This paper deals with the Resource-Constrained Project scheduling Problem (RCPSP) under
activity duration uncertainty. Based on scenarios, the object is to minimize the worst-case
performance among a set of initial scenarios which is referred to as the min-max robustness
objective. Due to the complexity of the tackled problem, we propose the application of the
GRASP method which is qualified as a simple and effective multi-start metaheuristic. The
proposed approach incorporates an adaptive greedy function based on priority rules to
construct new solutions, and a local search with a forward-backward heuristic in the
improvement phase. Two different benchmark data sets are investigated, the Patterson set and
the PSPLIB J30 set. Comparative results show that the proposed enhanced GRASP outperforms
the basic procedure in robustness optimization.
2. 138 Computer Science & Information Technology (CS & IT)
One of the basic approaches to deal with uncertainty [4] is the robust approach having the object
to find a schedule that remains with a highest quality across a set of scenarios. A scenario
represents a problem realization which is founded by matching fixed values to uncertain problem
parameters. Inspired from the decision analysis, the scenario-based approach is a simple and
effective way to model uncertainty. With the absolute robustness objective, referred to as the
min-max objective, the aim is to minimize the maximum performance degradation among all
scenarios. However, the regret robustness objective is to minimize the maximum deviation of
solutions from optimality across all scenarios.
In the literature [5], Kouvelis and Yu have investigated the cited robustness objectives for
different combinatorial optimization problems. Although robust scheduling problems are more
blinded to reality, solution methods for robust RCPSP are not exhaustive. In [6], Al Fawzen and
Haouari have proposed a bi-objective model for RCPSP with the minimization of the makespan
and the maximization of the robustness. The problem was solved by a tabu search heuristic.
Chtourou and al. [7] have studied various robustness measures based on priority rules when
activity durations vary. The work of Artigues and Leus [8] deals with RCPSP under activity
uncertainty. Based on PLNE, the authors proposed a scenario-based bi-level problem formulation
that minimizes the absolute and relative regret robustness. In this model, a solution depends on
priority rule, also called scheduling policy. The authors have applied, in first, exact method which
has taken excessive computational time considering medium sized instances. So they were
directed towards heuristic procedures. In addition, the Genetic algorithm was simply adapted to
robust optimization problems, such as the one machine problem [9] and the robust RCPSP [10].
Heuristics and metaheuristics are also approved as efficient methods for stochastic RCPSP
(SRCSP) where the uncertainty is modeled by probabilistic distributions, and the robustness is
evaluated in terms of expected makespan. We cite the work of [11] in which metaheuristics were
well investigated to SRCPSP. Recently, the work of [12] gives promising results for RCPSP
under uncertainty.
In this context, we are encouraged to use the GRASP to the scenario-based robust RCPSP. Our
tackled optimization problem aims to maximize the absolute robustness objective. We propose a
GRASP algorithm enhanced with a forward-backward heuristic.
The next section focuses on the problem definition in deterministic and non deterministic version.
Section 3 describes the main phases of the GRASP method. In section 4, we explicit the
application of the latter method to the robust RCPSP. Computational results are given in section
5. Section 6 concludes the paper.
2. ROBUST PROJECT SCHEDUING PROBLEM
2.1. Deterministic RCPSP
A deterministic version of RCPSP consists in performing a set A of n activities on a set K of m
resources. Every activity i has a fixed processing time denoted by pi and requires rik units of
resource type k which is characterized by a limited capacity Rk that must not be exceeded during
the execution, and activities must not be interrupted. Two additive dummy activities 0 and n + 1
are used that represent to start and the end of the project, respectively. Dummy activities have
null time duration and null resource requirement. The objective of the standard RCPSP is to
construct a precedence and resources feasible schedule with the minimum makespan.
3. Computer Science & Information Technology (CS & IT) 139
Precedence constraints perform that the start time of an activity i is permitted only when all its
previous activities are finished. The Resource constraints satisfy that the use of every resource
type, at every instant, does not exceed its capacity.
A schedule S referred to the baseline schedule which is given by the list of activity finish times
(start times); let Fi (>=0) denotes the finish time of an activity i, then S=(F0, F1, …, Fn, Fn+1) and
the total project duration corresponds to the end project finish time Fn+1.
Therefore, the conceptual formulation of the RCPSP is given by the following formula:
1min +nF sc. (1)
0;;
;11;
)(
≥∈≤∑
∈+=−≤
∈
tKkRri
PhnjpFF
k
tAi
k
jjjh Λ (2)
(3)
with A(t) denotes the set of activities which are executing at time t, and Pj denotes the set of
predecessors of the activity i.
