Computers & Industrial Engineering 112 (2017) 537–550

Contents lists available at ScienceDirect

Computers & Industrial Engineering

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

Resource constrained project scheduling with uncertain activity

durations
Ripon K. Chakrabortty ⇑, Ruhul A. Sarker, Daryl L. Essam
School of Engineering and Information Technology, University of New South Wales, Canberra 2600, Australia

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

Article history: In this paper, we consider Resource Constrained Project Scheduling Problems (RCPSPs) with known deter-
Available online 3 January 2017 ministic renewable resource requirements but uncertain activity durations. In this case, the activity dura-
tions are represented by random variables with different probability distribution functions. To deal with
Keywords: this problem, we propose an approach based on the robust optimization concept, which produces reason-
RCPSP ably good solutions under any likely input data scenario. Depending on different uncertainty character-
SRCPSP istics, we have developed six different heuristics to incorporate the uncertain duration as a deterministic
Branch and cut algorithm
constraint in a robust optimization model. The resulting optimization model is then solved by using a
Robust optimization
Heuristics
Coin-Branch & Cut (CBC) solver. To judge the performance of the algorithm, we solved 30, 60, 90 and
120-activity benchmark problems from the project scheduling problem library (PSPLIB). Our proposed
approach guarantees the feasibility of solutions and produces high-quality solutions, particularly for lar-
ger activity instances, compared to other existing approaches.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction reported some important merits and demerits of different

state-of-the-art exact approaches (i.e. mixed integer linear pro-
The resource constrained project scheduling problem (RCPSP) gramming models-MILP) and concluded that the exact approaches
aims to minimize the duration of a project, subject to satisfying are not competitive enough for higher activity instances for
the precedence and resource availability constraints. The RCPSP RCPSPs. Relevant surveys on different heuristic and exact
assumes complete information of resource usage and activity dura- approaches for RCPSPs can be found from Hartmann and Briskorn
tion, and determines a feasible baseline schedule which encom- (2010), Huang, Chen, and Hu (2013) and Zhou, Love, Wang, Teo,
passes a list of activity starting times while minimizing the and Irani (2013). Among a few exact approaches, Zhu, Bard, and
makespan (Bruni, Beraldi, Guerriero, & Pinto, 2011). For many Yu (2006) presented an exact branch and cut procedure for solving
years, the RCPSP has been of interest for practitioners and RCPSP directly while using an integer linear programming (ILP)
researchers. There has been a tremendous increase in research model. However, in their paper, they only considered 30 and 60
for the RCPSP, both in terms of heuristic and optimal procedures activity instances from the project scheduling problems library
over the last few decades. Among them, Hartmann and Kolisch (PSPLIB). Therefore, B&B or B&C for higher instances is still an open
(2000) considered an in-depth literature survey on state-of-the- area to perform research.
art heuristics and analysed the behaviour of different heuristics Because of a growing awareness that traditional deterministic
with respect to their components, such as priority rules and project scheduling models are in plain conflict with a reality that
meta-heuristic strategy. Apart from that, from a modelling per- is characterized by uncertainty, a recent research track focuses
spective, RCPSPs can be formulated by following discrete or contin- on the stochastic resource constrained project scheduling problem
uous time (i.e., even based) depending upon the number of (SRCPSP). The SRCPSP is an extension of the RCPSP that involves
variables (Chakrabortty, Sarker, & Essam, 2014). Alongside of that, the minimization of the expected makespan of a project with
very recently Koné, Artigues, Lopez, and Mongeau (2013) have stochastic activity durations (problem m; 1jcpm; di jEðC max Þ). As the
SRCPSP tries to incorporate uncertainty in the formulation itself
by expressing the random parameters of the problems as random
⇑ Corresponding author.
variables, the resulting stochastic problem is challenging from both
theoretical and computational point of view (Bruni, Beraldi, &
E-mail addresses: [email protected], ripon_ipebuet@
yahoo.com (R.K. Chakrabortty). Guerriero, 2015). Despite its applicability, the literature concerning

http://dx.doi.org/10.1016/j.cie.2016.12.040
0360-8352/Ó 2017 Elsevier Ltd. All rights reserved.
538 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550

SRCPSPs is virtually void (van de Vonder, Ballestin, durations and renewable/non-renewable resource constraints is a
Demeulemeester, & Herroelen, 2007). completely new and unique contribution of this work.
Uncertainties during project execution may lead to considerable Artigues et al. (2013) proposed a few project scheduling models
schedule disruptions. This uncertainty may arise from a number of with considerable uncertainty in their activity durations, for when
possible sources: activities may take more or less time than origi- a decision maker does not have sufficient confidence in the subjec-
nally estimated (i.e., activity disruptions), resources may become tive probabilities that can be attributed to different duration sce-
unavailable (i.e., resource disruptions), material may arrive behind narios. They developed a scenario-relaxation algorithm, based on
schedule, ready times and due dates may have to be changed, new the robust optimization concept, and proposed two heuristic pro-
activities may have to be incorporated or activities may have to be cedures for solving up to medium sized instances. Our work differs
dropped due to changes in the project scope, weather conditions from this line of work in two aspects. Firstly, our approach is cap-
may cause several delays, etc. (Herroelen & Leus, 2005). Existing able of solving not only small scale problems but also large activity
research in project scheduling has considered activity disruptions instances. Secondly, instead of using either mean durations or
(durational variability for activities) (Artigues, Leus, & Nobibon, omitting them for heuristic procedures, we propose six different
2013; Bruni et al., 2011; Deblaere, Demeulemeester, & Herroelen, heuristic models that depend on different possible uncertainty
2011a; Lamas & Demeulemeester, 2014; Ma, Che, Huang, & Ke, levels, and solve them with the proposed algorithm. The contribu-
2015; van de Vonder, Demeulemeester, & Herroelen, 2008), tions in this paper are threefold: (i) depending on various forms of
resource disruptions (due to breakdowns) (Chakrabortty, Sarker, parametric uncertainty levels, we have developed six different
& Essam, 2016; Lambrechts, Demeulemeester, & Herroelen, 2007, heuristics to incorporate the uncertain activity durations as deter-
2010) or even combination of both activity and resource disrup- ministic constraints in a robust optimization model; (ii) we pro-
tions (Deblaere, Demeulemeester, & Herroelen, 2011b; Fu, Lau, & pose a robust optimization based approach for solving resource
Varakantham, 2015). Although the sources of variability in the pro- constrained project scheduling under durational uncertainties;
ject environment are diversiform, the main scheduling objectives and (iii) we apply a distinct branch-and-cut algorithm (B&C) for
are mostly pertinent to the activities’ starting (or ending) times, solving those robust optimization models for stochastic-RCPSPs.
with project makespan being the single most studied objective Primarily, the focus of this paper is on the application of a newly
(Artigues et al., 2013). This rationalizes a restriction to the study defined branch-and-cut algorithm (B&C) for solving deterministic
of uncertainty in activity processing times or durations only, RCPSP models particularly for those which cannot be solved by
although different sources may be the basis of this uncertainty. using traditional exact approaches. In this case, a constrained logic
Even though a significant amount of theoretical work has programming (CLP) technique was employed with a branch-and-
already been done to conceptualize and measure uncertainty in cut algorithm to satisfy the constraints and to facilitate a comput-
project scheduling, still the most proficient managers have diffi- ing capability. The deterministic RCPSP model was solved by using
culty in handling them. The lack of practitioner’s knowledge on a commercial optimization solver from the OPTI toolbox that
adequate answers to uncertainty in RCPSP is surrounded by a rel- comes with Matlab (R2012b). The proposed algorithm is capable
atively scant scientific interest for this hot topic. As a matter of fact, of dealing with large activity instances of RCPSPs within a reason-
whilst RCPSPs have been widely investigated in the academic liter- able computing time (1000 s). To demonstrate the performance of
ature, the issue of the incorporation of uncertainty in project B&C algorithm, the benchmark instances with 30, 60, 90 and 120
scheduling has received growing attention in the last decade activities from the PSPLIB were solved. The solutions were com-
(Bruni et al., 2015). The uncertainty that arises during project exe- pared with traditional heuristic and exact approaches for some
cution can be explicitly taken into account through preventive randomly chosen 30-activity instances. Later, we studied the
approaches, such as two-stage stochastic programming, paramet- stochastic-RCPSP with the possibility of activity duration uncer-
ric programming, fuzzy programming, chance constraint program- tainty. Depending on different durational uncertainty levels/forms,
ming and robust optimization techniques (Ye et al., 2014). Since for we propose six different heuristic procedures to determine the
robust optimization approaches most of the material on scheduling realized activity durations and to incorporate the deterministic
policies developed for stochastic scheduling can be transferred constraints into the robust optimization model. To develop differ-
without major alterations, and the whole uncertainty space can ent deterministic robust counterpart optimization models for
be easily addressed instead of using parametric scenarios, the those uncertain activity durations, we then employed a robust
authors restrict this paper work to that approach only. optimization framework (Janak, Lin, and Floudas (2007), Li, Ding,
The term robust optimization has come to encompass several and Floudas (2011), Li and Floudas (2014), Lin, Janak, and
approaches to protecting the decision maker against parameter Floudas (2004)). Finally, those robust optimization models were
ambiguity and stochastic uncertainty (Gabrel, Murat, & Thiele, solved by using our B&C algorithm for different benchmark
2014). instances from PSPLIB. The solution from the robust optimization
Because of its applicability, robust optimization has been framework was guaranteed to be feasible for the whole space of
applied in diversified areas such as classical logistic problems the uncertain duration parameters.
(Remli & Rekik, 2013), facility location problems (Baron, Milner, The remainder of this paper is organizes as follows: in Section 2,
& Naseraldin, 2011), inventory management (Song, 2010), finance we define the deterministic RCPSP and stochastic RCPSP under var-
(Fabozzi, Huang, & Zhou, 2010), revenue management ied activity durations with a brief discussion of the scheduling
(DAVENDRALINGAM, 2011), stochastic systems (Bertsimas, methodology. The traditional deterministic mathematical formula-
Litvinov, Sun, Zhao, & Zheng, 2013) and even in energy systems tion for RCPSP is shown in Section 3. In Section 4, an overview of
(Jiang, Wang, & Guan, 2012). The detailed reviews on recent the robust optimization approach/framework is discussed with dif-
advances in robust optimization can be referred to (Gabrel et al., ferent uncertainty levels. Our robust optimization approach for
2014). Although the concept of robust optimization has already durational uncertainty is addressed in Section 5, followed by a
been utilized in different practical areas, the application on RCPSPs detailed discussion on our proposed B&C algorithm and its solution
is still meagre. Apart from the earlier work of Artigues et al. (2013), methodologies (Section 6). The results of our computational exper-
the authors could not find out any works on robust optimization iments on different PSPLIB benchmark instances for both deter-
approach for RCPSPs. Therefore, the application of a robust opti- ministic and stochastic RCPSPs are presented in Section 7. The
mization approach for RCPSPs that consider uncertain activity computational behavioural patterns exhibited by the different
 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550 539

