A Two-stage Stochastic Programming Model for the Resource
Constrained Project Scheduling Problem under Uncertainty
Maria Elena Bruni, Luigi Di Puglia Pugliese, Patrizia Beraldi and Francesca Guerriero
Department of Mechanical, Energy and Management Engineering, Unical, Italy
Keywords:
RCPSP, Uncertainty, Integer L-shaped, Stochastic Programming.
Abstract:
Due to the increasing competitiveness of businesses, project planning and scheduling have become a chal-
lenging theme in the last years. In this paper, we propose a two-stage stochastic programming model for the
resource constrained project scheduling problem, taking into account the stochasticity of activity durations. In
this formulation, assuming that some activity duration scenarios are known, resource allocations are taken in
the first stage, while scheduling decisions are postponed in the second stage. The resulting problem is a mixed
integer problem with recourse, where binary variables appear in the first stage. In order to efficiently solve
the problem, a decomposition algorithm is developed, based on the well-known integer L-shaped method.
Detailed computational results are presented for a set of benchmark instances taken from the literature.
1 INTRODUCTION
The resource-constrained project scheduling problem
(RCPSP) deals with the sequencing of activities that
are usually related by precedence constraints and by
the simultaneous use of scarce resources. The RCPSP
has attracted the attention of scientists and practitio-
ners, given also its importance in many fields (Bruni
et al., 2011a,b), who have made considerable efforts
to enhance the efficiency of the solution process. In
the last decade, increasing financial pressures have
put more emphasis on the project execution and, gi-
ven the uncertainty characterizing the project envi-
ronment, the role of deterministic static schedules has
been challenged. Due to machine breakdowns, em-
ployee absenteeism and delay in materials supply, bad
whether conditions and many other uncontrollable
factors, one or more project activities may experience
a delay, threatening the operational viability of the
planned schedule and its implementation. To address
these challenges, a flourishing stream of literature has
focused on the RCPSP under uncertainty in which
activity durations are assumed to be random variables.
Two different approaches can be used depending on
the genuine interpretation of the RCPSP under uncer-
tainty and the way the uncertainty is tackled. Indeed,
the problem can be viewed as a stochastic dynamic
optimization problem, where decisions are made each
time new information becomes available, or as a pro-
blem where a tentative plan, which can be changed
during project execution, should be determined and
agreed before knowing the realization of uncertainty.
As far as the first option is concerned, the literature
has focused on dynamic decision processes (called
policies) that define, at completion of some activities,
appropriate actions concerning the choice of a set of
activities that should be executed next. The vast ma-
jority of the research in this area is concerned only
with the expected makespan objective. M
¨
ohring and
Stork (2000) following Igelmund and Radermacher
studied the so-called linear preselective policies, that
were exploited by Stork (2001) to develop a branch-
and-bound procedure, to efficiently compute an opti-
mal policy minimizing the expected makespan. Po-
licies have also been used for determining predictive
activity starting times, with the objective of minimi-
zing costs related to positive and negative deviations
of actual starting times, from the predicted ones, and
to penalties/bonuses associated with late/early project
completion. Deblaere et al. (2007) and Golenko et al.
(1997) presented solution procedures for the RCPSP
with random activities duration and the expected ma-
kespan as objective. Starting from the concept of cri-
tical chain introduced by Goldratt (Goldratt, 1997),
in Rabbani et al. (2007) a new heuristic was pre-
sented for minimizing the expected project duration
and its variance. Later, Ballest
´
ın (2007) developed
regret-based biased random sampling procedures then
embedded into a genetic algorithm, whereas Ballest
´
ın
and Leus (2009) examined multiple possible objective
194
Bruni, M., Pugliese, L., Beraldi, P. and Guerriero, F.
A Two-stage Stochastic Programming Model for the Resource Constrained Project Scheduling Problem under Uncertainty.
DOI: 10.5220/0006612601940200
In Proceedings of the 7th International Conference on Operations Research and Enterprise Systems (ICORES 2018), pages 194-200
ISBN: 978-989-758-285-1
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
functions for project scheduling with stochastic acti-
vity durations. For a special case involving only one
renewable resource (the budget) a two-stage integer
linear stochastic program has been proposed in Zhu
et al. (2007), whereas Bruni et al. (2011a), proposed
a chance-constrained based heuristic aiming at buil-
ding a baseline schedule which is protected against
possible disruptions.