An instance of the RCPSP can be represented by a graph G = (V, E) where the set of nodes V is
defined by project activities and E contains arcs according to the precedence relations.
2.2. Min-Max robust RCPSP
The considered variability, for RCPSP under uncertainty, relies on activity durations. We use a
scenarios-based approach to model the problem variability. Hence, we construct a set of
scenarios, denotes by ∑, for optimization, let iσ be a single scenario that corresponds to a
problem realization. Each scenario is found by altering the initial activities durations with respect
to a maximal activity delay.
A feasible solution x for the robust scheduling problem is represented by an activity list; let f(x,
iσ ) denotes the makespan of the generated schedule according to x on scenario iσ . This value
defines the local performance of the solution x according to iσ . However, the global
optimization process has to find the robust schedule with the global performance across the
optimization set. Usually, the global performance is measured in terms of mean value, maximum
deviation, etc.
The object of the present work is to optimize the min-max robustness objective of RCPSP which
consists in minimizing the maximum makespan value over all scenarios. The optimization
objective is given by the following formula.
)),((maxmin ixf
i
σ
σ Σ∈
(4)
Resource and precedence constraints for the robust RCPSP are the same in the deterministic case
(Equations (2) and (3)).
4. 140 Computer Science & Information Technology (CS & IT)
2.3. Complexity
The robust scenario-based robust RCPSP with the min-max robustness objective is an NP-hard
problem as it can be reduced to the standard NP-hard deterministic version for a number of
scenarios equals to one [14].
3. GRASP METAHEURISTIC
The GRASP (Greedy Randomized Adaptive Search Procedures) is a multi-start metaheuristic
which was developed for combinatorial optimization [15, 16]. It consists of an iterative process,
in each of one, two phases are performed: a construction phase and a local search phase. The first
one permits the construction of a feasible solution iteratively, one element at once iteration.
However, the second phase performs the improvement of the recently constructed solution by a
simple local search heuristic. The best across all generated solutions is then retained.
In the construction phase, a Candidate List (CL) is generated that contains the set of the candidate
elements (edges) to be selected and added to the current partial solution. The CL is ordered with
respect to a greedy function that measures the benefit of selecting each element. Moreover, the
effective selection of one edge is done from an additive list: the Restricted Candidate List (RCL)
that regroups the best elements from the CL with highest greedy values.
The GRASP procedure combines crucial characteristics of search methods. In the one hand, it is
adaptive because the greedy function values are updated continuously depending on the current
partial solution and the considered schedule construction strategy. In the other hand, it is a
randomized-based method such that a selection of one element in the RC L is done randomly.
4. APPLICATION TO THE ROBUST RCPSP
We propose the application of the GRASP approach to the RCPSP with the optimization of the
absolute robustness so called the min-max robustness objective.
The main steps of the proposed approach are depicted in the following figure. As a multi-start
heuristic, the algorithm starts with generating gradually a new solution. This step integrates an
intensification strategy. Current solution is improved, in the second step, by a Forward-Backward
Improvement heuristic (FBI) and a Local Search heuristic (LS).
Figure 1. General steps of the enhanced GRASP approach
5. Computer Science & Information Technology (CS & IT) 141
Throughout this iterative process an elite set (ES) is generated containing the best encountered
solutions. The size of the ES is defined by a fixed parameter “nbElite”. Elite solutions are
updated iteratively.
4.1. Solution representation and robust fitness
A solution is represented by an activity list that satisfies precedence constraints. To evaluate the
global solution performance, a robust evaluation is made. We apply decoding procedure to
generate the schedule according to the activity list x and the scenario iσ . Then, the robust fitness
which measures the global solution performance is determined by the maximum makespan over
all obtained values for ∑. The Serial Schedule Generation Scheme (Serial SGS) is used as a
decoding procedure [2] to construct the schedule.
4.2. Construction phase with intensification strategy
At one iteration of the construction phase one activity is selected from an eligible set an added to
the current partial solution. We generate the list CL of candidate activities having all their
predecessors scheduled. For each activity in the CL, the corresponding greedy function value is
equals to the priority rule value. We propose the application of different priority rules: the
minimum Latest Finish Time (MLFT), the minimum of the activity free slack (MFLK), the
inverse free slack priority rule, and the critical activity based selection. The object is to study the
effect of the priority rule on robustness objective.
First activities of the CL are then copied in the RCL. The size of the latter list is denoted by
TRCL. From the constructed RCL, an activity is then chosen randomly and added at the latest
position in the partial activity list. A pool of elite solutions ES is constructed.