heuristic procedures are described in Section 7.2.4. Finally, we pro- schedule, with associated vectors of predictive starting and com-
vide conclusions in the last section. pletion times, to attempt to limit the risk of future perturbations,
is still a challenging research area for practitioners.
In uncertain environments, especially from a practical point of
2. Problem description view, project managers are mainly interested in the generation of
a proactive schedule with a quality that does not degrade during
The RCPSP can be stated as follows. A single project consists of project execution because of future perturbations. In this paper,
n þ 1 activities where each activity has to be processed in order to we used the unit penalty cost approach in addition to minimizing
P
complete the project. The dummy activities 0 and n þ 1 correspond the schedule makespan, as shown in Eq. (1), where Tt¼0 txit repre-
to the ‘project start’ and to the ‘project end’, respectively. The activ- sents the realized starting time of any activity i after uncertainty, si
ities are interrelated by two kinds of constraints. First, precedent represents the actual starting time of any activity i before uncer-
constraints force activity j not to be started before its immediate tainty is considered and the penalty cost or instability weight, b1i
predecessor activities have been finished. Second, all the activities and b2i , allocated to each activity i. b1i and b2i denotes the marginal
resource consumption at each particular time should satisfy their
costs of deviating the start time of activity i from its planned com-
maximum resource availabilities. In general there are two types
pletion time in the baseline schedule. Hence, the overall objective
of resources, i.e. renewable and non-renewable. Renewable
for this SRCPSP is to find precedence feasible and resource feasible
resources are limited for a particular time period and can be P
schedule s that minimizes both the project makespan Tt¼0 txðIþ1Þt
replenished (for example, machine and manpower). On the con-
and the sum of deviation penalties of all activities under uncertain
trary, non-renewable resources are limited throughout the project
activity duration set d ~i (as shown in Eq. (1)). Here we mention that,
tenure like money. The RCPSP under study in this paper is based on
the following assumptions (i) the activities composing a project for the formulation of RCPSP, we use a binary decision variable xit ,
have certain and known durations for the deterministic part only; which equals 1 if activity i starts at time period t, and 0 otherwise.
(ii) all predecessors must be finished before an activity can start; Hence, xðIþ1Þt represents the binary decision variable which is equal
(iii) although resources can be either of renewable or non- to 1 if the last activity (i.e., the dummy end activity, I þ 1) starts at
renewable, here we consider only the renewable resources; (iv) time period t, and 0 otherwise.
activities are non-preemptive (i.e., cannot be interrupted when in " #þ " #þ !
X XT X
T X
T
progress); and (v) the main objective is to minimize the project Min Z sc ¼ b1i txit  si þ bi si 
2
txit þ txðIþ1Þt
completion time. The proposed formulations for scheduling are i2I t¼0 t¼0 t¼0
discrete in manner, i.e., it depends on discretization of the time ð1Þ
horizon T, which is represented as time period t ðt ¼ 0 . . . TÞ. Each
resource type has a certain capacity limit which should be utilized Uncertainty in activity duration is usually represented by ran-
throughout the project life without exceeding that limit. dom variables with known probability distribution functions. As
Meanwhile for SRCPSPs, we associate with each activity 8i 2 I a the realized completion time is directly affected by duration uncer-
~  Rþ containing the possible realizations of the uncertain tainty, to determine the deterministic robust counterpart for
set d i
uncertain activity duration, we employed a robust optimization
duration of activity i (with Rþ the set of non-negative reals). The
framework only on durational uncertainty.
~ can be a discrete set fd ; d ; d ; . . . ; d ~ g or an interval
set of d i i1 i2 i3 ijdi j
min max min
½di ; di . In the first case, we also write di  mind~i dik and 3. Deterministic mathematical formulations
max
di  maxd~i dik . The dummy activities, 0 and n þ 1 have zero dura-
~i is also assumed In this section we present the Mixed Integer Programming
tion. Moreover, to handle uncertainty, the set of d
model for the RCPSPs under study. For the formulation of RCPSP,
to potentially follow different general distributions (i.e., uniform,
we use a binary decision variable xit which is equal to 1 if activity
normal and Poisson). Carrying out of a project with uncertain activ-
i starts at time t, and 0 otherwise. This formulation requires at
ity durations is a dynamic decision process. Under this scenario, a
most nT binary decision variables and Oðn2 þ mTÞ constraints,
solution is a policy which defines actions at decision times. A
where n is the number of jobs, m is the number of arcs, and T is
schedule is constructed gradually through time t. In this paper,
the maximum time required to complete the project. In spite of
we focus on activity duration uncertainty which is considered as:
various alternative models (Bianco & Caramia, 2011; Brucker &
(a) bounded, (b) symmetric and bounded, (c) following a known
Knust, 2000; Brucker, Knust, Schoo, & Thiele, 1998; Koné, 2012),
distribution such as uniform, Poisson or normal distribution, and
mainly based on the representation of decision variables, the initial
(d) following an unknown probability distribution. It should be
mathematical model from Pritsker, Waiters, and Wolfe (1969) and
noted that the random duration parameters considered in this
Talbot (1982) for the RCPSP is still in use and is hence used in this
paper are assumed to be fully independent.
paper. The basic formulation is presented below.
To the best of our knowledge, apart from the work of Artigues
et al. (2013), none of the methods proposed in the literature of
SRCPSPs considered a robust optimization framework that can rep- Sets
resent uncertain probabilistic parameters. Moreover, unlike the lit- I set of activities, i = 0, . . . , I+1
erature, instead of assuming that only one activity at a time T set of time periods, t = 0, . . . , T
disturbs the starting time of a successor activity, we considered Pprec Represents the precedence set
that any precedence activities may hamper the starting time of R set of renewable resources, r = 1, . . . , R
their successors. Secondly, our standpoint is rather unusual in Parameters
the literature on stochastic resource constrained project schedul- K rt Capacity of resource r at time t
ing, since our aim is to construct a proactive schedule under uncer- U ir Resource usage of activity i for resource r
tainty, partially bridging the gap between the stochastic di Duration of activity i
scheduling literature and the robust one. In uncertain environ-
ments, especially from a practical context, constructing a proactive
540 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550

