Decision Making in Manufacturing and Services

Vol. 3 2009 No. 12 pp. 4972

Robust Buer Allocation for Scheduling of a Project
with Predened Milestones
Marcin Klimek

, Piotr ebkowski

Abstract. The paper discusses the problem of robust buer allocation for Resource-
Constrained Project Scheduling Problem (RCPSP) with predened milestones
1
, for which
execution deadlines have been established. To solve the problem, an algorithm is proposed
supporting insertion of unit time buers, with the simultaneous maximisation of new metrics
of arrangement robustness. The presented results of experimental research speak for usabil-
ity of the solutions proposed. The eectiveness is studied with use of test tasks
2
included
in the Project Scheduling Problem Library (PSPLIB) with additionally specied project
milestones.
Keywords: resource-constrained project scheduling, predictive-reactive scheduling, robust
buer allocation, milestones
Mathematics Subject Classication: 90B35, 90C59
Revised: 19 December 2009
1. INTRODUCTION
The application of project management in production planning has been growing.
Make-to-Stock (MTS) processes have been displaced by Make-to-Order (MTO) pro-
cesses. The Make-to-Order production is primarily used in the production of non-
standard products, customised to the ever-changing and sometimes unpredictable
requirements of the customer. Each such production order should be treated as a sep-
arate project developed in consultation with the customer. The areas in which such
projects are implemented include research and development (R&D), public work, con-
struction, IT etc.

Pope John Paul II State School of Higher Vocational Education in Biala Podlaska, Poland.
E-mail: marcin [email protected]