To ensure solutions with a high quality, we incorporate in the construction phase an
intensification strategy based on the elite set.
In fact, when the ES attempts the fixed parameterized size, then, with probability pES, we select
randomly one element to be considered at the current iteration of the construction phase. Then,
the first activity, in the elite solution, that does not appear in the partial current solution is
selected and inserted in.
The above described process is repeated until the construction of the totality of the solution is
reached.
4.3. GRASP Improvement Phase
The proposed GRASP improvement phase combines a Local Search procedure (LS) with a FBI
heuristic. The proposed Local Search starts from the recently constructed and improved solution
x. Iteratively, a local move is applied to x to generate a neighbourhood set: N(x). The proposed
move consists of the permutation of one activity of x with others nodes. The activity to be
permuted is chosen at random. Obtained feasible solutions are saved to be compared with x. The
best element over all neighbours and the current solution is retained. After a maximum number of
iterations, the search method would stop with the best solution over all neighbourhoods.
6. 142 Computer Science & Information Technology (CS & IT)
The Forward-Backward Improvement (FBI) method is one of the basic heuristic for project
scheduling. It was successfully hybridized with others methods ensuring efficiency and an
acceptable computational time increment [17]. The Forward recursion is given by the serial SGS.
However, the backward recursion is the SGS algorithm applied to the precedence-reverse graph
starting from the end project activity where priority values are determined according to the lastly
generated schedule.
5. EXPERIMENTS
5.1. Data Sets and Scenario Generation
The proposed approach was implemented in Java and ran on a portable personnel computer
equipped with an Intel® Core™ i5-2450M CPU@ 2.50 GHz 773MHz, 2.70Go of RAM.
Experiments were performed on two benchmark project instances: the Patterson data set [18], and
the PSPLIB J30 data set [19]. The first data set contains 110 instances of various projects with 3
resource types and a number of activities that vary between 6 and 51. However, the second data
set contains 480 project instances which are generated by the ProGen generator. These instances
represent different projects with only 30 activities and 4 resource types.
We generated scenarios with limited size for both the optimization and the evaluation
(simulation) set. The optimization set contains nbScen scenarios, equals to 10, used to compute
the robustness objective. The evaluation set is used for simulation to estimate the expected
makespan. The size of the evaluation set is denoted by l. A scenario is an initial problem
realization where a set of activities are modified by altering their initial durations. In fact, for 10
percent of the total project activities, we add a time increment δ which is taken from a uniform
distribution U(1, maxDelay). The latter parameter indicates the maximum activity delay which is
fixed to 10.
In order to evaluate the performance of the proposed approach, we were interested by the
following performance measures:
- The estimate Expected makespan which is calculated over the evaluation set ቀEሺC୫ୟ୶ሻ =
ଵ
∑ ሺ݂ሺ,ݔ σ୧ሻሻ
ୀଵ ቁ;
- The Standard deviation of the makespan over the evaluation set;
- The Relative Optimality gap that measures the deviation between the estimate expected
makespan and the lower bound LB, or the optimal makespan if exists, for the
corresponding deterministic project ቀ
ாሺେౣ౮ሻି
ቁ.
All results are averaged by the number of tested project instances.
7. Computer Science & Information Technology (CS & IT) 143
5.2. Performance evaluation
5.2.1. Deterministic case
It is inevitable to study the algorithm behaviour on the deterministic case. Hence, the table 1
shows results of the GRASP implementation on the J30 data set. We performed the basic GRASP
approach which based on a Local Search (LS) in the improvement phase (column 2). Then, the
basic algorithm is improved with the FBI and tested for the same instances set. We vary the
number of maximum iteration for both GRASP process (Line 2) and the local search (Line 3).
Line 4 reported the average deviation from the well-known optimal solutions in percent, for both
two GRASP implementations. The number of the obtained optimums is given in Line 5.
Table 1. Average deviations from optimal solutions J30 data set instances (the deterministic case).
GRASP-LS GRASP-LS+BFI
Iterations for
GRASP
100 100 300 3000 1000
Iterations for
LS
100/2 10 10 10 1000/4
Optimality
deviation
0.51 0.57 0.34 0.24 0.20
Optimums 396 390 419 428 434
Results for static RCPSP show the performance of the applied GRASP procedure compared with
other methods in the literature [3], especially when combined with the forward-backward
heuristic.
5.2.2. Results for Robust case
Under uncertainty, we have performed different runs of the proposed GRASP on Patterson data
set. The basic algorithm denoted as (GRASP-LS(10)) is considered as the implementation of the
GRASP approach with the LFT priority rule in the construction phase and 10 iterations of local
search procedure.