X
T Let ascertain that the left-hand-side coefficients A and B, and
Minimize Z ¼ txðIþ1Þt the right-hand-side parameters p of the inequality constraints
t¼0
are uncertain parameters. The true realization of an uncertain
Constraints: parameter is represented by,
X
T ~ ¼ ð1 þ e  nÞ  a
a ð7Þ
xit ¼ 1 8i 2 I ð2Þ
t¼0 ~ being its true realization,
where, a is an uncertain parameter with a
X
T X
T e represents a given uncertainty level, and n stands for a random
txjt P ðt þ di Þxit 8ði; jÞ 2 Pprec ð3Þ variable. Then, according to the robust optimization conditions,
t¼0 t¼0 any inequality constraint l in Eq. (6) becomes:
X X X X
X
I X
t
alm xm þ ~lm xm þ
a blk yk þ b~ y
lk k
U ir xiq 6 K rt 8r 2 R and 8t 2 T ð4Þ mRMl m2Ml kRK l k2K l
i¼1 q¼tdi þ1
~l þ d max½1; jpl j 8l
6p ð8Þ
xit 2 f0; 1g 8i 2 I; t ¼ 1; . . . ; T ð5Þ
where l is the index of the uncertain inequality, m is the index of the
In the above model, the objective is to minimize the completion continuous terms, and k is the index of the binary terms. Thus, Ml
time of the last activity of the project, which represents the project and K l are the subsets that encompass those uncertain parameters
completion time (this objective function is for only our deterministic for the given inequality constraint l. During project tenure, there
RCPSP). Constraint (2) indicates that every job or activity must be may arise numerous unavoidable uncertainties such as resource
started exclusively in one time period. That means, no balking or pri- capacity fluctuations, infrequent arrival of raw materials, man-
ority violation is allowed (Non-preemptive case). The job which has power uncertainty, and activity duration uncertainty. Among them
started must be completed first, depending upon their corresponding resource uncertainty and activity’s processing time uncertainty, are
priority. Constraint (3) ensures the precedence relationship that most common and can be described using continuous or discrete
ensures that an activity cannot be started before the completion of distributions. In this paper, we focus on activity duration or pro-
its preceding activity. Finally, constraint (4) represents the capacity cessing time uncertainty which is considered as: (a) bounded, (b)
constraints of all renewable or non-renewable resources. symmetric and bounded, (c) following a known distribution such
as uniform, normal or poisson distribution, or (d) following an
4. Overview of the robust optimization framework unknown probability distribution. The considered random duration
parameters in this paper are assumed to be fully independent.
The robust optimization framework developed by Janak et al.
(2007), Li et al. (2011), Li and Floudas (2014), Li, Tang, and Floudas 4.1. Bounded uncertainty
(2012) is used to address duration uncertainty in this paper. Before
going into detail, we present a brief robust optimization approach, Let us assume that the uncertain parameters vary in a bounded
which explicitly takes into account the various forms of parameter interval and they are represented by,
uncertainty within constraints and/or objective functions. As shown ~lm  alm j 6 ejalm j
ja
in the previous section on RCPSP, the optimal solution of an MILP
~  b j 6 ejb j
jb
program may become infeasible, that is, one or more constraints lk lk lk
are violated substantially, if the nominal data is slightly disrupted. ~l  pl j 6 ejpl j
jp
Our objective here is to develop a robust optimization methodology
to generate ‘‘reliable” solutions to the MILP program, which are less Then according to Ye et al. (2014), the deterministic robust
vulnerable to uncertainty. This means that the optimal solution counterpart for Eq. (8) derived as follows,
0 1
should provide the best possible value of the original objection func- X X X X
tion while also being guaranteed to remain feasible in the range of alm xm þ blk yk þ e@ jalm jum þ jblk jyk A
the uncertainty set considered (Ben-Tal & Nemirovski, 2000). The m k m2M l k2K l
robust optimization framework has the advantage that it not only
6 pl  ejpl j þ d max½1; jpl j 8l ð9Þ
explicitly takes into account the various forms of parameter uncer-
tainty, but also ensures that the obtained solution is feasible for the where um 6 xm 6 um and variable um is an auxiliary positive vari-
nominal system conditions. In the robust optimization framework, able. Meanwhile, l is the index of that uncertain inequality con-
a solution ðx; yÞ is called robust if it satisfies the following conditions: straint, m is the index of the continuous terms, and k is the index
of the binary terms. Then, the deterministic robust counterpart of
(i) ðx; yÞ is feasible for the nominal problem; the original uncertain MILP problem can be derived by adding con-
(ii) whatever are the true values of the coefficients and right- straint (6) and the limiting constraint: um 6 xm 6 um 8m 2 Ml
hand-side parameters, ðx; yÞ must satisfy the l -th inequality
constraint with an error of at most d max½1; jpl j, where d is a 4.2. Bounded and symmetric uncertainty
given infeasibility tolerance (Ye et al., 2014).
For further study, more elaborate explanation and proofs can be Again, if the uncertain parameters are randomly and symmetri-
found in the work of Ye et al. (2014). Let us assume a generic deter- cally distributed around their nominal values, then they are repre-
ministic MILP problem: sented as follows:

Min=Maxx;y cT x þ d y
T ~lm ¼ ð1 þ enlm Þalm
a
s:t: Ex þ Fy ¼ e ~ ¼ ð1 þ en Þb
b lk lk lk
Ax þ By 6 p ð6Þ ~l ¼ ð1 þ enl Þpl
p
x 6 x 6 x
where, nlm , nlk ; nl are random variables distributed symmetrically in
y ¼ f0; 1g the interval [1, 1]. Now we can define a probabilistic version of
 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550 541

condition (ii), as described in the previous section for a robust solu- 4.3. Known probability distribution
tion. For the l-th inequality, the probability of the event of con-
P P ~
straint violation, i.e. ~ xm þ
a b y > p þ d max½1; jp j, is at
m lm k lk k l l
The true values of the uncertain parameters are obtained from
most j, where d > 0 is a given feasibility tolerance and j > 0 is a their nominal values by random variables:
9
given ‘‘reliability level” (Lin et al., 2004). ~lm ¼ ð1 þ enlm Þalm >
a =
~ ¼ ð1 þ en Þb
b ð17Þ
Lemma 1. Let qh be given reals and gh be independent random lk lk lk
>
;
variables symmetrically distributed in [1, 1]. Then for every X > 0 ~l ¼ ð1 þ enl Þpl
p
the following inequality holds: Now, if the probability distributions of the random variables nlm ,
( )
X qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X nlk and nl in the uncertain parameters are known, then the earlier
Pr g h qh > X q2
h h
6 expfX2 =2g ð10Þ MILP can be re-written as follows:
h T
Min=Maxx;y cT x þ d y
For the sake of better understanding, we provide here the proof of s:t: Ex þ Fy ¼ e
this Lemma presented by Ben-Tal and Nemirovski (2000). By homo- ~ þ By~ p
P Ax ~ 6 d:max½1; jpj
geneity arguments, it suffices to consider the case of h q2h ¼ 1. Now,
in this case x 6 x 6 x
( ) ( ( )) y ¼ f0; 1g
X X
Pr gh qh > X |{z}
6 expfX2 g exp X gh qh where the data set n  ½A; B; p varies in a given uncertainty set Z,
h
ðaÞ
h ~ B
A; ~ and p
~ represents the true value of the uncertain coefficients,
Y and d P 0 is an infeasibility tolerance. The inequality (17) can be
¼ expfX2 g fexp fXgh qh gg
|{z}
h
written in expanded form as
ðbÞ
X X
" # ~lm xm þ
a b~ y 6p~l þ d max½1; jpl j 8l ð18Þ
Y X1
ðXqh Þ2r
lk k

6 expfX g
2 m2M l k2K l
|{z} ð2rÞ!
ðcÞ
h r¼0
~ and p
~lm , b
Every constraint l, a ~l are again the true values of the
lk
Y uncertain coefficients. Substituting the expressions for the true
¼ expfX2 g expfX2 q2h =2g
|{z}
h
values of the uncertain coefficients given in constraint (17), the
ðdÞ
uncertain inequality in (18) can be rewritten as follows
0 1
¼ expfX2 =2g
|{z} X X X X
ðeÞ alm xm þ blk yk  pl þ e@ nlm alm xm þ nlk blk yk  nl pl A
m k m2Ml k2K l
With (a) Tschebyshev inequality; (b) independence of gh ; (c) Tay-
lor’s expansion and symmetric distribution of gh ½1; 1; (d) Taylor’s 6 d  max½1; jpl j ð19Þ
P
expansion and (e) h q2h ¼ 1. Therefore, according to this aforemen- where Ml and K l define the sets of uncertain parameters alm and blk ,
tioned Lemma, it can be justified that the reliability level respectively for constraint l. Now to transform the constraint into a
j ¼ expfX2 =2g. deterministic form, Janak et al. (2007) proposed the following
Now, for an infeasibility tolerance ðdÞ and a reliability level ðjÞ, formulation
8 0 1
the following ðe; d; jÞ - deterministic robust counterpart for Eq. (9) <X X X X
can be derived as follows: pr a x þ blk yk  pl þ e@ nlm alm xm þ nlk blk yk  nl pl A
2 : m lm m k m2Ml k2K l
X X X X )
alm xm þ blk yk þ e4 jalm jum þ jblk jwlk þ jpl juOl
m k m2M l k2K l > d  max½1; jpl j 6j ð20Þ
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X X ffi
þX m2M im lm
a 2 2
l
z þ b
k2K lk lk
l
2 2
v þ p 2
l zO 2
l This constraint enforces that the probability of violation of the
uncertain inequality is at most j, where d P 0 is a given feasibility
6 pl þ d max½1; jpl j8l ð11Þ tolerance and j P 0 is a given reliability level (i.e., the probability
where of violation of constraint l). Hence, if we know a probability distri-
bution function for the sum of the random variables,
ulm 6 xm  zlm 6 ulm 8m 2 Ml ð12Þ X X
n¼ nlm alm xm þ nlk blk yk  nl pl ð21Þ
wlk 6 yk  v lk 6 wlk 8k 2 K l ð13Þ m2Ml k2K l

we can use this information in the probabilistic constraint (20), to

uOl 6 1 þ zOl 6 uOl 8l ð14Þ
write a deterministic form for the uncertain constraint which is reli-
able, depending on the value of j. This case is demonstrated using
x 6 x 6 x ð15Þ
the definition of a probability distribution function and the follow-
ing relationship
yk ¼ f0; 1g 8k ð16Þ
F nl ðkÞ ¼ Prfnl 6 kg ¼ 1  Prfn > kg ¼ 1  j ð22Þ
where um , wk , uOl are auxiliary positive variables, l is the index of
the uncertain inequality, m is the index of the continuous terms, k Then the final form of the deterministic constraint (or robust
is the index of the binary terms and X is a positive parameter with counterpart problem) is simply determined using the inverse dis-
j ¼ expðX2 =2Þ. Then, the deterministic robust counterpart of the tribution function (quantile) of the random variable n
original uncertain MILP problem can be derived by adding con-
nl ð1  jÞ ¼ f ðk; jalm jxm ; jblk jyk ; pl Þ
F 1 8l ð23Þ
straint sets (11)–(16).
542 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550