Department of Operations Research and Information Technology, AGH University of Science and
Technology, Krakow, Poland. E-mail: [email protected]
1
The expressions milestone, phase of a project and project phase are used interchangeably
throughout the paper.
2
The expressions task, activity, operation and, occasionally, job are used interchangeably
throughout the paper.
49
50 M. Klimek, P. ebkowski
While implementing a project providing for production as per a customers order,
uncertainty arises connected, for instance, with changing requirements, diculty in
estimating task durations etc. In recent years, the number has been steadily growing of
studies including the uncertainty connected with the dynamics of production systems
in the changing environment. In numerous survey papers (Aytug et al., 2005; Her-
roelen and Leus, 2004; Klimek and ebkowski, 2007; Van De Vonder, 2006; Vieira et
al., 2003), the authors have attempted to structure the research in project scheduling
under uncertain conditions.
Project scheduling most often includes an analysis of uncertainty pertaining to
the activity duration time (Herroelen and Leus, 2004; Kobylaski and Kuchta, 2007;
Lambrechts et al., 2006; Leus, 2003; Van de Vonder et al., 2005; Van de Vonder et al.,
2006; Van De Vonder, 2006). It is assumed that numerous other disturbances (includ-
ing adverse weather conditions, resource unavailability, irregularities in material sup-
plies etc.) may aect task durations. The problem then reduces to minimising the ef-
fect of possible changes on task durations. This approach is also applied in this paper.
Two methods are used in project scheduling under uncertain conditions (Vieira
et al., 2003):
1) dynamic scheduling and
2) predictive-reactive scheduling.
In highly changing environments, the planning phase is omitted and dynamic
scheduling is used. No schedule is being developed in the dynamic task arrangement,
also known as online scheduling. The allocation of tasks to appropriate resources
(machinery) is carried out as soon as resources of a given type (a machine) are ready
to process a new task. All information available at the time is used. Most often, the
task processing queue is being created based on a specied criterion (dispatching
rules). Priorities are attached to tasks waiting for processing and the task with the
top priority at a given time is allocated to the resources. The criteria used include
Shortest Processing Time (SPT) and Earliest Due Date (EDD).
Where the number of disturbances is not large, predictive-reactive scheduling
guarantees better results. The method comprises two, often separate phases:
1) predictive scheduling phase, connected with the production planning phase, and
2) reactive scheduling phase, connected with the production plan execution phase.
In actual systems, both phases should be used jointly (Vieira et al., 2003).
In the production planning phase, a predictive schedule, also known as a baseline
schedule, is developed. Because task arrangement is determined before the pro-
duction process starts, this phase is also known as oine scheduling. The devel-
opment of a predictive schedule prior to the project execution plays an impor-
tant role in production planning. Its advantages include the possibility of schedule-
-driven production control and the possibility of monitoring the execution timeliness
for individual project phases.
The plan having been developed, in the course of the production process in
actual production systems, it is necessary to monitor the production process in order
to react to events disturbing the predictive schedule. In this phase, known as reactive
Robust Buer Allocation for Scheduling . . . 51
scheduling, changes are introduced in the task arrangement enabling the problems
arising in the baseline schedule (e.g., connected with machine failures) to be solved.
The task arrangement is subject to changes on an ongoing basis, during the production
process; therefore, this phase is also known as online scheduling. The advantages of
schedule development put emphasis on the importance of production system stability,
which may be achieved through the implementation of a specied plan included in the
schedule. Accordingly, among the objectives of reactive scheduling, there is schedule
stability, that is as precise as possible execution of the predictive schedule. The larger
the number of changes in the schedule, the larger the nervousness of those involved
in the production, which may disrupt the production process.
In this paper, predictive-reactive scheduling is used. The study focuses on the
predictive scheduling phase, which starts with the development of a nominal schedule,
to be later made robust during proactive scheduling, also known as robust scheduling.
A robust schedule is designed to prevent instabilities in nominal schedules through,
for instance, factoring in the knowledge of variability and uncertainty of production
system parameters. The development of a proactive schedule includes the insertion of
time buers and resource buers at key points in the task arrangement. The method
facilitates the inclusion of statistical knowledge concerning potential disturbances in
the production plan, gathered through the analysis of earlier production runs.
The following techniques are used in proactive scheduling (Van De Vonder, 2006):
redundancy-based techniques,
contingent scheduling,
development of Partial Order Schedules (POS) (Policella et al., 2004),
arrangement sensitivity analysis (Hall et al., 2004).
In this paper, redundancy-based techniques are used, which reduces to the gener-
ation of schedules including time buers (Van de Vonder et al., 2005; Van de Vonder
et al., 2006) or resource buers (Lambrechts et al., 2006). Time or resource buers are
inserted in order to render the schedule robust to possible disturbances in production
processes, with use of the knowledge on the uncertainty of production parameters.
Buer insertion protects the given arrangement against a delay, if any, in the execu-
tion of a given task, but it simultaneously increases the total time of execution of all
tasks and contributes to the deterioration of schedule quality. A relation does exist
between the schedule quality and its robustness (Leus, 2003; Van de Vonder et al.,
2006).
For the project scheduling problem, dierent redundancy-based methods are used
to render an arrangement robust, including:
Critical Chain/Buer Management (CC/BM) method or Critical Chain Schedul-
ing/Buer Management (CCS/BM) method (Goldratt, 1997);
methods of estimating buer size based on the statistical knowledge for the prob-
lem of minimising weighted cost of schedule instability: time buer allocation
algorithms with variable task execution times (Van de Vonder, et al., 2005; Van
de Vonder et al., 2006; Van De Vonder, 2006), time or/and resource buer allo-
cation algorithms with variable resource availability (Lambrechts et al., 2006).
52 M. Klimek, P. ebkowski
Basic information on the methods referred to above follows.