We firstly, vary the maximum number of iterations with the local search incorporated in GRASP.
Results are reported in table 2 with 1000 simulations as the size of the evaluation set. The
evaluation procedure was ran for each project instance.
Table 2. Robustness evaluation on Patterson data set (1000 simulations).
GRASP-LS(10) GRASP-LS(20) GRASP-LS(100)
Avg. Optimality
gap
0.2366 0.2384 0.2249
Avg. Standard
deviation
3.4423 3.4026 3.4121
Referring to the table 2, the local search procedure has an impact on the global performance. In
fact, an increment of the number of iterations yields to better results for robustness. However, this
8. 144 Computer Science & Information Technology (CS & IT)
parameter must be controlled to ensure the non degradation of the later objective, which is the
case with 100 iterations in the local search procedure. Table 3 contains numerical results of the
proposed GRASP approach on J30 data set under uncertainty, simulated over 100 replications.
Table 3. Robustness evaluation on J30 data set (100 simulations).
GRASP-LS(10) GRASP-LS(20) + BFI
Nunber of iterations 500 250
Optimality gap avg. 0.1349 0.1243
max. 0.2755 0.2603
Standard deviation avg. 4.2515 4.1428
max. 6.3863 5.8086
In order to study the efficiency of the enhanced GRASP approach for the robust RCPSP, we
evaluate the computational time on Patterson instances set for 1000 generated schedules, reported
in Table 4. As described in [17], the FBI heuristics needs two passes of the SGS procedure to
doubly justified the initial schedule. Thus, the number of generated schedules with the basic
GRASP algorithm and the enhanced version with a FBI heuristic is equals toሺnbIterMax × 10ሻ
andሺnbIterMax × ሺ10 + 3ሻሻ, respectively.
Table 4. Comparison between GRASP and GRASP-FBI on Patterson data set (1000 simulations).
GRASP-LS(10) GRASP-LS(10) + BFI GRASP-LS(10) + BFI
Nunber of iteration 100 75 333
Optimality gap avg. 0.2362 0.2304 0.2259
max. 1.3352 1.31197 1.3398
Standard deviation avg. 3.4106 3.4008 3.4089
max. 5.5832 5.6207 5.5526
Time(s) 2.04 1.775 7.926
In table 4, column 2 and 3 show that for maximum 1000 generated schedules, the enhanced
GRASP outperforms the basic GRASP in terms of robustness and computational time.
5.2.3. Priority rule
The idea of the present experiment is to study the effect of priority rules on robustness solution
quality. As described in section 4, the construction phase implements a priority rule to order the
Candidate List content, from which we select the TRCL best elements to the RCL.
We investigate in table 5 different priority rules based on critical path: Minimum Latest Finish
Time (MLFT), the minimum Slack (MSLK), the inverse MSLK, the critical Activity based rule.
The total activity slack is obtained by the difference between its latest and earliest start time. We
also propose to study a priority rule which based on graph structure which is the GPRW (Greatest
Rank Positional Weigth). We ran the GRASP algorithm with two different RCL size values
(TRCL).
9. Computer Science & Information Technology (CS & IT) 145
With limited size of the restricted list, the standard deviation of the estimated makespan is
decreased as we reinforce best elements in the RCL. The inverse MSLK gives better results than
the MSLK; this result can be interpreted as the inverse SLK favour activities having greatest
slack values, consequently generated schedule will be more flexible to absorb activity delays.
Table 5. The Standard deviation variation on Patterson data set (1000 simulations).
Priority rule
GRASP-LS(10)
TRCL=5 TRCL=3
MLFT 3.4721 3.4456
MSLK 3.4702 3.4871
inverse MSLK 3.4640 3.4454
Critical Activity 3.4533 3.4741
GPRW 3.4606 3.4499
6. CONCLUSIONS
This paper has presented the application of the GRASP approach to the robust RCPSP. Based on
scenarios, the object of the tackled optimization problem is to maximize the min-max
robustnesss. The proposed GRASP approach incorporates priority rules in the greedy
construction phase and two procedures in the improvement phase such as a local search and the
forward-backward heuristic. Experiments have shown the simplicity of the GRASP
implementation as a multi-start heuristic compared with other complex metaheuristic as
evolutionary-based approach. In addition, the presented meta-heuristic was efficient to deal with
uncertainty in acceptable computational time. Further works must be concentrated on the study of
the GRASP construction phase to explore diverse solution on the search space, and the
application of the algorithm to more large-sized project instances in robust scheduling.