Contrary to the stochastic project scheduling
(Bruni et al. 2015), robust project scheduling assumes
that the distribution of the uncertain activity duration
is not known or only partially known. Robustness can
be referred to either the project makespan (in this case
it is referred as quality-robustness) or to possible devi-
ations between the planned and realized starting times
of the projected schedule (we call this solution robus-
tness). The literature on robust project scheduling has
mainly dealt with the development of effective and
efficient proactive and reactive scheduling procedu-
res. Proactive scheduling aims at generating robust
baseline schedules, that incorporate some protection
against possible disruptions, whereas reactive schedu-
ling procedures can be invoked during the execution
of the project, to repair the baseline schedule by de-
viating as little as possible from the original baseline
schedule. For an extensive review of research in this
field, the reader is referred to Demeulemeester and
Herroelen (2011).
Robust optimization approaches have been re-
cently proposed for the RCPSP, under general poly-
hedral uncertainty (Bruni et al., 2017a,b). Assuming
that scenarios represent different realizations of the
activity durations, a min-max absolute-regret problem
is proposed by Artigues et al. (2011), to minimize the
maximum absolute difference between the makespan
obtained by the robust solution and the scenario de-
pendent optimal solutions.
In this paper, we study a stochastic programming
optimization approach for the RCPSP, assuming that
the random variables are discretely distributed. We
observe that such an assumption is quite general,
since discrete distributions arise either naturally in
many real-world applications, or as empirical approx-
imations of the continuous ones derived, for example,
by taking a Monte Carlo sample from a general distri-
bution.
The two-stage stochastic RCPSP (TSRCPSP) is
investigated by assuming that scenarios represent dif-
ferent realizations for the activity durations. The de-
cision variables are divided into wait-and-see varia-
bles, that must be determined before the realization
of the uncertain parameters, and here-and-now varia-
bles, that can adjust to the uncertain data when they
become known. The objective is to find a schedule
that minimizes the expected makespan over all sce-
narios. To solve the problem, we have designed and
implemented an integer L-shaped decomposition ap-
proach, both in the single cut and the multi-cut versi-
ons.
The remainder of the paper is organized as fol-
lows: in Section 2, a formal definition of the TSR-
CPSP is given. Section 3 presents a detailed descrip-
tion of the proposed exact algorithm. Section 4 dis-
cusses the computational results obtained on a set of
benchmark problems. Finally, some conclusions are
drawn in Section 5.
2 PROBLEM FORMULATION
Project activity duration is typically unknown when
scheduling decisions need to be taken; under uncer-
tainty, specialized models able to cope with this un-
certainty should be developed. Specifically, these
models should allow to make some decisions about
the schedule before the actual activities duration is
known. Then, after the uncertainty is disclosed, a
recourse action can be implemented to compensate
deficiencies in the previously made schedules. Sto-
chastic programming formulations extend and adapt
deterministic models to allow this kind of schedule
modifications. In these models, some decisions must
be made in the first-stage under uncertainty. Then, in
the second-stage, a recourse action can be made after
observing the actual values of the random variables.
In real cases, it is a very common practice to de-
cide upon the resource allocation well in advance,
since often resources (e.g. expert staff) cannot be ea-
sily transferred between activities at short notice, (for
instance in a multi-project environment), nor alloca-
ted without sufficient time lapse. In our model, we
assume that the resource allocation is a static here-
and-now decision, that can be made in advance, un-
der uncertainty about the actual duration of activities.
The starting times, on the contrary, can be decided la-
ter on.
Let define a project as a set of activities V =
{0, . . . , n +1} (where 0 and n +1 are two dummy acti-
vities representing the project start and end, respecti-
vely) that have to be scheduled. A set of renewable
resources K = {1, . . . , m}, each with finite capacity
R
k
for all k belonging to K is used during the project
execution. Precedence constraints between different
activities within the project are modeled by a set of
project arcs E such that (i, j) E means that activity
j has to start after the completion of activity i. Each
activity i V requires a non-negative amount r
ik
and
for the activities 0 and n + 1 r
0k
= r
n+1 k
= R
k
for all
A Two-stage Stochastic Programming Model for the Resource Constrained Project Scheduling Problem under Uncertainty
195
k K.
The TSRCPSP can be defined on the basis of the
flow network model, originally proposed by Artigues
and Roubellat (2000). Different sets of variables are
used. Variables f
i jk
denote the number of units of re-
source k directly transferred from an activity i to an
activity j. Binary variables y
i j
assume the value 1 if
activity j starts after the completion of activity i and
0 otherwise and define an extended set of precedence
relations with the aim of resolving possible resource
conflicts. Finally variables S
i
define the starting times
of each activity i.