The CC/BM method is a frequently used approach to project scheduling. The
idea of critical chain, proposed by Goldratt (1997) is based on the Theory Of Con-
straints (TOC). It has commonly been applied in the area of project management,
both in theory and in practice (e.g., in the management of IT or construction
projects). A new project management philosophy emerged, known as Critical Chain
Project Management (CCPM), which has in practice displaced the project manage-
ment methodology based on such techniques as Performance Evaluation Review Tech-
nique/Critical Path Method (PERT/CPM).
The CCPM methodology introduces the idea of a critical chain, dened as the
set of activities determining the total project execution time, with the arrangement
relations and resource-related constraints factored in the determination. With unlim-
ited resources, the denition of a critical chain coincides with that of a critical path.
The critical chain method supports a new way to manage a project. No milestones are
used. Instead of security margins for individual tasks, common buers are inserted at
strategic points along the project execution path. It is the protection of the deadline
for the entire project execution that becomes particularly important, and not timely
execution of individual tasks. The additional time buer is introduced in the rst
place, at the end of the project critical chain; this buer is referred to as the Project
Buer (PB). The introduction of project buers supports the protection of timely
project execution.
For activities out of the critical chain, feeding buers (FB) are introduced. In the
PERT/CPM methodology, for tasks out of the critical path, a time play is introduced,
which according to Parkinsons Law is wasted. In the CCPM methodology, those
time reserves are removed from all tasks and retained in the form of feeding buers.
Such buers are inserted at the points where a non-critical chain connects to the
critical chain, thus protecting the progress of critical activities against disturbances.
A feeding buer is a common time buer for an entire group of tasks comprising a
non-critical chain.
Earlier availability (availability in advance) of resources for critical activities is
achieved through the introduction of Resource Buers (RB). Resource buers are
inserted so as to guarantee the availability of resources for tasks included in the
critical chain in appropriate quantities and on appropriate times.
An advantage of the CC/BM method is that it supports easy monitoring of
the project execution progress through tracking the current use ratio of the critical
chain buer (project buer). This procedure is known as the Buer Management. The
CC/BM method focuses on the protection of tasks along the critical chain. Tasks out
of the critical chain are not buered. Non-critical tasks are executed at appropriate
times. Non-critical activities of the project are commenced on the ALAP (As Late
As Possible) basis, with time provided for buers. The ALAP approach shifts capital
expenditure in time, thus increasing the project Net Present value (NPV).
The CC/PM method has primarily been developed for projects executed by hu-
mans. Its objective is to mitigate the human-related risks of untimely project execu-
tion. There exist production environments in which the stability of execution of all
Robust Buer Allocation for Scheduling . . . 53
elements of production plan is of importance, given, for instance, the requirement to
deliver materials exactly at the task commencement time.
Therefore, it is reasonable to develop buer arrangement algorithms which would
guarantee production stability. The critical chain method is not applied to the prob-
lem considered also because of the scheduling objective set in the paper: in addition
to timely execution of the entire project, meeting other contractual deadlines for
individual project phases is also important.
The CC/BM method is not suitable in production systems where not only the
timely execution of the entire project, but also schedule stability is important. Algo-
rithms are developed designed to estimate sizes of buers protecting both the sched-
ule and the timely project completion. In project scheduling, the most often analysed
problem is uncertainty related to activity durations. The algorithms for buer size
denition include:
algorithms using information on Fixed Relative Deviation (FRD) (Herroelen and
Leus, 2004; Leus, 2003; Van de Vonder et al., 2005; Van de Vonder et al., 2006),
algorithms using information on weights attached to activities and arrangement-
related relations among activities: Resource Flow Dependent Float Factor
(RFDFF) and RFDFF+ (Herroelen and Leus, 2004; Leus, 2003; Van de Vonder
et al., 2005; Van de Vonder et al., 2006),
algorithms using information on weights attached to activities and knowledge of
activity duration variability: Starting Time Critically (STC) and STC+ (Herroe-
len and Leus, 2004; Leus, 2003; Van de Vonder et al., 2005; Van de Vonder et al.,
2006),
integer programming algorithms which arrange time buers evenly, thus maximis-
ing for all tasks free slack (FS) (Al-Fawzan and M. Haouari, 2005) or minimum
ratio of FS to activity duration (Kobylaski and Kuchta, 2007).
In this paper, a new RCPSP model is proposed, with predened project phases (mile-
stones), for which execution deadlines are dened. For the RCPSP problem with
milestones, robustness metrics are developed and an algorithm of robust buer al-
location is proposed; an algorithm which insets unit time buers so as to maximise
arrangement robustness. The algorithm has been tested with use of test tasks which
were sourced from the PSPLIB library and for which milestones were additionally
dened.
2. STATING OF THE PROBLEM
A project is a set of inter-related tasks (operations, activities) executed with use
of resources in order to achieve set objectives. Table 1 presents symbols standing
for parameters and variables used in dening the project scheduling problem with
predened milestones.
Networks (graphs) are used to graphically represent project scheduling problems.
For the considered RCPSP problem, the AON network, known also as the operation
network, is used, suitable for scheduling problems including time optimisation cri-
terion. With use of the operation network, a project is represented as an acyclic,
54 M. Klimek, P. ebkowski
connected, simple, oriented graph G(V, E). The tasks are numbered from 1 to n in
such a way that the predecessors number is always less than the successors.
Table 1. Parameters and variables symbols
G(V, E) graph describing a project in the AON (Activity-On-Node) representa-
tion,
V set of vertices corresponding to tasks,
E set of edges describing the arrangement relations between tasks,
s
i
start time of the activity i (decision variable),
d
i
duration of the activity i,
a
k
number of available resource of the type k,
r
ik
demand of the activity i for the type k resource,
z
i
planned completion time for the activity i,

