionary optimization in uncertain environments
survey. IEEE Transactions on Evolutionary Computation 9 (3), 303–317.
Kacem, I., Hammadi, S., Borne, P., 2002a. Approach by localization and multi-
objective evolutionary optimization for flexible job-shop scheduling problems.
IEEE Transactions on Systems, Man, and Cybernetics Part C: Applications and
Reviews 32 (1), 1–13.
Kacem, I., Hammadi, S., Borne, P., 2002b. Pareto-optimality approach for flexible
job-shop scheduling problems: hybridization of evolutionary algorithms and
fuzzy logic. Mathematics and Computers in Simulation 60, 245–276.
Kobylański, P., Kuchta, D., 2007. A note on the paper by M. A. Al-fawzan and M.
Haouari about a bi-objective model for robust resource-constrained project
scheduling. International Journal of Production Economics 107, 496–501.
Lambrechts, O., Demeulemeester, E., Herroelen, W., 2008. A tabu search procedure
for developing robust predictive project schedules. International Journal of
Production Economics 111, 493–508.
Lei, D., 2010. A genetic algorithm for flexible job shop scheduling with fuzzy
processing time. International Journal Production Research 48 (10),
2995–3013.
Lei, D., 2011. Scheduling stochastic job shop subject to random breakdown to
minimize makespan. International Journal of Advanced Manufacturing Tech-
nology 55, 1183–1192.
Leon, V.J., Wu, S.D., Storer, R.H., 1994. Robustness measures and robust scheduling
for job shops. IIE Transactions 26 (5), 32–43.
Leung, S., Yue, W., 2004. A robust optimization model for stochastic aggregate
production planning. Production Planning & Control 15 (5), 502–514.
Li, J.-Q., Pan, Q.-K., Gao, K.-Z., 2011. Pareto-based discrete artificial bee colony
algorithm for multi-objective flexible job shop scheduling problems. Interna-
tional Journal of Advanced Manufacturing Technology 55, 1159–1169.
Liao, C.-J., Chen, W.J., 2003. Single-machine scheduling with periodic maintenance
and nonresumable jobs. Computers & Operations Research 30, 1335–1347.
Lin, C.-H., Liao, C.-J., 2007. Makespan minimization for two parallel machines with
an unavailable period on each machine. International Journal of Advanced
Manufacturing Technology 33, 1024–1030.
Liu, L., Gu, H.-Y, Xi, Y.-G, 2007. Robust and stable scheduling of a single machine
with random machine breakdowns. International Journal of Advanced Manu-
facturing Technology 31, 645–654.
Mahdavi, I., Shirazi, B., Solimanpur, M., 2010. Development of a simulation-based
decision support system for controlling stochastic flexible job shop manufac-
turing systems. Simulation Modelling Practice and Theory 18, 768–786.
Mehta, S.V., Uzsoy, R.M., 1998. Predictable scheduling of a job shop subject to
breakdowns. IEEE Transactions on Robotics and Automation 14, 365–378.
Mehta, S.V., Uzsoy, R.M., 1999. Predictable scheduling of a single machine subject
to breakdowns. International Journal of Computer Integrated Manufacturing
12, 15–38.
Moradi, E., Ghomi, S.M.T.F., Zandich, M., 2011. Bi-objective optimization research
on integrated fixed time interval preventive maintenance and production for
scheduling flexible job-shop problem. Expert Systems with Application 38,
7169–7178.
O’Donovan, R., Uzsoy, R.M., McKay, K.N., 1999. Predictable scheduling of a single
machine with breakdowns and sensitive jobs. International Journal of Produc-
tion Research 37, 4217–4233.
Pezzella, F., Morganti, G., Ciaschetti, G., 2008. A genetic algorithm for the flexible
job-shop scheduling problem. Computers & Operations Research 35,
3202–3212.
Shadrokh, S., Kianfar, F., 2007. A genetic algorithm for resource investment project
scheduling problem, tardiness permitted with penalty. European Journal of
Operational Research 181, 86–101.
Tseng, C.-T., Liao, C.-J., Huang, K.-L., 2009. Minimizing total tardiness on a single
machine with controllable processing times. Computers & Operations
Research 36, 1852–1858.
Wang, K., Choi, S.H. A decomposition-based approach to flexible flow shop
scheduling under machine breakdown. International Journal of Production
Research, in press, http://dx.doi.org/10.1080/00207543.2011.571456.
Wang, S., Yu, J., 2010. An effective heuristic for flexible job-shop scheduling
problem with maintenance activities. Computers & Industrial Engineering 59,
436–447.
Wu, S.D., Byeon, E., Storer, R.H., 1999. A graph-theoretic decomposition of the job
shop scheduling problem to achieve scheduling robustness. Operations
Research 47 (1), 113–124.
Xia, Y., Chen, B., Yue, J., 2008. Job sequencing and due date assignment in a single
machine shop with uncertain processing times. European Journal of Opera-
tional Research 184, 63–75.
Xing, L., Chen, Y., Wang, P., Zhao, Q., Xiong, J., 2010. A knowledge-based ant colony
optimization for flexible job shop scheduling problems. Applied Soft Comput-
ing 10 (3), 888–896.
Xiong, J., Liu, J., Chen, Y., Abbass, H.A., 2011. An evolutionary multi-objective
scenario-based approach for stochastic resource investment project schedul-
ing. In: 2011 IEEE Congress on Evolutionary Computation (CEC), New Orleans,
LA, USA, pp. 2767–2774.