We assume that the random vector, represen-
ting stochastic activity durations, has a finite sup-
port. Henceforth, we define Σ as the set of its
possible realizations (scenarios) and we denote by
{0, d
σ
1
, . . . , d
σ
n
, 0}, σ Σ the durations of the activi-
ties under scenario σ occurring with probability p
σ
.
The TSRCPSP may be formulated as follows.
min
y
Q(y) (1)
y
i j
= 1 (i, j) E, (2)
y
i j
+ y
ji
1 i, j V, i < j, (3)
y
il
y
i j
+ y
jl
1 i, j, l V, i 6= j 6= l, (4)
f
i jk
min(r
ik
, r
jk
)y
i j
0 i, j V, i 6= j,
i 6= n + 1, j 6= 0, k K, (5)
jV\{i,0}
f
i jk
= r
ik
i V \ {n + 1}, k K, (6)
iV \{ j,n+1}
f
i jk
= r
jk
j V \ {0}, k K, (7)
f
i jk
0 i, j V, i 6= j, i 6= n + 1, j 6= 0,
k K, (8)
y
i j
{0, 1} i, j V, i 6= j. (9)
It should be noted that the constraints (3) and (4)
are valid inequalities preventing cycles into the ex-
tended graph, and constraints (5), (6) and (7) are re-
source flow-conservation constraints, ensuring feasi-
ble resource allocations. Preprocessing constraints
(2) impose the preexisting precedence constraints.
Hence, the solution of this first stage problem, gi-
ves a resource allocation, i.e., the sequence of the acti-
vities and the resource flow among them. It should
be clear that it is often possible to make different re-
source allocation decisions, each represented by a dif-
ferent resource flow, for the same schedule, having se-
rious impacts on the expected makespan of the sche-
dule. For this reason, the TSRCPSP minimizes in
the second stage the expected value of the makespan,
Q(y) =
σΣ
p
σ
Q(y, σ), resulting from the resource
allocation decisions made in the first stage. For a gi-
ven realization σ, Q(y, σ) is the optimal solution of a
scenario subproblem that can be obtained by solving
the following model:
Q(y, σ) = minS
σ
n+1
(10)
S
σ
j
S
σ
i
My
i j
d
σ
i
M i, j V
σ Σ (11)
S
σ
0
= 0 (12)
here M is large enough constant used to model the
disjunctive precedence constraints (11).
For a fixed y = ˆy, the subproblem (10)–(12) is
equivalent (for strong duality) to the following model.
Q( ˆy, σ) = max
(i, j)|ˆy
i j
=1
d
σ
i
α
i j
jV\n+1
α
ji
+
jV\0
α
i j
= 0 i V \ {0, n + 1}
iV \n+1
α
in+1
= 1
iV \0
α
0i
= 1
α
i j
ˆy
i j
i, j V
α
i j
0 i, j V
3 SOLUTION APPROACH
Problem (1)–(12) is a two-stage stochastic mixed
integer model with continuous recourse, where the
first-stage problem includes binary decision variables,
whereas the recourse problem is a longest path pro-
blem. In general, solving two-stage stochastic pro-
grams is computationally demanding and decomposi-
tion techniques are usually used to solve the problem
efficiently. Specifically, for the continuous recourse
case, the L-shaped algorithm can be used to solve the
two-stage stochastic programming problem based on
Benders decomposition, a well-known technique for
solving large scale MIP problems that exhibit a spe-
cial structure. This approach decomposes the original
problem into an integer master problem and one or
more linear subproblems. In particular, after solving
the master problem, the current optimal master solu-
tion is fed into the recourse problems. Subsequently,
a series of cuts is generated based on the subproblems
solutions and new constraints are added to the master
problem which is then solved again. The above pro-
cess is executed iteratively until an optimal solution
is found. The L-shaped method was first proposed in
Van Slyke and Wets (1969) for solving two-stage sto-
chastic linear problems. Its main idea is to approxi-
mate the expected recourse function Q(y) (whose eva-
luation requires the solution of all the second-stage
ICORES 2018 - 7th International Conference on Operations Research and Enterprise Systems
196
recourse problems) in the objective function of the
two-stage stochastic problem by adding cuts into the
master problem. The cuts are classified into two types
known as optimality cuts and feasibility cuts. The for-
mer drive the master solution toward optimality, while
the latter ensure the feasibility of the subproblems. In
the case of complete recourse (which is also our case)
it is not longer required to take care of feasibility is-
sues.