i
deadline for the activity i; precisely the deadline for the execution of
the earliest project phase during which the activity has to be executed,
tm
i
deadline for the execution of the milestone i,
m number of milestones (n > m > 0).
The scheduling problem for a resource-constrained project consists in nding the
task start (or end) time vector for an accepted optimisation criterion. Additionally,
the following constraints are dened:
no-delay end-start relations between operations occur - the next operation can
start immediately upon the end of the previous one having ended (arrangement
constraints):
s
i
+d
i
s
j
(i, j) E (1)
at any time t, the use of resources by operations does not exceed the available
quantities (resource constraints), resources (e.g., workforce of machinery) are re-
newable, that is the quantity of given resource is constant irrespective of their
respective loads during earlier periods:

iS
t
r
ik
a
k
t, k (2)
The most common optimisation criterion used for the problems considered is the
minimisation of the total execution time for the entire project. Given the uncertainty
encountered in the course of project execution, the authors suggest that milestones
(perhaps contractually specied project phases dened by the projects principal)
should be dened, which would reduce the risk of unsuccessful or untimely execution
of the project. Predened milestones enable the customer (principal) to monitor the
progress of project execution. Any delay in the execution of individual milestones
results in a contractual penalty being imposed. On the other hand, timely execution
Robust Buer Allocation for Scheduling . . . 55
of a given milestone may trigger the payment of consideration for the completion of
a given project phase.
Milestones are dened by way of setting time constraints for individual activities.
For each task, a deadline is dened connected with the deadline for the closest project
phase during which a given task is to be performed (3).
z
i
< (3)
Figure 1 presents an example of an AON-type network with dened times of
monitoring the execution of project phases.
d
i
, r
i
,
i
2, 4, 18
2, 6, 15
3, 8, 18
5, 9, 10
2, 5, 18
2, 3, 15
3, 2, 15
0, 0, 0
2, 1, 18
0, 0, 18
i activity identier
0
9
1
2
3
5
6
4
7
8
i
d
i
duration of activity i
r
i
resource requirements of
activity i

i
due date for job i
1
Fig. 1. An example of an AON-type network for a project with single resource and
predened deadlines for some tasks
Subsequent phases of a project (milestones) are connected with tasks which
have the same execution time dened (
j
= 0). Let denote the set of tasks di-
rectly connected with the i-th milestone by M
i
. M
i
includes all activities with
the same deadline
j
:
M
i
= {j :
j
= tm
i
, j V } (4)
The sets M
i
are pairwise disjoint, that is each activity lies in one of
the sets M
i
only:
M
i
M
j
= , i = j, i, j = 1, 2, ..., m (5)
The deadlines tm
i
for the arranged project milestones are determined based on
the predened task deadlines
j
in such a way that the sequence is strictly increasing,
that is the following inequalities hold:
tm
i
< tm
i+1
, i < 1, m) (6)
56 M. Klimek, P. ebkowski
Let KM
i
denote the set of all activities upon the performance of which the
performance of the milestone km
i
is conditional:
KM
i
= {j : j M
i
(j P
k
, k M
i
)} (7)
where P
k
stands for the set of all (direct or indirect) predecessors of activity k.
The schedule looked for should take into consideration deadlines for individual
milestones. For the project illustrated in Fig. 1, the minimum total execution time
for the entire project is 12. The requirement of timely execution of all project phases
lengthens the total project execution time. Job 1 should be completed by the absolute
deadline (
1
= 10). Given resources availability requirement, during the execution of
Job 1, no other activity may be performed, with the exception of Job 8, which in
turn may start no sooner than at t = 7. Thus, the minimum project execution time
is the aggregate of the duration of Job 1 and duration of Jobs 3, 4, 7 and 8 (along
the critical path) and equals 14 (t
1
+t
3
+t
4
+t
7
+t
8
= 5 + 2 + 2 + 3 + 2 = 14). An
example of a schedule for the execution time 14 is presented in Fig. 2.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Time
1
2
3
4
5
6
7
8
9
10
Resources
Job 1
Job 2
Job 3
Job 4
Job 6
Job 5
Job 7
Job 8
1
Fig. 2. Schedule taking into consideration deadlines for individual activities
Minimisation of the project duration and other objective functions commonly
used for the RCPSP problem are not suitable objective functions for the considered
problem of timely execution of all phases of a project. It is necessary to dene a new
objective function. Such a denition is proposed below as formula (9).
The protection (security) level for timely execution of all phases of a project has
a decisive eect on the quality of a schedule. Let pb
i
denote the protection level for
the timely execution of the i-th phase of a project. Then the security level may be
computed according to formula (8):
Robust Buer Allocation for Scheduling . . . 57
pb
i
=
rez
i

jKM
i
d
j
(8)
where: rez
i
stands for the dierence between the absolute completion deadline
j
(dened for the i-th milestone) and the earliest possible completion time for all tasks
comprising the set KM
i
.
For the project illustrated in Fig. 1, we thus obtain:
tm
1
= 10, M
1
= {1}, KM
1
= {1}, tkm
1
= 5, rez
1
= 5, pb
1
= 1
tm
2
= 15, M
2
= {3, 4, 7}, KM
2
= {3, 4, 7}, tkm
2
= 7, rez
2
= 3, pb
2
= 3/7
tm
3
= 18, M
3
= {2, 5, 6, 8, 9}, KM
3
= {1, 2, 3, 4, 5, 6, 7, 8, 9}, tkm
3
= 21, rez
3
= 4, pb
3
= 4/21
The authors-proposed objective function for nominal scheduling, taking into consid-
eration the protection of timely execution of all phases of a project, is maximisation
F
n
dened as a weighted average of time buers rez
i
, as in formula (9).
F
n
=
m