Thus, the generic deterministic robust counterpart of the prob- Now, an upper bound of the probability of the sum of the uncer-
abilistic constraint for random variables following any distribution tainty right-hand side parameters taking on a value greater than k1
can be formulated as follows: can be provided by means of the Markov inequality (Verderame &
X X Floudas, 2009), as follows:
alm xm þ blk yk  pl þ e:f ðk; jalm jxm ; jblk jyk ; pl Þ ( ) P
X i2Il pli
m k
Pr ~
pli > k 6 ð33Þ
6 d max½1; jpl j 8l ð24Þ i2Il
k1

where l is the index of the uncertain inequality constraint, m is To determine the value of k1 ,
( ) P P
the index of the continuous terms, and k is the index of the binary X p p
terms. k is determined from j using constraint (22) and the prob- j ¼ Pr ~li > k 6 i2Il li :So; k1 6 i2Il li
p ð34Þ
i2I
k 1 j
ability distribution function for n. For example, if we make the l

assumption that all the relevant random variables are independent The upper bound on k1 represents the largest expected value that
and follow a normal distribution, the function f can be represented P
~li can take on for a given value of j, and as a result, it
the term i2Il p
in the following manner:
can be utilized in order to formulate a relaxed deterministic robust
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X X ffi
2 2 counterpart constraint for Eq. (29) (see Eqs. (35) and (36)).
f ðk;jalm jxm ;jblk jyk ;pl Þ ¼ k1 a
mM lm lm m
2
r 2 2
x þ b
kK lk lk k
r y þ p r
2 2
l l " #
l l
X X X X
ð25Þ alm xm þ blk yk P pli þ k1  dmax 1; pli 8l ð35Þ
m k iRIl i2Il
where rlm , rlk and rl are the standard deviations for the indepen-
dent normal random variables used to represent the uncertainty X
pli " #
associated with the parameters alm , blk and rl respectively. Replac- X X X i2Il X
ing the function f with its equivalent form for random variables fol- alm xm þ blk yk P pli þ  dmax 1; pli 8l ð36Þ
m k iRIl
j i2Il
lowing independent normal distributions, the deterministic robust
counterpart constraint can be written as seen in Eq. (26). For the sake of computational ease, we assumed that the
X X rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
X X 2 unknown distribution is approximately symmetric and the
alm xm þ blk yk  pl þ ek1 a 2 r2 x 2 þ
mM lm lm m
b r2 y þ p2l r2l
kK lk lk k
l l extremes are equidistant from the mean. Then a lower bound on
m k P
the expected value of the term ~
i2Il pli can be determined for a
6 dmax½1; jpl j 8l ð26Þ
given value of j, and it is equivalent to the value of k2 defined by
Eq. (37). The max operator is used in order to signify the fact that
4.4. Unknown probability distribution
the underlying distribution is nonnegative, as the duration values
are always non-negative. The given lower bound can then be used
In this section, we discuss the scenario of uncertain right-hand
in order to formulate a relaxed deterministic robust counterpart
side parameters which follow an unknown probability distribution.
constraint for Eq. (30), as seen in Eqs. (38) and (39).
The nominal values for the right hand side parameters of Eq. (6) " !#
(i.e., which follow an unknown probability distribution) will be con- X
sidered equivalent to the average values for the same right hand side k2 ¼ max 0; 2 pil  k1 ð37Þ
i2Il
parameters. The inequality constraints for this type are like,
X X X " #
alm xm þ blk yk P pli 8l ð27Þ X X X X
m k i alm xm þ blk yk 6 pli þ k2 þ dmax 1; pli 8l ð38Þ
m k iRIl i2Il
X X X
alm xm þ blk yk 6 pli 8l ð28Þ  
m k i
X X X 2j  1 X
alm xm þ blk yk 6 pli þ max 0; pli
Eqs. (29) and (30) represent the realization of the uncertain m k iRIl
j i2Il
" #
parameters, pli , where Il represents the subset of the right-hand side X
parameters which are stochastic in nature for a given constraint l. þ dmax 1; pli 8l ð39Þ
" # i2Il
X X X X X P
alm xm þ blk yk P pli þ p~li  dmax 1; pli 8l ð29Þ Here, pl ¼ i pli and l is the index of the uncertain inequality
m k iRIl i2Il i2Il constraint. Meanwhile, pli represent the uncertain parameters for
" # element iði 2 Il Þ, where Il represents the subset of right-hand side
X X X X X
alm xm þ blk yk 6 pli þ p~li þ dmax 1; pli 8l ð30Þ parameters which are stochastic in nature for a given constraint
m k iRIl i2Il i2Il l. For better understandings, interested readers are referred to
the work of Verderame and Floudas (2009).
The upper mentioned inequalities can be reformulated into Therefore, the deterministic robust counterpart optimization
probabilistic inequalities, like Eqs. (31) and (32), while j is defined model for the original uncertain MILP problem can be derived by
as a reliability level. adding constraint set (39).
( " #)
X X X X X
Pr alm xm þ blk yk < pli þ p~li  dmax 1; pli 6 j 8l 5. Robust optimization approach for durational uncertainty
m k iRIl i2Il i2Il

ð31Þ In the case of RCPSPs, activity duration uncertainty may lead to

( " #) increases in the project completion time or makespan time. Hence,
X X X X X in the case of designing robust RCPSPs, handling those durational
Pr alm xm þ blk yk > pli þ ~
pli þ dmax 1; pli 6 j 8l
uncertainties is a must. In its practical aspect, activity duration
m k iRIl i2Il i2Il
uncertainty replicates processing time uncertainties and may have
ð32Þ numerous forms to represent. Keeping that in mind, in this paper
 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550 543

we present six heuristics, based on the way the activity duration 5.2. Bounded and symmetric uncertainty
uncertainties are measured or estimated. Here, the activity dura-
tion information is implicitly lumped into set dc , which denotes For the case of RCPSPs, if the X demand parameters are uncer-
the processing times for an activity set c. As mentioned earlier, tain and distributed around the nominal values randomly and
the overall objective for this SRCPSP is to find precedence feasible ~ ¼ ½1 þ e  n   d where n are random
symmetrically as follows,d i i i; i
and resource feasible schedule s that minimizes both the project variables distributed symmetrically in the interval [1, 1]. Then by
PT
makespan t¼0 txðIþ1Þt and the sum of deviation penalties of all following a similar guideline from Eq. (11), the robust counterpart
activities under uncertain activity duration set d ~ (as shown in for Eq. (41) is derived as follows,
i
Eq. (1)). However, since the realized project makespan increases jdc j P di þ e  di  minðX; 1Þ  d max½1; di  8ði; cÞ 2 I ð43Þ
with increases in activity durations, the cardinality of set dc should
be higher than or equal to any typical di . Here, is a positive parameter with j ¼ expðX =2Þ and the vio-
2

lation probability or reliability level is j. Similarly,

jdc j P di 8ði; cÞ 2 I ð40Þ
jdc j ¼ ceilfdi þ e  di  minðX; 1Þ  d max½1; di g. Meanwhile the
~ represent the true realization of the uncertain demand
Let d second heuristic is named Heuristic 2 and is outlined below:
i
parameter of activity i. Then Eq. (40) becomes:
Heuristic 2
~
jdc j P d 8ði; cÞ 2 I ð41Þ
i
1: determine the realized activity duration vector d ~ for each
i
Now we can apply the robust optimization techniques to Eq. e
activity i, while assuming that d ¼ ½1 þ e  n   d .
i i i
(41). For simplicity, universal values for e, j and d were adopted;
2: apply robust optimization framework (i.e., Eq. (43)) to
however, as noted by Janak et al. (2007), the robust optimization
measure or identify the cardinality of set jdc j
framework can easily be extended in order to take into account
3: solve the corresponding robust optimization based RCPSP
scenarios where e can vary from parameter to parameter and j
model by using our proposed B&C algorithm as seen in
and d can vary from constraint to constraint. Here in this
Sections 6.1–6.4
research, we set two different uncertainty levels (U.L), e as 0.05
4: output the solution found and generate a robust schedule
and 0.15; the infeasibility tolerance d as 0%; and the reliability
5: end
level j as 5%.

5.1. Bounded uncertainty

5.3. Robust counterpart with known probability distribution

If the activity duration parameter varies in a bounded interval

As of Eq. (24), the generic robust counterpart of the probabilistic
min max
½di ; di , then according to Eq. (9), the robust counterpart for constraints for our targeted RCPSPs can be formulated as follows
Eq. (41) is given by: for random variables following any distribution,
max
jdc j P di  d max½1; di  8ði; cÞ 2 I ð42Þ jdc j P di þ ef ½k; di   d max½1; di  8ði; cÞ 2 I ð44Þ

and d ~ ¼ max
di d max½1; di  Similar to that with bounded uncertainty, the constraint can be,
i
Here we include the robust counterpart (i.e., Eq. (42)) into the jdc j ¼ ceilfdi þ ef ½k; di   d max½1; di g
deterministic formulation to generate the robust optimization for-
mulation. To solve this robust optimization formulation, we need 5.3.1. Uncertainty with uniform probability distribution
to predetermine set dc . Recall that the objective function is to min- Assume the duration parameters are uncertain and follow a
imize the makespan, which increases as processing time increases. uniform continuous distribution. Given an uncertainty level (e),
Therefore, only the equality part of Eq. (42) needs to be considered, an infeasibility tolerance (d), and a reliability level (j), to generate
max
which stands when, jdc j ¼ di  d max½1; di , which is the maxi- robust solutions, the following (e; d; j) - robust counterpart of the
mum true realization of the uncertain duration parameters. Since original uncertain MILP problem can be derived.
the values of set dc need to be integer, then Min=Maxx;y cT x þ d y
T
max
jdc j ¼ ceilfdi  d max½1; di g should be considered, where function
s:t: Ex þ Fy ¼ e
ceil½z provides the smallest integer that is greater than z. Later on,
Ax þ By 6 p
we present how this robust optimization model was solved by X X
using our proposed B&C approach. The first heuristic is referred alm xm þ blk yk  pl þ eð1  2jÞjpl j 6 d max½1; jpl j 8l ð45Þ
to as Heuristic 1 and is described in pseudo code below. m k