The general scheme of the L-shaped approach, in
this case, is the following.
Set L B := ; UB := + t = 0.
while UB > LB do
Solve the master problem obtaining y
, f
.
Assign to LB the optimal objective function va-
lue of the master problem.
Solve the subproblems for each scenario σ Σ
and let Q(y
)
t
be the optimal expected make-
span.
Update UB = min(UB, Q(y
)
t
).
Add an optimality cut and update t := t + 1.
end while
Return L B
By defining η as an additional variable, the master
problem in our case assumes the following form:
minη (13)
s.t.(2) (9) (14)
3.1 Optimality Cuts
At a given iteration of the algorithm, once the master
problem has been solved and the corresponding opti-
mal solution obtained, we should solve all the scena-
rio related subproblems (10)–(12). Let Q(y, σ)
t
, for
all σ Σ the optimal second stage objective function,
representing the makespan under scenario σ. Let π
t
σ
the optimal longest path for the scenario σ. Let denote
with π
t
=
σΣ
π
t
σ
the subset of arcs belonging to the
union of the optimal longest paths for each scenario.
Given a finite global lower bound L of the TSRCPSP,
and defining Q = (Q(y)
t
L), the following valid
cut can be added to the master problem formulation:
η Q
(i, j)|(i, j)π
t
y
i j
[Q (|π
t
| 1) L] (15)
Since there is a finite number of such cuts the al-
gorithm converges to an optimal solution to the TSR-
CPSP.
Proposition 1. The cut (15) is a valid optimality cut.
Proof. The quantity
(i, j)π
t
y
i j
is always less than or
equal to |π|, taking exactly the value |π
t
| when y
i j
=
1, (i, j) π
t
. In this case, the right-hand side takes
the value Q(y)
t
, otherwise the right-hand side takes
a value less than or equal to L. Since the first stage
decision variables y are binary, there is only a finite
number of feasible first stage solutions and therefore,
the number of cuts that can be added to the master is
finite.
We have also implemented a multicut L-shaped al-
gorithm (Birge and Louveaux, 1988). In general, the
multicut L-shaped algorithm has less major iterations
than the L-shaped algorithm. However, solving the
master problem requires more computation time. In
particular, by defining an additional set of free vari-
ables η
σ
and denoting with Q
σ
= Q(y, σ)
t
L, the
following optimality cut can be added for each σ Σ:
η
σ
Q
σ
(i, j)|(i, j)π
t
σ
y
i j
[Q
σ
(|π
t
σ
| 1) L] (16)
Optimality cuts (16) ensure that the value of each va-
riable η
σ
is larger than or equal to the optimal value
of its corresponding second-stage problem for each
scenario.
Proposition 2. The cut (16) is a valid optimality cut.
Proof. The proof is omitted since trivial.
In this case, the master problem objective function
is min
σΣ
p
σ
η
σ
.
4 COMPUTATIONAL
EXPERIMENTS
This Section reports on the computational experi-
ments carried out to assess and compare the perfor-
mance of the proposed approaches. The algorithms
have been coded in Java and run on a PC with 16
GB RAM and 2.50 GHz Intel Core i7 - 4710 HQ
CPU. The numerical results have been collected on
120 instances generated from the benchmark determi-
nistic instances of the PSPLIB (Kolisch and Sprecher,
1997) (available at http://www.om-db.wi.tum.de/psplib)
with 30 activities. The instances differ essentially
for three main parameters, namely the network com-
plexity (NC {1.50, 1.80, 2.10}) , the resource fac-
tor (RF {0.25, 0.50}) and the resource strength
(RS {0.70, 1.00}). The first parameter measures
the average number of non-redundant arcs per nodes
including the dummy activities, the second reflects
the average percentage of different resource types for
which a non-dummy activity has a non-zero resource
A Two-stage Stochastic Programming Model for the Resource Constrained Project Scheduling Problem under Uncertainty
197
demand and the last one concerns the strength of the
resource constraints.
A time limit of 20 minutes has been imposed for
both algorithms. We have assumed that each activity
either may have a normal duration or it can be de-
layed. In this case the duration assumes a peak value.
In practical settings, it is unlikely that all the activities
exhibit longer durations than expected. Henceforth,
we have considered single disruption scenarios, that
is in each scenario only one activity can be disrupted.
Consequently, the number of scenarios is 30. The nu-
merical results are collected in Tables 1–4.