i=1
rez
i
wm
i
(9)
where wm
i
stands for the weight ascribed to the i-th phase of a project.
The value of the weight wm
i
depends on the current (at any given time) level of
protection of the i-th milestone and is determined based on the list of milestones sorted
in the increasing order. Weights are ascribed in such a way that the less the level of
milestone protection, the larger the weight ascribed to that milestone. In this paper,
milestones are arranged in the descending order of levels pb
i
(multiple milestones with
the same level pb
i
are arranged in the ascending order of the respective numbers of
project phases); then the weights are computed as follows:
for the milestone with the minimum protection level pb
i
: wm
i
= m
2
;
for the milestone with the k-th pb
i
(according to the order assumed): wm
i
=
(mk)2;
for the milestone with the maximum protection level pb
i
: wm
i
= 1.
For instance, in the case of the problem illustrated in Fig. 1, indices pb
i
are
arranged as follows: pb
3
, pb
2
, pb
1
. Consequently, wm
1
= 1, wm
2
= 4, wm
3
= 9. The
value of the objective function F
n
is now computed as follows:
F
n
=
3

i=1
rez
i
wm
i
= 5 1 +3 4 +4 9 = 53 (10)
Using the objective function F
n
dened by formula (9) guarantees:
uniform distribution of security buers according to the levels pb
i
owing to the
appropriate denition of the weights wm
i
,
distribution of the time buer proportionally to the durations of individual mile-
stones the longer the execution time of all tasks comprising a given project
phase, the larger the buer ascribed to that phase (milestone).
58 M. Klimek, P. ebkowski
The maximum possible protection of milestone deadlines is material, given the
uncertainty (connected, e.g., with possible occurrence of production disturbances) ob-
served in actual production systems. The creation of a nominal schedule with large
time buers rezi (with large values of the objective function F
n
) would also sup-
port the insertion of time buers protecting the tasks most vulnerable to production
disturbances in the phase of robust scheduling.
3. SCHEDULING IN A SCHEDULING PROBLEM FOR A PROJECT WITH
MILESTONES
The phase of nominal scheduling includes the arrangement (determination of task
starting times) with use of the schedule quality criterion (in the case discussed, the
maximisation of protection of timely execution of milestones). The next scheduling
phase is devoted to making the schedule robust to possible production disturbances;
therefore, this phase is known as robust scheduling.
The robust scheduling has with growing frequency been used in task arrangement.
A robust schedule is developed in the phase of project planning, prior to project ex-
ecution, and is dened as an arrangement which, owing to its properties, is robust to
disturbances which may occur during the production process. Jensen (2001) denes
a robust schedule as an arrangement whose quality remains acceptable even when
unpredictable events have occurred. Al-Fawzan and Haouari (2005) have limited the
scope of the notion of robustness by suggesting that schedule robustness should be un-
derstood as the schedules ability to counteract such minor increases in task durations
which may be caused by uncontrollable factors. This paper follows his approach.
The robust scheduling phase for the RCPSP problem includes the following two
optimisation steps (Van De Vonder, 2006):
1) robust resource allocation appropriate allocation of resources for the purposes
of performing individual activities in order to develop a schedule which would
be as robust to disturbances as possible; the relevant algorithms include those
minimising the number of additional arrangement-related relations resulting from
resource allocation, e.g., ISH, ISH2 (Policella et al., 2004) or MABO (Van De
Vonder, 2006),
2) robust buer allocation performed with the xed resource allocation to tasks
and consisting in the insertion of resource or time buers before (or after) tasks
in order to make the schedule robust to temporary unavailability of resources and
variability of activity duration.
Robust allocation of time buers is subject of analysis in his paper. The ap-
propriate localisation of buers is material in the context of, inter alia, the project
execution cost. For a less disturbed schedule, lower storage cost is observed or a lower
cost of loan nancing the purchase of materials supplied just-in-time. Time buers
should be inserted at critical points of the nominal schedule, those most exposed
to disturbances, and with the largest projected instability cost. Individual tasks are
commenced as under the CC/BM methodology, in line with the ALAP principle, but
with task and milestone buers included in the schedule.
Robust Buer Allocation for Scheduling . . . 59
Time buers are designed to prevent minor increases (or, more generally,
oscillations) in task durations, where such increases are triggered by uncontrol-
lable factors. Other production disturbances are excluded from considerations as
it has been assumed that the majority of disturbances aect task durations. In
the course of project execution, task durations may lengthen for reasons dicult
to identify during project planning, such as errors in duration estimates, adverse
weather conditions, failures etc.
The need to buer tasks also results from the principle, currently applied in prac-
tice, of estimating task durations in project execution, for instance, under the CCPM
methodology. Each project is a unique undertaking, harmonised with the customers
specications. Frequently, no statistical data is available on durations of identical
tasks. However, information may be available concerning the execution of similar
tasks. Based on such information, an analysis of task complexity or other input, in
the production planning phase, duration is estimated for each task, in line with the
reasonable assessment of execution feasibility. Minimum estimate values, known as
aggressive estimates, are assumed. This is justied by the results of current research
into the eld of project management. These results indicate that assuming safe esti-
mates is less eective than assuming aggressive estimates. Where safe estimates are
used, the execution of a task, as a rule, consumes the entire time available anyway
(as predicted by Parkinsons Law), and even delays occur with a frequency slightly
lower than when aggressive estimates are used (the student syndrome). Assuming
shorter task durations enables the development of a schedule with time buers se-
curing timely execution of milestones. It is also often possible to insert additional
time buers securing task arrangement at points most vulnerable to disturbances.
The following assumptions were adopted for buer arrangement optimisation:
the larger is the number of resource types used by an activity, the more is the
activity vulnerable to disturbances; e.g., the probability is larger of temporary
resource unavailability (machine failure, an employees sickness etc.);
the longer is activity duration, the larger is the probability that it lengthens: tasks
with longer execution times reveal larger absolute variability;
a disturbance in the commencement of a more resource-consuming activity gen-
erates larger instability cost (as it disorganises the operation of a larger number
of resource types).
The assumptions adopted result from the fact that in actual production systems,
the longest and most resource-consuming tasks are critical for project execution - they
are more dicult to estimate and more vulnerable to disturbances.
The buer allocation algorithms described in Section 2 cannot be applied to the
problem of timely execution of milestones. Schedule buering should be performed
in such a way that predened proportions should be followed as best as possible in
protecting timely execution of all project milestones. The authors have adopted the
following assumptions for buer allocation in the problem considered:
these activities may be buered only which are connected with milestones for
which rez
i
> 0;
60 M. Klimek, P. ebkowski
buer insertion must not result in contractual deadlines of individual milestones
being missed.
Under the assumptions adopted, the number of dierent buer arrangements depends
on the sizes rez
i
of individual milestones. The larger is rez
i
, the larger is the number
of possible distributions of time buers for individual tasks.
Buer allocation is designed to determine the list of buers after activities B =
(B
1
, B
2
, . . . , B
n
) or before activities b = (b
1
, b
2
, . . . , b
n
), so that the resulting proactive
arrangement is as robust to disturbances as possible. Thus it is of key importance
that the level of schedule robustness be determined, that is its ability to absorb
disturbances occurring during the operation of a production system. The following
two types of robustness are dened:
quality robustness under this approach, the key task is to meet the performance
criterion, which is referred to as stability of makespan; anticipating production
disturbances is designed to minimise the deviation of the actual makespan from
the planned one;
solution robustness this approach endeavours to execute all details of the ar-
rangement as planned; the deviations are taken into consideration concerning,
for instance, starting and ending times of individual tasks, manner of resource
allocation to individual activities etc.; anticipating production disturbances is de-
signed to minimise, e.g., the total deviation of actual task starting times from the
planned ones or the number of activities in the actual arrangement rearranged
when compared with the planned arrangement.
For objective functions of robust scheduling, quality robustness metrics and solu-
tion robustness metrics are used. The metrics used in the study of the RCPSP problem
cannot be used, as they do not cover contractual deadlines of milestones. Accordingly,
the authors have developed robustness metrics dedicated to the considered problem
of timely execution of agreed milestones.
For the considered RCPSP problem with milestones, quality robustness is meant
as the ability to meet all agreed deadlines of milestones. The problem of achieving
quality robustness consists in protecting deadlines of individual milestones and thus
of the entire project. The metric of arrangement quality robustness, appropriate for
the problem considered, is the objective function of nominal scheduling dened by
formula (9). Therefore, the initial schedule developed during nominal scheduling is
robust in terms of arrangement quality.
The objective function for robust scheduling is the measure of robustness of ar-
rangement details. With the robustness of arrangement details, the schedule stability
is connected, with the proviso that robustness is the metric of arrangement quality
in the planning phase, while schedule stability is such a metric during the implemen-
tation of the plan (in reactive scheduling). Robust scheduling may be described as
endeavour to minimise the eect of disturbances onto the stability indicators of the
implemented arrangement. Also in this paper, the authors suggest that solution ro-
bustness should be measured by way of simulation and dened as a schedule stability
indicator determined through computational experiments with random generation of
production disturbances.
Robust Buer Allocation for Scheduling . . . 61
In the studies concerning the RCPSP problem, dierent reactive-scheduling ob-
jective functions are used. Some of them are also used in the development of stability
indicators for the problem of timely execution of milestones.
For a stochastic model (Lambrechts et al., 2006; Van de Vonder et al., 2005;
Van de Vonder et al., 2006; Van De Vonder, 2006), the minimisation of the weighted
instability cost is the objective function:
(S
R
, S
0
) =
n