x 6 x 6 x
Heuristic 1 y ¼ f0; 1g
~ for every
1: determine the realized activity duration vector d Therefore, for right-hand-side parameter uncertainty (i.e., dura-
i
involved activity by considering bounded interval tion uncertainty), Eq. (45) becomes,
min max
½di ; di . jdc j P di þ eð1  2jÞdi :  d max½1; di  8ði; cÞ 2 I ð46Þ
2: apply robust optimization framework (i.e., Eq. (42)) to
measure or identify the cardinality of set jdc j Here, f ½k; di  ¼ ð1  2jÞdi for uniform continuous distribution.
3: solve the corresponding robust optimization based RCPSP
model by using our proposed B&C algorithm as seen in 5.3.2. Uncertainty with normal probability distribution
Sections 6.1–6.4 Again, we assume that the distributions of the random variables
4: output the solution found and generate a robust schedule nlm , nlk and nl in (17) are all standardized normal distributions with
5: end zero as the mean and one as the standard deviation. Then the dis-
tribution of n defined in (21) is also a normal distribution with
544 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P P 2 ⇑ (continued)
mMl alm xm þ kK l blk yk þ pl as the standard
2 2 2
mean zero and
deviation (Janak et al., 2007). Heuristic 5
According to Eq. (22), k ¼ F 1 1
n ð1  jÞ and F n is the inverse dis- 2: apply robust optimization framework (i.e., Eq. (48)) to
tribution function of a random variable with standardized normal measure or identify the cardinality of set jdc j
distribution. Thus, 3: solve the corresponding robust optimization based RCPSP
Z model by using our proposed B&C algorithm as seen in
k
1 x2
j ¼ 1  F n ðkÞ ¼ 1  Prfn 6 kg ¼ 1  pffiffiffiffiffiffiffi expð Þdx Sections 6.1–6.4
1 2p 2
4: output the solution found and generate robust a schedule
For this considered RCPSP, if the activity duration parameters 5: end
follow a normal distribution, the deterministic robust counterpart
(i.e., Eq. (44)) is represented by,

jdc j P di þ eF 1
n ð1  jÞdi :  d max½1; di  8ði; cÞ 2 I ð47Þ
5.4. Unknown probability distribution
Therefore, jdc j is determined as ceilfdi þ e F 1
n ð1  jÞdi :
d max½1; di g. When the activity duration parameters are uncertain and follow
an unknown probability distribution, then according to the Eq.
5.3.3. Uncertainty with Poisson probability distribution (39), the robust counterpart for Eq. (41) is represented as,
 
On the other hand, when the uncertain activity duration param- 2j  1
eters follow a Poisson probability distribution, then the determin- jdc j 6 max 0; di þ d max½1; di  8ði; cÞ 2 I ð49Þ
j
istic robust counterpart for Eq. (44) is given by,
The violation probability or reliability level is j and jdc j is deter-
jdc j P di þ eF 1
n ð1  jÞdi :  d max½1; di  8ði; cÞ 2 I ð48Þ mined as ceilfmax½0; 2jj1di þ d max½1; di g: Under this circum-
where k ¼ F 1 stance, Heuristic 6 is carried out by following the pseudo code
n ð1  jÞ is the inverse distribution function of a dis-
crete variable with a Poisson distribution. Here we categorize provided below:
Heuristics 3, 4 and 5 depending on different uncertainty distribu-
tions and are outlined below: Heuristic 6

1: determine the realized activity duration vector d~ for each

i
Heuristic 3 activity i, when the activity duration parameters are
1: determine the realized activity duration vector d ~ for each uncertain and follow an unknown probability distribution
i
activity i, while assuming that the activity duration 2: apply robust optimization framework (i.e., Eq. (49)) to
parameters are uncertain and follow a uniform continuous measure or identify the cardinality of set jdc j
distribution. 3: solve the corresponding robust optimization based RCPSP
2: apply robust optimization framework (i.e., Eq. (46)) to model by using our proposed B&C algorithm as seen in
measure or identify the cardinality of set jdc j Sections 6.1–6.4
3: solve the corresponding robust optimization based RCPSP 4: output the solution found and generate a robust schedule
model by using our proposed B&C algorithm as seen in 5: end
Sections 6.1–6.4
4: output the solution found and generate a robust schedule
5: end
6. Solution methodology

Branch and cut (B&C) is a generalization of B&B where the LP

Heuristic 4 relaxation of the ILP is used to obtain a bound at each node in
the search tree. If a node has a fractional solution and cannot be
1: determine the realized activity duration vector d ~ for each
i fathomed, this approach finds cuts that are violated by the frac-
activity i, while assuming that the activity duration tional solution but are satisfied by all feasible integer solutions
parameters are uncertain and follow a normal distribution (Zhu et al., 2006). If no cuts can be found, branching is performed
2: apply robust optimization framework (i.e., Eq. (47)) to to create new nodes in the tree. B&C has proven to be effective in
measure or identify the cardinality of set jdc j solving different combinatorial optimization problems (Bianco &
3: solve the corresponding robust optimization based RCPSP Caramia, 2012; Naber & Kolisch, 2014; Zhu et al., 2006). Our solu-
model by using our proposed B&C algorithm as seen in tion approach makes use of the cut generating features built into
Sections 6.1–6.4 the CBC solver that comes with Matlab (R2012b). In addition, for
4: output the solution found and generate a robust schedule accelerating convergence, we provide problem specific cuts to
5: end tighten the LP bounds, and also alternative branching rules. In this
section, we detail the overall solution techniques for the proposed
RCPSPs.
Heuristic 5
6.1. Fixing variables
1: determine the realized activity duration vector d ~ for each
i
activity i, while assuming that the activity duration Here for each activity i 2 I, we define the decision variable xit for
parameters are uncertain and follow a Poisson probability t bounded by the earliest finish time (EFT) and latest finish time
distribution (LTF) instead of considering the upper bound on the project
 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550 545

completion time T, which drastically reduce the number of vari- up each special ordered set were partitioned into two subsets
ables. The EFT of activity i is defined as the earliest time period t and all the variables in one subset were set to zero on each branch-
such that there exist a feasible schedule in which activity i finishes ing. Predominantly, the branching was performed on the finish
at time t. Similarly, the LFT is defined as the latest time period t in time of an activity. We employed the CBC solver defined default
which activity i can finish for a given upper bound of the project bound tightening scheme every time a SOS was used for branching.
makespan. However, finding those EFT and LFT values is an NP- Despite having several techniques to accelerate the solution pro-
hard problem. Although for any case, it is not necessary to know cess, initial testing indicated that many problem instances were
EFT and LFT exactly for all activities to formulate valid MILP mod- still difficult to solve due to the large number of binary variables.
els (2)–(6), as calculating lower and upper bounds on those values In response, we adopted a solver that defined a greedy heuristic
is sufficient to lead to smaller models (Zhu et al., 2006). Here, we approach for faster divergence while the neighbourhoods were
employed an iterative critical path method (CPM) for obtaining defined by linear inequalities. However, if the heuristic results do
better lower bounds for more reduction of the number of variables. not yield a new solution with makespan less than the incumbent
makespan, we switched the heuristic and directly solved with
6.2. Cuts from resource conflicts the solver (with its default local branching scheme).
Benchmark problems developed by Kolisch and Sprecher (1997)
As from constraint set (5), all renewable resource constraints have been widely used to test various exact and heuristic methods
are in the form of knapsack constraints with generalized upper for RCPSPs and can be found in the PSPLIB. In this paper, the B&C
bound (GUB) variables (Gu, Nemhauser, & Savelsbergh, 1998). That approach was evaluated by solving instances from that library.
is, all variables are divided into mutually exclusive subsets, and in For solving those MILP models, a commercial optimization solver
each subset one variable must be present. For details on the com- from the Computational Infrastructure for Operations Research
plexity and the computation of GUB cover cuts see Gu et al. (1998). (COIN-OR) community, named as Coin-OR Branch and Cut (CBC)
These cuts were developed whenever there was a resource conflict (version 2.9.3), was employed. It comes with Matlab (R2012b),
and were generated automatically by the CBC solver through Mat- and was executed on an Intel core i7 processor with 16 GB RAM
lab when requested. For a renewable resource, a GUB cover cut and 3.40 GHz CPU. All relevant documents and solver specified
prevents two or more activities from being executed at the same information can be found from the OPTI toolbox.
time if the combined resource usage exceeds the maximum limit.

7. Computational results
6.3. Cuts from precedence relations
7.1. For deterministic RCPSPs
Precedence cuts were obtained by explicitly enumerating the
possible finish times for activities with precedence relations. For
To analyse the B&C algorithm, all benchmark instances from
instance, if activity i precedes activity j, which has a duration of
different activity numbers, ranging from lower to higher activities
dj , then according to constraint (3), if activity i finishes at time t
(30–90 activities), available in PSPLIB were chosen, except for 120-
then activity j must finish after t þ dj . Each of the enumerated cases activity instances where some random instances were chosen. In
was written in the form of a cut which was used to tighten the LP Table 1, we give the results of our B&C algorithm compared with
bounds. Meanwhile, it was not necessary to consider cuts from all the tightest bound on the KSD instances proposed by Brucker
of the precedence relations, because many were redundant accord- and Knust (2000) (BK: column 4), with the destructive lower
ing to the proposition mentioned in Zhu et al. (2006). bound (DLB: column 5) proposed by Demassey, Artigues, and
Michelon (2005) and with the B&B algorithm solved using Lingo.
6.4. Branching rules and bound tightening Here, %Optimal represents the ratio of number of instances solved
optimally among all 480-instances of different activity numbers
In this case, the bound on one variable at a time is fixed in the (except for the 120-activity instances where we solved 53 ran-
construction of the search tree. We avoided branching on a single domly chosen instances). Despite the higher computation time,
variable, which is supposedly very inefficient, because it gives our B&C algorithm is comparable to, and even gives better percent-
identical results for the two sub problems of the original model. age of optimal results than any other exact algorithms particularly
In special ordered set (SOS) branching, the variables that make for higher activity instances. Moreover, except for the 30-activity

Table 1
Computational results on the KSD instance sets.