Columns NC, RF, and RS correspond to the cha-
racteristics of the problem instance. The total compu-
tational times and the time spent for solving the sub-
problems, in seconds, are indicated in the table with
headers time and timeSP, respectively. The average
number of iterations-calculated over the instances sol-
ved within the time limit (column #solved)-is repor-
ted in column iter. The number of cuts is reported in
columns #cuts.
The behavior of the Benders approach is strongly
influenced by the characteristics of the instances, as
it happens also in the deterministic case. In the next
Sections,we present a detailed analysis of the compu-
tational performance of the proposed approaches de-
pending on the network characteristics.
4.1 Sensitivity to the Network
Characteristics
For the sake of comprehension, let us consider the pa-
rameters NC, RF, and RS separately.
Parameter NC. Tables 1 and 2 show the average
results over all solved instances varying NC, for the
single cut and the multi-cut algorithm, respectively.
Table 1: Average computational results over all the instan-
ces solved within the time limit varying NC, single cut.
NC time iter #cuts timeSP #solv
1.5 197.74 53.60 52.60 0.011 30
1.8 227.52 73.48 72.48 0.016 27
2.1 140.25 66.66 65.66 0.017 32
From the analysis of the results, we can observe
that, on average, 75% of the instances are solved to
optimality within the time limit. As can be observed,
the total computational time is similar for both algo-
rithms, but the number of cuts that should be added
into the master problem is higher for single cut than
for the multi-cut version. In general, the subproblems
do not represent a computational challenge since they
Table 2: Average computational results over all the instan-
ces solved within the time limit varying NC, multi-cut.
NC time iter #cuts timeSP #solv
1.5 192.90 24.90 23.90 0.005 25
1.8 229.68 38.65 37.65 0.014 27
2.1 145.13 35.70 34.70 0.044 26
can be solved in polynomial time. This behaviour is
observed for all values of NC.
Parameter RF. Referring to the values of RF, Ta-
bles 3 and 4 highlight the strong relation between the
average portion of the resources used and consumed
and the behaviour of the proposed solution approach.
Indeed, the lower the value of RF, the higher the num-
ber of solved instances to optimality.
Table 3: Average computational results over the instances
solved within the time limit, single cut.
RF time iter #cuts timeSP #solv
0.25 63.78 26.44 25.44 0.01 59
0.5 426.68 138.83 137.83 0.03 30
Table 4: Average computational results over the instances
solved within the time limit, multi-cut.
RF time iter #cuts timeSP #solv
0.25 128.63 14.92 13.92 0.00 47
0.5 288.78 64.09 63.09 0.05 31
The single cut version is more efficient on pro-
blem with a small RF, whereas when the RF incre-
ases the computational time increases as well. The
multi-cut version of the algorithm is able to solve one
more instance with RF = 0.5. The higher is the va-
lue of the RF, the harder are the instances to be sol-
ved, since when very few resource types are required
by the activities (low values for RF) it is easy to take
good resource allocation decisions.
The last part of the Section is devoted to the analy-
sis of the impact of the resources strength. The higher
the value of RS, the higher the number of instances
solved to optimality, revealing the difficulty of the in-
stances with low RS values. The analysis of the re-
sults shows that the most time consuming instances
for the single-cut algorithm are those with RS = 1,
whereas for the multi-cut version is the opposite. On
the other hand, the multi-cut version is able to solve
the instances with RS = 1 more efficiently than the
single-cut algorithm, with an average time of around
148 seconds.
This might be due to the fact that problems with
RS = 1 have less resource conflicts than problems
with lower values of RS. We notice that this has a di-
ICORES 2018 - 7th International Conference on Operations Research and Enterprise Systems
198
Table 5: Average computational results over the instances
solved within the time limit, single-cut.
RS time iter #cuts timeSP #solv
0.7 131.26 52.68 51.68 0.01 34
1 220.01 71.53 70.53 0.02 55
Table 6: Average computational results over the instances
solved within the time limit, multi-cut.
RS time iter #cuts timeSP #solv
0.7 252.13 33.44 32.44 0.04 27
1 148.62 33.09 32.09 0.01 51
rect impact on the master problem, where a low num-
ber of extra precedences need to be added in order to
eventually solve resource conflicts. Since the master
it is easy to be solved, the multi-cut version turns out
to be more efficient than the single cut on average.
For the instances not solved to optimality within the
time limit, the gap is below 1% for both the single cut
and the multi-cut L-shaped algorithm.