i=1
w
i
|s
R
i
s
0
i
| (11)
where:
S
R
the realized schedule,
S
0
the planned schedule,
s
0
i
stands for the planned starting time for the operation i; it serves as the
decision variable during the development of a predictive schedule,
s
R
i
stands for the actual starting time for the operation i.
Now Al-Fawzan and Haouari (2005) consider a bi-objective function enabling an
analysis to be performed of the relation between schedule robustness to disturbances
and stability on the one hand and its optimality expressed in terms of project ex-
ecution time on the other. A parameter is dened specifying relative importance of
individual optimisation criteria and subject to the planners decisions.
Following Al-Fawzan and Haouaris approach, the RCPSP problem with mile-
stones may be dened as the problem of maximising a bi-objective function admitting
a following general description:
F

= Stab +(1 ) Qual 0 1 (12)

where:
F

stands for a function of arrangement stability and quality,

stands for a parameter specifying the importance of schedule robustness
metrics with respect to its quality,
Stab stands for the metric of arrangement stability,
Qual stands for the metric of schedule performance (quality).
For the resource-constrained scheduling problem with predened deadlines of key
activities, considered herein, the objective function should reect both the quality
criterion Qual and the arrangement stability criterion Stab. The authors propose to
minimise the following bi-objective functions (Klimek and ebkowski, 2008a):
F
1
=
n

i=1
(
K

j=1
r
ij

s
R
i
s
0
i

) +(1)
m

i=1
f
i
max
jM
i
(0, z
j
tm
i
)) (13)
F
2
=
n

i=1

s
R
i
s
0
i

+(1)
m

i=1
f
i
max
jM
i
(0, z
j
tm
i
)) (14)
62 M. Klimek, P. ebkowski
where f
i
stands for the penalty function dening the penalty for missing the deadline
of the milestone km
i
), and reecting, e.g., cost connected with milestone execution
delay by 1 unit of time.
Under the approach using the function F
1
, it is assumed that the larger the
number of resources undergoing changes, the less the stability of the production sys-
tem a delay in the commencement of a more resource-consuming activity causes
larger instability of the production system. Under the approach using the function
F
2
, it is assumed that all disturbances have the same eect on production stability,
irrespective of the activity disturbed.
The objective functions F
1
and F
2
enable an analysis of the RCPSP problem for
stability (schedule robustness) and performance (quality robustness). The coecient
lies in the interval 0, 1 and determines the importance (weight) of the stability of
production plan execution for a given project (taking into consideration a number of
factors, including the nature of the project, industry prole etc.):
when = 0, the problem consists in the timely execution of all milestones (quality
criterion);
when = 1, the problem consists in rendering the schedule robust to possible
disturbances in order to minimise the total deviation of the actual task staring
times from the planned ones (stability criterion).
For an individual milestone, the function f
i
depends on the eect of delays on
project execution. In this paper, the penalty f
i
is such a transformation of the cost
related to the delay in execution of the milestone i which renders similar values of the
metrics of stability Stab and quality Qual.
The objective functions F
1
and F
2
proposed above may be used for both pre-
dictive scheduling (in the course of which they are determined by way of simula-
tion) and reactive scheduling. They both serve as comprehensive criteria for schedule
assessment.
4. BUFFER ALLOCATION ALGORITHM FOR THE PROBLEM OF TIMELY
MILESTONE EXECUTION
In this section, the buer allocation algorithm BufR is presented, developed by the
authors and executing the procedure of unit buer insertion with view to optimising
static metrics of robustness. The BufR algorithm uses an auxiliary objective func-
tion of proactive scheduling (schedule robustness metric); the function is computable
without production system simulation.
The use of this auxiliary objective function should lead to making the schedule
robust in such a way that the instability cost (see formulae (13) and (14)) of the
actual arrangement is as low as possible.
A schedule developed in the nominal scheduling phase provides for time redun-
dancy protecting the schedule against untimely execution of milestones. The redun-
dancy may partially be used for time buers in order to render the arrangement
robust at the points most exposed to disturbances and thus generating the highest
Robust Buer Allocation for Scheduling . . . 63
instability cost. The metric of solution robustness should reect a specic nature of
a given production system. In this paper, the authors assume that a task performed
with/by a larger number of resources and more time-consuming is more exposed to
disturbances.
The authors propose the following general form of the metric of solution robust-
ness:
R =
_
n