No. of Act Parameters B&B BK DLB B&C

KSD 30 (480) % Gap(avg.) 0.00 1.5 0.7 7.16
CPU time (avg.) 12.20 0.4 3.2 41.47
%Optimal 100 66.25 83.95 94.37
KSD 60 (480) % Gap(avg.) 11.59 7.8 7.7 6.59
CPU time (avg.) 153.17 5.0 168 89.82
%Optimal 66.04 71.04 75.00 80.00
KSD 90 (480) % Gap(avg.) 17.97 7.2 7.0 6.86
CPU time (avg.) 376.67 72 379 135.68
%Optimal 57.91 72.91 75.83 76.75
KSD 120 (53) % Gap(avg.) 22.27 21.4 19.1 15.70
CPU time (avg.) 437.25 21,300 1388 284.35
%Optimal 9.434(53) 34.67 (600) 38.17(600) 68.87(53)

The values indicated by bold marks represent the best obtained values among all alternative heuristics or algorithms, under each parameter setting.
546 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550

Table 2
Test results for some randomly chosen 30-activity instances from PSPLIB.

Group RS B&B GA GA_LR B&C

LB %Gap t LB %Gap t LB %Gap t LB %Gap t
1 0.2 47.6 0.00 14.8 48.4 1.53 830 48.2 1.11 145.9 47.6 0.00 68.43
2 0.5 45.2 0.00 1.0 45.2 0.00 913 45.2 0.00 74.8 45.2 0.00 5.27
3 0.7 64 0.00 20.8 64.6 0.76 1000 64.2 0.20 460.3 64 0.00 2.77

instances, our proposed algorithm shows the least average%GAP computing time. When B&C and B&B are compared, the both have
than the other approaches, even for larger activity instances. 0% Gap, and B&C is faster for group 3, and B&B is faster for groups 1
Although the average number of integer variables and generated and 2. On the contrary, for higher activity instances (activity num-
nodes for higher activity instances are much higher than the lower bers 60 and 90), this is not always true.
ones, their average solution time is always lower for the B&C algo-
rithm than for B&B and even BK and DLB which further proves the
efficiency of our proposed algorithm in comparison to exact 7.2. For RCPSPs with uncertain activity durations
approaches, particularly for higher activity instances.
For better comparison, three other algorithms were executed on As mentioned earlier, the second target of this work is to deal
the same computer configuration for those randomly chosen 30- with SRCPSPs under uncertain activity duration scenarios. Depend-
activity instances (only for Table 2). To do this, we used four types ing on the uncertain characteristics, we proposed six robust opti-
of indicators which have emerged to characterize the RCPSP mization based heuristics, which are mainly based on different
instances: precedence-oriented; time-oriented; resource-oriented ways to represent uncertainties and have already been discussed
and hybrid (Artigues et al., 2008). Kolisch, Sprecher, and Drexl in Section 6. For comparing those heuristics and different ways
(1995) observed that the hardness of the RCPSP instances of handling uncertainties, computational experiments have been
decreases as network complexity (NC) increases from 1.5 to 2.1. carried out on a set of randomly selected benchmark problems/
Another important indicator is resource strength (RS), which is a instances from the PSPLIB, for 30, 60 and 120 activities [10 prob-
hybrid indicator which mixes both resource and time parameters. lems with 30 activities (j301_1, j302_1, j305_1, j306_1, j307_1,
After experimenting with the 30-activity instances of PSLIB, it has j308_1, j309_7, j3010_1, j3037_1 and j3044_1), 10 problems with
been observed that the average solution time reduces with increas- 60 activities (j601_1, j602_1, j605_1, j606_1, j607_1, j608_1,
ing RS values from 0.20 to 1. Again the resource factor (RF) reflects j609_1, j6010_1, j6037_1, j6044_1), 10 problems with 120 activi-
the average portion of resources requested per job, where RF = 1 ties (j1201_1, j1202_1, j1205_1, j1206_1, j1207_1, j1208_1,
means all resources are required by a job. Considering all those j1209_1, j12030_1, j12037_1, j12060_1)].
important instance characteristics, all the randomly chosen For solving those deterministic robust counterparts of the
instances (only for Table 2) were chosen from those that has [NC, robust optimization formulation (i.e., six heuristics), we employed
RF, RS] of [1.50, 0.25, 0.2–0.7], which reflects the difficult ones our proposed B&C algorithm which was coded in Matlab (R2012b)
among the available instances. under the built-in CPLEX platform, and was executed on an Intel
For finding optimal results, at first all the MILP models were core i7 processor with 16 GB RAM and 3.40 GHz CPU. Note that
solved by using the commercial optimization software LINGO for the sake of computational swiftness, we used limits of maxi-
v10.0, which predominantly use the branch and bound (B&B) algo- mum 10,000 nodes, 2000 iterations and 500 s solving time for
rithm as default and is exact in nature. Meanwhile, all those our B&C algorithm. Note that the results obtained by our B&C algo-
instances were also solved by using a genetic algorithm (GA) and rithm may not sufficiently be optimal for all instances due to its
a Lagrange relaxation based genetic algorithm (GA_LR) under the solution time limitation. Moreover, for larger activity instances,
Matlab (R2012b) platform. Here we mention that for solving with some infeasible results were found due to this defined solution
GA, the authors only employed the built-in GA toolbox with minor time limit. In that case, we excluded those results for our future
customization, while for GA_LR, all the equality constraints of the analysis. Hence, the average value of makespan after durational
MILP model were relaxed and acted as a penalty function in the variance for SRCPSP (i.e., MADV-makespan after durational vari-
objective values. To compare with GA and GA_LR, it was infeasible ance) presented in Tables 3–5, are only for those instances, for
to use the large activity instances. So only a subset of the 30- which we got optimal or at least feasible and integer results. For
activity instances was used. All the results for the four different solving heuristic 1, we assumed that the uncertain processing
algorithms, within or before 1000 s, are summarized in Table 2. times vary in a bounded interval, while the lower bound of that
min
As instance hardness depends on its NC, RF and RS values, all the interval ðd Þ was measured by multiplying each activity’s deter-
max
chosen instances were grouped in three different cases by varying ministic duration by 0.5 and the upper bound ðd Þ was measured
their RS values. Here in Table 1, LB represents the lower bound or by multiplying by 2.25. For better understanding, we set two dif-
best obtained results within 1000 s, t represents the average ferent uncertainty levels (U.L), e as 0.05 and 0.15; the infeasibility
computing time in seconds, while% Gap is calculated by using the tolerance d as 0%; and the reliability level j as 5%. Here to recall
following equation. that the overall objective for this SRCPSP is to minimize both the

Best obtained result  Optimal or best known result according to PSPLIB

%Gap ¼  100%
Optimal or best known result according to PSPLIB

PT
Compared with GA and GA_LR, both B&C and B&B are more con- project makespan t¼0 txðIþ1Þt
and the sum of deviation penalties
sistent for solving 30-activity instances (0% Gap) with much lower ~
(SODP) of all activities under uncertain activity duration set d i
 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550 547

Table 3
Comparison for KSD 30 benchmark problem.

U.L Para. Heuristic 1 Heuristic 2 Heuristic 3 Heuristic 4 Heuristic 5 Heuristic 6

5% MBDV (avg.) 56.1 56.1 56.1 56.1 56.1 56.1
MADV (avg.) 95.4 72.075 101.4 77.7 62.475 77.625
SADV 16.35 81.78 36.67 29.08 12.11 108.14
SODP (avg.) 412.48 241.88 470.22 282.48 114.96 753.24
CPU (s) 192.91 351.985 363.903 400.97 203.919 232.603
15% MBDV (avg.) 56.1 56.1 56.1 56.1 56.1 56.1
MADV (avg.) 95.4 106.65 103.875 105.375 66.225 77.625
SADV 16.35 79.54 34.26 37.5 15.71 109.81
SODP (avg.) 412.48 570.32 481.16 431.2 206.24 753.24
CPU (s) 197.0207 424.907 382.689 374.912 255.126 233.209

The values indicated by bold marks represent the best obtained values among all alternative heuristics or algorithms, under each parameter setting.

Table 4
Comparison for KSD 60 benchmark problem.

U.L Para. Heuristic 1 Heuristic 2 Heuristic 3 Heuristic 4 Heuristic 5 Heuristic 6

5% MBDV (avg.) 97.2 97.2 97.2 97.2 97.2 97.2
MADV (avg.) 125.7 81.45 138.375 110.625 81.6 43.575
SADV 17.42 77.71 1.08 72.14 26.23 23.71
SODP (avg.) 1007.44 443.2 1141.74 703.1 443.2 735.88
CPU (s) 414.1 448.238 482.78 488.234 450.864 492.375
15% MBDV (avg.) 97.2 97.2 97.2 97.2 97.2 97.2
MADV (avg.) 125.7 90.825 138.45 153.525 101.475 43.575
SADV 17.42 80.20 3.52 75.58 36.8 25.68
SODP (avg.) 1007.44 578.12 1141.9 1295.08 830.82 735.88
CPU (s) 387.9006 409.454 482.679 344.107 455.712 492.678

The values indicated by bold marks represent the best obtained values among all alternative heuristics or algorithms, under each parameter setting.

Table 5
Comparison for KSD 120 benchmark problem.