5 CONCLUSIONS
In this paper, we have studied the RCPSP under un-
certainty. In summary, the main contributions of this
paper are the following. First, we have proposed a
two stage stochastic programming approach, where
resource allocation is a first stage decision, whereas
the starting times are second stage decisions. An
integer L-shaped algorithm has been tested on a set
of benchmark instances generated from deterministic
benchmark instances.
The analysis of the results shows that the single-
cut version provides better results than the multi-cut
one. This consideration would suggest the use of dif-
ferent cut aggregation strategies. The design of these
strategies together with the definition of tailored ap-
proaches for the master problem solution is the sub-
ject of ongoing research.
REFERENCES
Artigues C and Roubellat F (2000) A polynomial activity
insertion algorithm in a multi-resource schedule with
cumulative constraints and multiple modes. European
Journal of Operational Research 127: 297–316
Artigues C, Leus R, Nobibon FT (2013) Robust optimiza-
tion for resource-constrained project scheduling with
uncertain activity durations. Flexible Services and
Manufacturing Journal 25(12):175205
Ballest
´
ın F (2007) When it is worthwhile to work with the
stochastic RCPSP? J Sched 10:153–166
Ballest
´
ın F and Leus, R (2009) Resource-constrained pro-
ject scheduling for timely project completion with sto-
chastic activity durations. Prod Oper Manag 18:459–
474
Birge J and Louveaux F (1988)A multicut algorithm for
two-stage stochastic linear programs. European Jour-
nal of Operational Research 34:384–392
Bruni ME, Di Puglia Pugliese L, Beraldi P, Guerriero F,
(2017a) An adjustable robust optimization model for
the resource-constrained project scheduling problem
with uncertain activity durations. Omega 71: 66–84
Bruni ME, Di Puglia Pugliese L, Beraldi P, Guerriero F,
(2017b) A computational study of exact approaches
for the robust resource-constrained project scheduling
problem. Submitted
Bruni ME, Beraldi P, Guerriero F. (2015) The stochas-
tic resource-constrained project scheduling problem.
Handbook on Project Management and Scheduling
Vol. 2:811–835
Bruni ME, Beraldi P, Guerriero F, Pinto E (2011a) A heu-
ristic approach for resource constrained project sche-
duling with uncertain activity durations. Comput Oper
Res 38:1305–1318
Bruni ME, Beraldi P, Guerriero F, Pinto E (2011b) A sche-
duling methodology for dealing with uncertainty in
construction projects. Eng Computation 28(8):1064–
1078
Deblaere F, Demeulemeester E, Herroelen W, VandeVon-
der S (2007) Robust resource allocation decisions in
resource-constrained projects. Decision Sci 38:1–37
Demeulemeester E and Herroelen W (2011) Robust Project
Scheduling. Foundations and Trends in Technology,
Information and Operations Management: Vol. 3: No.
34, pp 201-376.
Goldratt EM (1997) Critical Chain. The North River Press,
Great Barrington
Golenko-Ginzburg D, Gonik A (1998) A heuristic for net-
work project scheduling with random activity durati-
ons depending on the resource allocation. Int J Prod
Econ 55:149–162
Kolisch R and Sprecher A (1997) Psplib - a project schedu-
ling problem library. European Journal of Operational
Research 96(1): 205–216
Laporte G, Louveaux FV (1993) The integer L-shaped met-
hod for stochastic integer programs with complete re-
course. Operations Research Letters, 13(3): 133-142
M
¨
ohring RH and Stork F (2000) Linear preselective stra-
tegies for stochastic project scheduling. Math Method
Oper Res 52(3):501–515
Rabbani M, Fatemi Ghomi SMT, Jolai F, Lahiji NS (2007)
A new heuristic for resource-constrained project sche-
duling in stochastic networks using critical chain con-
cept. Eur J Oper Res 176:794–808
Stork F (2001) Stochastic resource-constrained project
scheduling. PhD Dissertation, Technische Universit
¨
at
Berlin, Berlin, Germany.
Tavares LV, Ferreira JAA and Coelho JS (1998) On the opti-
mal management of project risk. European Journal of
Operational Research 107:451–469
A Two-stage Stochastic Programming Model for the Resource Constrained Project Scheduling Problem under Uncertainty
199
Van Slyke RM and Wets RJ (1969) L-shaped linear pro-
grams with applications to optimal control and sto-
chastic programming. J. Applied. Math 17(4):638–
663
Zhu G, Bard JF, Yu G (2007) A two-stage stochastic pro-
gramming approach for project planning with uncer-
tain activity durations. J Sched 10:167–180
ICORES 2018 - 7th International Conference on Operations Research and Enterprise Systems
200