i=1
w
i

bB
i

j=1
1
j
a
+
m

i=1
_
wm
i

bm
i

j=1
1
j
a
_

i=1
fm
i
max
jM
i
_
0, z
j
bm
i
tm
i
_
_
(15)
where:
a stands for a parameter determining buer allocation (a > 0, a = 2 has
been assumed in the computations),
bB
i
stands for the time buer of the task i(before the activity b
i
or after the
activity B
i
),
w
i
stands for the weight ascribed to the activity i,
wm
i
stands for the weight ascribed to the i-th milestone,
fm
i
stands for the cost relating to missing the deadline for the i-th milestone,
it may include, for instance, actual contractual penalties for delays,
bm
i
stands for the time buer of the i-th milestone, i.e., the number rez
i
of time units provided for the protection of timely execution of the i-th
project phase:
bm
i
=
_

_
0, if rez
i
0,
_
rez
i
_
, if rez
i
> 0
(16)
stands for the parameter specifying what portion of the time redundancy
rez
i
has been applied towards the time buer bm
i
of the i-th milestone.
The size of bm
i
depends on the buering strategy. The development of a strategy
appropriate for production environments characterised by dierent variability of task
durations has been subject to experimental analysis discussed herein.
The weights w
i
and wm
i
ascribed to individual activities and milestones may be
based on an analysis of costs connected with the instability of project execution and
may include storage cost, additional organisational cost and contractual penalties for
delays.
For the metric R, such task time buer allocation is preferred in which the buers
bB
i
are as little as possible. The even distribution of buers throughout the schedule
is striven at (the idea akin to the one proposed by Kobylaski and Kuchta (2007). For
instance (the importance of instability cost for a given activity having been neglected):
an increase in the time buer bB
i
before an activity for which bB
i
= 0 instead of before
an activity for which bB
i
= 2, increases the value of the objective function by the
factor of 3
a
. The parameter a controls the evenness of buer distribution (the larger
is a, the more even is the distribution).
64 M. Klimek, P. ebkowski
The general form of the robustness metric R is used to develop the following
metrics R
1
, R
2
and R
3
:
R
1
=
_
n

i=1
_
k

j=1
r
ij
_

bB
i

j=1
1
j
a
+
m

i=1
_
r
max

bm
i

j=1
1
j
a
_

i=1
sumr
i
max
jM
i
_
0, z
j
bm
i
tm
i
_
_
(17)
R
2
=
_
n

i=1
d
i

bB
i

j=1
1
j
a
+
m

i=1
_
d
max

bm
i

j=1
1
j
a
_

i=1
sumd
i
max
jM
i
_
0, z
j
bm
i
tm
i
_
_
(18)
R
3
=
_
n

i=1
_
bB
i

j=1
1
j
a
_
+
m

i=1
_
bm
i

j=1
1
j
a
_

i=1
n
i
max
jM
i
_
0, z
j
bm
i
tm
i
_
_
(19)
where:
n
i
stands for the number of tasks executed in the i-th project phase, i.e.,
the number of elements in the set KM
i
;
r
max
stands for the maximum aggregate demand for resources from project
activities;
sumri stands for the demand for resources from the activities comprising the
execution of the i-th milestone;
d
max
stands for the maximum duration of project activities;
sumd
i
stands for the aggregate duration of the activities comprising the exe-
cution of the i-th milestone.
Maximisation of R
1
, R
2
and R
3
renders the arrangement robust. The metric R
1
protects primarily the most resource-consuming activities. The metric R
2
is designed
to prefer buer insertion after activities with longer durations. The use of the metric
R
3
renders all tasks equally important. The metric R
1
has been designed for the
objective function dened by formula (13), while the metrics R
2
and R
3
for objective
function (14). All these metrics include milestone deadlines; missing the deadlines is
penalised.
The BufR algorithm inserts unit buers one after another. In a single step of the
algorithm, adding a time buer to each activity is tested. A time buer is increased
for the job for which the robustness metric R used (i.e., R
1
, R
2
or R
3
) takes the
largest value. The algorithm stops when adding such a buer does not enhance the
schedule robustness.
The BufR algorithm supports the insertion of buers both before and after
an activity, with the proviso that a given proactive schedule inserts buers either
exclusively after activities or exclusively before them.
The BufR algorithm is supported by the metrics R
1
, R
2
, R
3
, which have been
designed for the considered problem RCPSP with milestones and for the objective
functions of the predictive-reactive scheduling used herein (see formulae (13) and
(14)). However, the BufR algorithm is versatile: it may operate with other robust-
ness metrics based on other assumptions with respect to variability of production
parameters; it may also reect considerations of various production environments.
Fig. 3 illustrates the run of the algorithm BufR maximising the metric R.
Robust Buer Allocation for Scheduling . . . 65
//Step 1: Initialization S
0
(actual schedule with time buffers)
//as nominal schedule S
nom
without time buffers
S
0
= S
nom
//Iteration steps: in one step 1-unit buffer is inserted
repeat
bestR := 0 //best value of robustness measure in actual step
for j ob 1 to n
buf f e r ( j ob ) = buf f e r ( j ob ) + 1
ModifyS
0
i f (R(S
0
) > bestR) then
bestJob := j ob
bestR := R(S
0
)
end i f
buf f e r ( j ob ) = buf f e r ( j ob ) 1
ModifyS
0
end for
i f (R(S
0
) < bestR) then
buf f e r ( bestJob ) = buf f e r ( bestJob ) 1
end i f
unti l (R(S
0
) >= bestR)
//no improve after adding buffer in actual step
Fig. 3. Flowchart of the operation of the buer allocation algorithm BufR
5. CONSTRUCTION OF TEST TASKS
This section discusses the rules for the construction of test tasks for the RCPSP
problem with milestones. Experiments are conducted with use of the PSPLIB library
0, most commonly applied in research into the RCPSP problem and generated with use
of the project generator ProGen/max. From the PSPLIB library, test instances have
been selected comprising 30 or 90 tasks, with a single execution mode (single-mode
RCPSP), with renewable resources and minimisation criterion for project duration.
A total of 960 instances (480 30-task instances and 480 90-task ones) have been used.
For projects sourced from the PSPLIB library, the authors suggest using the
LOSM procedure of unequivocal denition of individual milestones, as this procedure
guarantees inspection of project progress over the entire execution period through an
even distribution of inspection times tm
i
:
tm
i+1
tm
i
=
tm
m
m
, i < 1, m) (20)
where tm
m
stands for the entire project deadline (tm
m
=
n+1
).
The LOSM procedure runs as follows:
Step 1. a) Development of a schedule H for the RCPSP problem with use of prior-
ity algorithm with random task priorities (a random number generator
used to generate task priorities is initialised with the seed identical to
the one used for the given test problem by ProGen). Computation of
the project duration Cmin for the random schedule H thus developed.
66 M. Klimek, P. ebkowski
b) Determination of milestone deadlines (tm
1
, tm
2
, . . . , tm
m
) according
to the formula:
i 1, m) : tm
i
= i
_
C
min
(1 +)
m
_
(21)
where stands for a parameter describing the project duration
relative lengthening (expressed as percentage) which may be used for
task buering.
Step 2. Creation of the task list L and supplementing it with all inconspic-
uous/minor tasks arranged in the ascending order of their respective
starting times specied in the schedule H generated in Step 1. When
multiple tasks have the same starting time, they are added to the list L
in the ascending order of their numbers.
Step 3. Downloading the subsequent task j from the list L and including it in the
set M
i
if the following condition holds true:
s
H
j
+d
j
< (1
i
) tm
i
(22)
where:
s
H
j
stands for the starting time of the activity j
in the schedule H,