U.L Para. Heuristic 1 Heuristic 2 Heuristic 3 Heuristic 4 Heuristic 5 Heuristic 6

5% MBDV (avg.) 133.9 133.9 133.9 133.9 133.9 133.9
MADV (avg.) 161.1 100.65 137.175 133.5 97.575 37.2
SADV 14.85 3.56 2.58 5.00 3.53 1.82
SODP (avg.) 2270.6 1511.8 1707.98 1820.5 1506.24 2338.4
CPU (s) 273.508 468.034 326.937 500.00 500.00 500.00
15% MBDV (avg.) 133.9 133.9 133.9 133.9 133.9 133.9
MADV (avg.) 161.1 125.55 138.075 200.025 143.1 44.7
SADV 14.85 4.85 3.67 7.42 2.98 2.83
SODP (avg.) 2270.6 2196.2 1727.68 2481.8 1836.22 2432.6
CPU (s) 279.366 460.964 436.825 375.922 500.00 500.00

The values indicated by bold marks represent the best obtained values among all alternative heuristics or algorithms, under each parameter setting.

(as shown in Eq. (1)). The penalty coefficients are: and CPU is the average of ten values obtained, after applying the
b1i ¼ b2i ¼ 1 8i 2 I n fI þ 1g; b1Iþ1 ¼ 1; b2Iþ1 ¼ 8, which are mainly B&C algorithm for a maximum of 500 s. Table 3 shows that the
motivated from Zhu, Bard, and Yu (2005). Below, we present our makespan obtained after durational variance (MAVD) and SODV
computational results for KSD (Kolisch and Sprecher, 1997) for heuristic 5 (i.e., uncertainty follows known and Poisson distri-
instance sets with 30, 60 and 120 activities. bution) is the least among all the heuristics under both uncertainty
levels, although the computational time is a little higher than
heuristic 1. The rationale behind this is that for heuristic 1, the
7.2.1. Comparison of different uncertainty handling techniques for 30- uncertain processing times vary in a bounded interval, which is
activity instances independent of uncertainty or risk levels. As a result, the MADV
Table 3 displays for every heuristic, the mean value of make- values for heuristic 1 are the same for both uncertainty levels. Also
span under the deterministic condition (MBDV-makespan before for heuristic 6, which assumed that the uncertainty distributions of
durational variance), the average value of makespan after dura- the activity durations are unknown, we obtained slightly higher
tional variance for SRCPSP (MADV-makespan after durational vari- MADV values, but compared to the other heuristics, the PSC values
PT
ance, t¼0 txðIþ1Þt ), standard deviation of obtained MADV values are significantly higher. This is perhaps because heuristic 6 deals
(SADV), sum of deviation penalties (SODP) of all activities and with the most uncertain distribution. This efficacy of heuristic 5
the average CPU time (second) needed to solve the set of ten ran- further proves that, for KSD 30 instances under stochastic condi-
domly chosen instances with thirty non-dummy activities. In the tion, selecting a Poisson distribution for representing uncertainties
table, each value reported in the columns for MBDV, MADV, SODV is a sound approach.
548 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550

7.2.2. Comparison of different uncertainty handling techniques for 60- (i.e., KSD 60 and KSD 120), representing uncertainties by an
activity instances unknown probability distribution gives better result than others,
In this section, we focus on the set of instances with 60 non- while for lower activity instances (i.e., KSD 30), a Poisson distribu-
dummy activities. As shown in Table 4, our proposed B&C algo- tion is the better choice.
rithm took almost all our dedicated time limit (500 s) for solving
these instances. In spite of being larger activity instances than
7.2.4. Impact of risk level on different heuristics
before in Section 7.2.1, similar conclusion can be drawn, apart from
To analyse the impact of risk levels on both expected/realized
those about heuristic 6. Surprisingly for this case, heuristic 6 pro-
makespan after considering duration uncertainties (here MADV is
duced the least MADV values in comparison to other heuristics,
termed as EXPMAK) and sum of deviation penalties (SODP), we
although a precise justification of this scenario cannot be given
considered all ten above mentioned randomly chosen KSD 30
because of its’ uncertainty characteristics. Meanwhile for two dif-
instances from PSPLIB under four different risk levels, e (0.05,
ferent uncertainty levels, Table 4 shows that both the MADV values
0.10, 0.15, 0.20). Here, Figs. 1 and 2 reports the SODP and EXMAK
and SODP values for uncertainty level 5%, is quite lower than val-
under varied risk levels for all six different heuristics. As we can
ues for 15%, which is expected. More detailed discussion on the
observe in Fig. 1, apart from heuristics 1 and 6, the average SODP
impact of risk or uncertainty levels will be covered in Section 7.2.4.
increases with increasing values of e up to a certain limit. This is
an expected result, since for decreasing values of e we impose a
7.2.3. Comparison of different uncertainty handling techniques for more prudent project manager’s position, and thereby impose a
120-activity instances higher risk aversion level. Mathematically speaking, as the value
We have performed similar comparisons for the set of 120 non- of e decreases, probabilistic constraints are somehow more binding
dummy activities. While most exact algorithms failed to solve the and the schedule is more robust, since it is less exposed to disrup-
KSD 120 instances within our dedicated time frame (500 s), our tions. Moreover, with decreasing risk levels, a project manager can
B&C algorithm produced feasible and integer optimal (6 out of 10 decide more prudently, which also leads to lower SODP vales. The
instances on average) results, even though it took higher CPU time. reader may notice a seemingly strange trend, particularly for
As of Section 7.2.2, similar conclusion can be drawn for all six heuristics 2 and 3, in that the SODP started to decrease after a cer-
heuristics. Table 5 shows, for uncertainty level 5%, that heuristic tain point. In effect, the SODP seems to have a non-monotonic
5 shows better solution robustness (i.e., least SODP value) than behaviour, with an increasing slope up to a maximum and a
any other heuristics, although it produces higher makespan than decreasing slope afterwards. This behaviour is more evident for
heuristic 6. These analyses further justify that, if someone consid- the KSD 30 instances. The unforeseen descendant behaviour of
ers that the activity duration parameters are uncertain and follow the expected makespan is due to the influence of two opposing
an unknown probability distribution, he/she can obtain better forces (i.e., the expected makespan whose values increases as e
makespan performance than with the other five ways of uncer- decreases and on the other hand the expected tardiness that dras-
tainty handling techniques. Therefore, for larger activity instances tically reduces as long as the risk we are willing to bear decreases)

800
700 Heuristic 1
600 Heuristic 2
500 Heuristic 3
SODP

400 Heuristic 4
300
Heuristic 5
200
Heuristic 6
100
0
0.05 0.1 0.15 0.2
Risk Level
Fig. 1. Risk level versus SODP.

180

160 Heuristic 1
Heuristic 2
140
EXPMAK

Heuristic 3
120 Heuristic 4
100 Heuristic 5
Heuristic 6
80

60
0.05 0.1 0.15 0.2
Risk Level
Fig. 2. Risk level versus EXPMAK.
 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550 549