REFERENCES
[1] Blazewicz, J., Lenstra, J., Rinnooy Kan, A., (1983), “Scheduling subject to resource-constraints:
Classification and complexity”, Discrete Applied Mathematics, Vol. 5, pp. 11–24.
[2] Kolisch, R., Hartmann, S. ,(1999), “Heuristic algorithms for the resource-constrained project
scheduling problem: Classification and computational analysis”, International Series in Operations
Research and Management Science, Weglarz,J. (ed.), Vol. 14, pp. 147-178.
[3] Kolisch, R., Hartmann, S., (2006) ,“Experimental investigation of heuristics for resource-constrained
project scheduling: An update”, European Journal of Operational Research, Vol. 174, pp. 23-37
[4] Davenport, A.J., Beck, J.C., (2000), “A survey of techniques for scheduling with uncertainty”,
Available from http://tidel. mie. utoronto.ca/publications.php.
[5] Kouvelis, P. , Yu., G., (1997), “Robust Discrete Optimization and Its Applications”. Klower
Academic Publisher.
[6] Al-Fawzan, M., Haouari, (2004), “M.: A bi-objective model for robust resource-constrained project
scheduling”, International Journal of production economics, Vol. 18, pp. 1-13.
10. 146 Computer Science & Information Technology (CS & IT)
[7] Chtourou, H., Haouari, M., (2008), “A two-stage-priority-rule-based algorithm for robust resource-
constrained project scheduling”, Comput. Ind. Eng., Vol. 55, pp. 183-194.
[8] Artigues, C., Leus, R., Talla Nobibon, F., (2013), “Robust optimization for resource-constrained
project scheduling with uncertain activity durations”, Flexible Systems and Management Journal,
Vol. 25(1-2), pp. 175-205.
[9] Sevaux, M. , Sorensen, K., (2004), “A genetic algorithm for robust schedules in a just-in-time
environment”, 4OR ,Quaterly journal of Operations Research Societies, Vol. 2(2), pp.129-147.
[10] Mogaadi, H., Fayech, B., (2015), “Scenario-Based Evolutionary Approach for Robust RCPSP”,
International Afro-European Conference for Industrial Advancement AECIA.
[11] Ballestin, F., R. Leus, (2009), “Resource-constrained project scheduling for timely project completion
with stochastic activity durations", Production and Operations Management, Vol. 18, pp. 459-474.
[12] Creemers, S, (2015), “Minimizing the expected makespan of a project with stochastic activity
durations under resource constraints”, Journal of Scheduling, Vol. 18 (3), pp. 263–273.
[13] S. Horroelen, Leux, R.., (2005), “Project scheduling under uncertainty: Survey and research
potentials”, European Journal Of Operational Research, Vol. 165(2), 289–306.
[14] Aissi, H., Bazgan, C., Vanderpooten, D. (2009), “Min–max and min–max regret versions of
combinatorial optimization problems: A survey”, European journal of operational research, Vol.
197(2), pp. 427-438.
[15] Feo, T.A., Resende, M.G.C., Smith, S., (1994), “A Greedy Randomized Adaptive Search Procedure
for Maximum Independent Set”, Operations Research, Vol. 42, pp. 860-878.
[16] Festa, P., Resende, M. G. (2009), “Effective application of GRASP”, Wiley encyclopedia of
operations research and management science.
[17] Valls, V., Ballestin, F., Quintanilla, S., (2005), “Justification and rcpsp: A technique that pays”,
European Journal of Operational Research, Vol. 165, pp. 375-386.
[18] Patterson, J. H.,( 1984) ,“A comparison of exact approaches for solving the multiple constrained
resource, Project Scheduling Problem”, Management Science, Vol. 30, p854-867.
[19] Kolisch, R., A. Sprecher, (1996), “PSPLIB - A project scheduling problem library”, European Journal
of Operational Research, Vol. 96, pp. 205-216.
AUTHORS
Hayet Mogaadi received a diploma of Engineer in Computer Science from the National School of
Computer Sciences (Tunisia) in 2003, and a Master degree in Automatic and Signal Processing from the
National Engineering School of Tunis (Tunisia) in 2005. She is a Ph.D. student in Electrical engineering at
the National Engineering School of Tunis. Her interest’s area is project scheduling.
Besma Fayech Chaar received the diploma of Engineer in Industrial Engineering from the National
Engineering School of Tunis (Tunisia) in 1999, the D.E.A degree and the Ph.D degree in Automatics and
Industrial Computing from the University of Lille (France), in 2000, 2003, respectively. Currently, she is a
teacher assistant at the University of Tunis (Tunisia). Her research interests include artificial intelligence,
decision-making systems, scheduling, and transportation systems.