(Bruni et al., 2011). However depending on the uncertainty distri- Artigues, C., Leus, R., & Nobibon, F. T. (2013). Robust optimization for resource-
constrained project scheduling with uncertain activity durations. Flexible
butions, heuristics 1 and 6 show constant SODP values for all four
Services and Manufacturing Journal, 25, 175–205.
uncertainty levels, which further proves their independence con- Baron, O., Milner, J., & Naseraldin, H. (2011). Facility location: A robust optimization
cerning risk parameters. approach. Production and Operations Management, 20, 772–785.
A similar trend can be observed in Fig. 2, for EXPMAK, which Ben-Tal, A., & Nemirovski, A. (2000). Robust solutions of linear programming
problems contaminated with uncertain data. Mathematical Programming, 88,
also increases with increasing risk levels. However, the increasing 411–424.
trend here is also non-monotonic, particularly for heuristics 2 and Bertsimas, D., Litvinov, E., Sun, X. A., Zhao, J., & Zheng, T. (2013). Adaptive robust
3. According to Figs. 1 and 2, the stability costs and EXPMAK show optimization for the security constrained unit commitment problem. IEEE
Transactions on Power Systems, 28, 52–63.
better performance at risk levels of 0.1–0.15, and so it should be Bianco, L., & Caramia, M. (2011). A new formulation for the project scheduling
preferred in terms of both stability cost and makespan perfor- problem under limited resources. Flexible Services and Manufacturing Journal, 25,
mance. Therefore, the range [0.1, 0.15] constitutes a meaningful 6–24.
Bianco, L., & Caramia, M. (2012). An exact algorithm to minimize the makespan in
choice for moderately risk averse project managers who wish to project scheduling with scarce resources and generalized precedence relations.
avoid unnecessary extra cost. European Journal of Operational Research, 219, 73–85.
Brucker, P., & Knust, S. (2000). A linear programming and constraint propagation-
based lower bound for the RCPSP. European Journal of Operational Research, 127,
355–362.
8. Conclusions Brucker, P., Knust, S., Schoo, A., & Thiele, O. (1998). A branch and bound algorithm
for the resource-constrained project scheduling problem. European Journal of
Operational Research, 107, 272–288.
In the past, exact methods for solving the RCPSP have mainly Bruni, M. E., Beraldi, P., & Guerriero, F. (2015). The stochastic resource-constrained
focused on enumerating all possible activity sequences within a project scheduling problem handbook on project management and scheduling.
branch and bound framework. In the first part of this work, we pro- Handbook on project management and scheduling (Vol. 2). Springer.
Bruni, M. E., Beraldi, P., Guerriero, F., & Pinto, E. (2011). A heuristic approach for
posed a B&C procedure for general linear objective functions, with
resource constrained project scheduling with uncertain activity durations.
integer variables and extended types of resource constraints. Main- Computers & Operations Research, 38, 1305–1318.
taining all the precedence and resource constraints, the proposed Chakrabortty, R. K., Sarker, R. A., & Essam, D. L. (2014). Event based approaches for
solving multi-mode resource constraints project scheduling problem. Computer
algorithm was employed for small to large activity instances. The
information systems and industrial management. Springer.
primary contribution of the research pivoted on the application Chakrabortty, R. K., Sarker, R. A., & Essam, D. L. (2016). Multi-mode resource
of the B&C algorithm while using the CBC solver, which predomi- constrained project scheduling under resource disruptions. Computers &
nantly is focused on the RCPSP discipline. Empirical tests on bench- Chemical Engineering, 88, 13–29.
Davendralingam, N. (2011). Formulation and demonstration of a robust mean
mark problems showed that our approach gives markedly variance optimization approach for concurrent airline network and aircraft
improved results with respect to conventional exact algorithms design.
and even some meta-heuristic approaches, under minimum mod- Deblaere, F., Demeulemeester, E., & Herroelen, W. (2011a). Proactive policies for the
stochastic resource-constrained project scheduling problem. European Journal of
ifications. The Second aspect of this article covers more realistic Operational Research, 214, 308–316.
project scheduling problems, which are termed as stochastic- Deblaere, F., Demeulemeester, E., & Herroelen, W. (2011b). Reactive scheduling in
RCPSPs. In practical project management, a project’s parameters, the multi-mode RCPSP. Computers & Operations Research, 38, 63–74.
Demassey, S., Artigues, C., & Michelon, P. (2005). Constraint-propagation-based
such as activity duration and resource requirements, are scarcely cutting planes: An application to the resource-constrained project scheduling
ever precisely known and are usually subject to estimation errors. problem. INFORMS Journal on Computing, 17, 52–65.
In this paper we utilized the robust optimization framework to Fabozzi, F. J., Huang, D., & Zhou, G. (2010). Robust portfolios: Contributions from
operations research and finance. Annals of Operations Research, 176, 191–220.
develop a deterministic robust counterpart optimization model Fu, N., Lau, H. C., & Varakantham, P. (2015). Robust execution strategies for project
for activity duration uncertainty. It was assumed that the uncer- scheduling with unreliable resources and stochastic durations. Journal of
tainty can be of different forms and show different statistical distri- Scheduling, 1–16.
Gabrel, V., Murat, C., & Thiele, A. (2014). Recent advances in robust optimization: An
butions. Depending on those different uncertainty characteristics,
overview. European Journal of Operational Research, 235, 471–483.
we have developed six different heuristic models for SRCPSP the Gu, Z., Nemhauser, G. L., & Savelsbergh, M. W. (1998). Lifted cover inequalities for
under robust optimization framework. All those heuristic models 0–1 integer programs: Computation. INFORMS Journal on Computing, 10,
were then employed for small to large activity instances, and were 427–437.
Hartmann, S., & Briskorn, D. (2010). A survey of variants and extensions of the
solved by using our proposed B&C algorithm. The robust solution resource-constrained project scheduling problem. European Journal of
from the robust optimization framework is guaranteed to be feasi- Operational Research, 207, 1–14.
ble for the nominal parameters. The results on benchmark prob- Hartmann, S., & Kolisch, R. (2000). Experimental evaluation of state-of-the-art
heuristics for the resource-constrained project scheduling problem. European
lems proved the dominance of certain heuristics (i.e., certain Journal of Operational Research, 127, 394–407.
uncertainty handling ways) over others and helped to give proper Herroelen, W., & Leus, R. (2005). Project scheduling under uncertainty: Survey and
guidance to risk reluctant project managers and practitioners. research potentials. European Journal of Operational Research, 165, 289–306.
Huang, L., Chen, H.-S., & Hu, T.-T. (2013). Survey on resource allocation policy and
Using a genetic algorithm or any other advanced meta-heuristic job scheduling algorithms of cloud computing1. Journal of Software, 8, 480–487.
(i.e. simulated annealing, particle swarm optimization, etc.) for Janak, S. L., Lin, X., & Floudas, C. A. (2007). A new robust optimization approach for
local branching, could be an important future work for getting scheduling under uncertainty: II. Uncertainty with known probability
distribution. Computers & Chemical Engineering, 31, 171–195.
good feasible solutions in the early stages of the computations. Jiang, R., Wang, J., & Guan, Y. (2012). Robust unit commitment with wind power and
For projects with relatively long makespan or activities whose fin- pumped storage hydro. IEEE Transactions on Power Systems, 27, 800–810.
ish time intervals are large, more efficient solution methods are Kolisch, R., & Sprecher, A. (1997). PSPLIB-a project scheduling problem library: OR
software-ORSEP operations research software exchange program. European
still needed. Along with durational uncertainty, there may be other
Journal of Operational Research, 96, 205–216.
sorts of uncertainties in project execution. Therefore, designing Kolisch, R., Sprecher, A., & Drexl, A. (1995). Characterization and generation of a
novel disruption handling strategies for SRCPSPs with those uncer- general class of resource-constrained project scheduling problems. Management
tainties could be another important area for future study. Science, 41, 1693–1703.
Koné, O. (2012). New approaches for solving the resource-constrained project
scheduling problem. 4OR, 11, 93–94.
Koné, O., Artigues, C., Lopez, P., & Mongeau, M. (2013). Comparison of mixed integer
References linear programming models for the resource-constrained project scheduling
problem with consumption and production of resources. Flexible Services and
Manufacturing Journal, 25, 25–47.
Artigues, C., Koné, O., Lopez, P., Mongeau, M., Néron, E., & Rivreau, D. (2008).
Lamas, P., & Demeulemeester, E. (2014). A purely proactive scheduling procedure
Benchmark instance indicators and computational comparison of methods.
for the resource-constrained project scheduling problem with stochastic
Resource-Constrained Project Scheduling: Models, Algorithms, Extensions and
activity durations. Journal of Scheduling, 1–20.
Applications, 107–135.
550 R.K. Chakrabortty et al. / Computers & Industrial Engineering 112 (2017) 537–550

Lambrechts, O., Demeulemeester, E., & Herroelen, W. (2007). Proactive and reactive Remli, N., & Rekik, M. (2013). A robust winner determination problem for
strategies for resource-constrained project scheduling with uncertain resource combinatorial transportation auctions under uncertain shipment volumes.
availabilities. Journal of Scheduling, 11, 121–136. Transportation Research Part C: Emerging Technologies, 35, 204–217.
Lambrechts, O., Demeulemeester, E., & Herroelen, W. (2010). Time slack-based Song, M. (2010). Applications of stochastic inventory control in market-making and
techniques for robust project scheduling subject to resource uncertainty. Annals robust supply chains. Massachusetts Institute of Technology.
of Operations Research, 186, 443–464. Talbot, F. B. (1982). Resource-constrained project scheduling with time-resource
Li, Z., Ding, R., & Floudas, C. A. (2011). A comparative theoretical and computational tradeoffs: The nonpreemptive case. Management Science, 28, 1197–1210.
study on robust counterpart optimization: I. Robust linear optimization and van de Vonder, S., Ballestin, F., Demeulemeester, E., & Herroelen, W. (2007).
robust mixed integer linear optimization. Industrial & Engineering Chemistry Heuristic procedures for reactive project scheduling. Computers & Industrial
Research, 50, 10567–10603. Engineering, 52, 11–28.
Li, Z., & Floudas, C. A. (2014). A comparative theoretical and computational study on van de Vonder, S., Demeulemeester, E., & Herroelen, W. (2008). Proactive heuristic
robust counterpart optimization: III. Improving the quality of robust solutions. procedures for robust project scheduling: An experimental analysis. European
Industrial & Engineering Chemistry Research, 53, 13112–13124. Journal of Operational Research, 189, 723–733.
Li, Z., Tang, Q., & Floudas, C. A. (2012). A comparative theoretical and computational Verderame, P. M., & Floudas, C. A. (2009). Operational planning of large-scale
study on robust counterpart optimization: II. Probabilistic guarantees on industrial batch plants under demand due date and amount uncertainty. I.
constraint satisfaction. Industrial & Engineering Chemistry Research, 51, Robust optimization framework. Industrial & Engineering Chemistry Research, 48,
6769–6788. 7214–7231.
Lin, X., Janak, S. L., & Floudas, C. A. (2004). A new robust optimization approach for Ye, Y., Li, J., Li, Z., Tang, Q., Xiao, X., & Floudas, C. A. (2014). Robust optimization and
scheduling under uncertainty:I. Bounded uncertainty. Computers & Chemical stochastic programming approaches for medium-term production scheduling
Engineering, 28, 1069–1085. of a large-scale steelmaking continuous casting process under demand
Ma, W., Che, Y., Huang, H., & Ke, H. (2015). Resource-constrained project scheduling uncertainty. Computers & Chemical Engineering, 66, 165–185.
problem with uncertain durations and renewable resources. International Zhou, J., Love, P. E. D., Wang, X., Teo, K. L., & Irani, Z. (2013). A review of methods and
Journal of Machine Learning and Cybernetics, 1–9. algorithms for optimizing construction scheduling. Journal of the Operational
Naber, A., & Kolisch, R. (2014). MIP models for resource-constrained project Research Society, 64, 1091–1105.
scheduling with flexible resource profiles. European Journal of Operational Zhu, G., Bard, J. F., & Yu, G. (2005). Disruption management for resource-constrained
Research, 239, 335–348. project scheduling. Journal of the Operational Research Society, 56, 365–381.
Pritsker, A. A. B., Waiters, L. J., & Wolfe, P. M. (1969). Multiproject scheduling with Zhu, G., Bard, J. F., & Yu, G. (2006). A branch-and-cut procedure for the multimode
limited resources: A zero-one programming approach. Management Science, 16, resource-constrained project-scheduling problem. INFORMS Journal on
93–108. Computing, 18, 377–390.