i
stands for a parameter controlling the possibility of buer-
ing the i-th milestone.
If condition (22) is not met, the task j is added to the next set M
i
+1 and
the then current set of milestones is changed to the next one: i := i + 1.
If i = m, the procedure goes to Step 4; otherwise, Step 3 is repeated for
the next activity in the list L.
Step 3 is executed for all tasks considered in the order dened in the list
L. Step 3 supports the generation of the sets M
i
(i = 1, 2, . . . , m 1).
First, the set M
1
(i := 1) is created. In a single run of Step 3, a sin-
gle task is allocated to the appropriate set M
i
. Step 3 is rerun until i = m.
Step 4. Adding, to the set M
m
, the remaining tasks included in the list L and not
yet allocated to any of the sets M
i
(i = 1, 2, . . . , m1).
For the purposes of computational experiments, four project phases (milestones:
m = 4) are dened for each test problem. The parameters tm
i
are dened as the
common deadlines for tasks included in the same set M
i
, as well as the tasks included
in individual sets M
i
according to the LOSM procedure (in Step 1: the parameter
= 30%, in Step 3
1
= 10%,
2
= 10% and
3
= 10%).
Beside the denition of conventional milestones, the denition of task duration
variability is also material in simulations. Project execution is inescapably connected
with the uncertainty of task duration estimations, related to numerous factors, in-
cluding a large number of uncontrollable factors, unique nature of the tasks involved
etc. The authors have assumed that the planned task durations may be subject to
Robust Buer Allocation for Scheduling . . . 67
factors which are not identiable in the planning phase, such as duration estimate
errors, adverse weather conditions, failures etc. The analysis of task durations for real
projects (especially those executed by human resources) reveals that the probability
curve for task ending takes the shape of the curve. The cumulative beta distribu-
tion is used in project planning to model possible ending times at the given expected
ending time and its variability.
In this paper, the authors adopt the same parameters describing task duration
variability which have been used in other studies (Herroelen and Leus, 2004; Lam-
brechts et al., 2006; Leus, 2003; Van de Vonder et al., 2005; Van de Vonder et al., 2006;
Van De Vonder, 2006). For each task, the actual duration is randomly selected from
the discretised positively skewed beta distribution with the parameters 2 and 5, and
with the expected (average) value equal to the planned task duration. The unequivo-
cal parameter denition requires that minimum and maximum values should also be
dened. Two cases are considered of large and small task duration variability. For
each task i(i = 1, 2, . . . , n):
with high variability, the minimum value is 0.25 d
i
, while the maximum value is
2.85 d
i
;
with low variability, the minimum value is 0.75 di, while the maximum value is
1.625 d
i
.
6. RESULTS OF COMPUTATIONAL EXPERIMENTS
In this section, the eectiveness of the proposed robust buer allocation is analysed.
Experiments were performed on a computer with a Pentium 1.7 GHz processor, sup-
ported by a program implemented in C# in the Visual Studio.NET environment.
A nominal schedule was generated with use of the Simulated Annealing (SA)
metaheuristics, with the objective function dened by formula (9). The best values
of parameters of the algorithm and best solution search techniques were determined
experimentally (Klimek and ebkowski, 2008b). The next phase comprised robust
allocation of resources. Resources allocation was executed with use of the ISH
2
(It-
erative Sampling Heuristic) algorithm (Policella, et al., 2004; Policella, 2005). The
use of that algorithm enables the reduction of the number of additional arrangement
relations (known as synchronisation points) in the project activity network.
The ISH
2
algorithm allocates the activity j just analysed to these resources
(chains) in which the last activity immediately precedes the activity j. If the demand
of the activity j for resources exceeds the demand of its predecessors, the remain-
ing chains are selected by the ISH algorithm, which allocates an activity demanding
a given type of resources (r
jk
> 1) so as to maximise the number of common chains
shared with the most recent activities in the available chains.
The objective functions F
1
and F
2
dened by formulae (13-14) serve as assess-
ment criteria for robust buer allocation algorithms. The functions F
1
and F
2
are
determined by way of simulation for durations generated from the beta distribution
with a large or small variability. In the project execution phase, the simplest right-
shift rescheduling method is used, consisting in shifting the disturbed activities to the
right with resources allocation unchanged.
68 M. Klimek, P. ebkowski
The following parameters are tested in the BufR algorithm:
robustness metric used (R
1
, R
2
or R
3
),
various values of (0.25, 0.5, 0.75),
buer insertion before activities (b buers) or after activities (B buers).
For each project, three production process scenarios are generated: random task
durations sourced from the distribution with appropriate parameter values (selected
separately for large variability and for small variability). Projects (their test instances)
have dierent buering possibilities. A large variability of instability cost with respect
to the objective function of nominal scheduling renders it more dicult to analyse
the eciency of the buer allocation algorithm BufR against the parameters used
(large values of standard deviations for the instability costs obtained are observed).
Tables 2 and 3 set forth the results of the operation of the BufR algorithm with
small and large variability of durations.
Table 2. Average values of the objective functions F
1
and F
2
for selected values of
the parameters of the BufR algorithm 30-job projects (J30 set)
Algorithm
Small duration variability Large duration variability
F
1
F
2
F
1
F
2
= 0.25 = 0.75 = 0.25 = 0.75 = 0.25 = 0.75 = 0.25 = 0.75
R
1
, = 0.25, b 75.5 72.0 3.7 4.6 4,467.9 2,066.7 184.8 99.9
R
1
, = 0.50, b 71.5 112.8 3.9 7.5 4,032.2 2,091.5 168.9 105.5
R
1
, = 0.75, b 101.5 213.2 5.9 13.7 3,906.8 2,265.8 165.2 117.7
R
2
, = 0.25, b 78.9 81.1 3.8 4.8 4,508.1 2,092.6 186.6 100.7
R
2
, = 0.50, b 74.1 122.4 4.0 7.7 4,085.3 2,122.1 170.8 106.4
R
2
, = 0.75, b 102.2 215.4 5.8 13.7 3,922.9 2,268.9 165.8 117.6
R
3
, = 0.25, b 77.3 74.1 3.7 4.4 4,491.9 2,071.7 186.0 99.4
R
3
, = 0.50, b 74.0 121.3 3.9 7.6 4,055.1 2,108.0 170.0 105.7
R
3
, = 0.75, b 103.4 219.1 5.9 13.7 3,936.8 2,281.8 166.4 118.0
R
1
, = 0.25, B 67.7 81.1 3.5 5.2 4,220.1 2,040.0 174.9 100.6
R
1
, = 0.50, B 76.9 136.2 4.3 8.9 3,903.3 2,117.8 164.2 109.0
R
1
, = 0.75, B 112.5 247.4 6.6 16.0 3,824.4 2,310.3 162.7 122.4
R
2
, = 0.25, B 61.8 68.1 3.1 4.3 4,162.0 2,003.0 172.8 99.0
R
2
, = 0.50, B 70.0 118.0 3.9 7.9 3,881.0 2,088.9 163.0 107.4
R
2
, = 0.75, B 107.9 233.9 6.4 15.4 3,800.3 2,282.4 161.7 121.2
R
3
, = 0.25, B 69.1 78.4 3.6 5.0 4,228.3 2,027.2 175.6 100.0
R
3
, = 0.50, B 77.3 139.7 4.4 9.3 3,889.4 2,114.1 164.0 109.1
R
3
, = 0.75, B 112.9 248.5 6.7 16.2 3,839.4 2,316.4 163.4 122.9
Nominal 162.7 399.2 10.1 26.6 3,850.1 2,531.8 166.0 138.9
Robust Buer Allocation for Scheduling . . . 69
Table 3. Average values of the objective functions F
1
and F
2
for selected values of the
parameters of the BufR algorithm 90-job projects (J90 set)
Algorithm
Small duration variability Large duration variability
F
1
F
2
F
1
F
2
= 0.25 = 0.75 = 0.25 = 0.75 = 0.25 = 0.75 = 0.25 = 0.75
R
1
, = 0.25, b 482.1 379.4 20.4 21.3 35,469.8 15,680.6 1,315.9 683.5
R
1
, = 0.50, b 456.3 634.7 21.5 37.6 33,702.9 16,377.4 1,258.5 745.3
R
1
, = 0.75, b 660.9 1,322.4 34.3 78.8 33,247.7 17,722.1 1,254.2 840.0
R
2
, = 0.25, b 508.1 421.2 21.2 22.6 36,069.2 15,993.7 1,335.5 694.9
R
2
, = 0.50, b 466.3 653.6 21.6 37.6 34,177.3 16,542.5 1,274.9 749.6
R
2
, = 0.75, b 655.5 1,304.2 33.7 76.9 33,383.1 17,732.0 1,258.7 838.4
R
3
, = 0.25, b 522.9 427.4 22.0 22.5 35,658.1 15,772.1 1,325.1 684.3
R
3
, = 0.50, b 502.1 756.8 23.4 42.3 33,751.9 16,479.5 1,264.3 748.2
R
3
, = 0.75, b 699.0 1,422.1 35.7 82.4 33,504.9 17,876.4 1,264.9 845.2
R
1
, = 0.25, B 401.8 394.4 17.6 22.4 33,982.9 15,460.9 1,260.6 685.2
R
1
, = 0.50, B 463.6 722.2 22.3 42.4 32,756.8 16,365.7 1,227.7 758.5
R
1
, = 0.75, B 707.9 1,491.5 37.5 89.5 32,711.3 17,869.6 1,239.5 861.4
R
2
, = 0.25, B 362.1 330.9 15.5 18.6 33,955.5 15,409.5 1,258.2 682.2
R
2
, = 0.50, B 421.0 625.9 20.2 37.3 32,841.6 16,282.6 1,227.1 752.1
R
2
, = 0.75, B 671.4 1,386.8 35.8 84.4 32,618.9 17,709.7 1,235.1 853.6
R
3
, = 0.25, B 433.5 429.8 19.1 24.2 34,028.7 15,453.9 1,263.5 684.1
R
3
, = 0.50, B 501.4 833.4 24.4 48.6 32,886.8 16,495.6 1,232.8 764.1
R
3
, = 0.75, B 742.3 1,582.5 39.6 95.2 32,947.7 18,007.3 1,248.9 868.3
Nominal 1,104.4 2,690.3 64.7 171.2 33,047.8 19,434.0 1,270.1 977.4
F
1
, F
2
stand for the average value of the objective function (instability cost) F
1
and F
2
, respectively.
For the majority of parameter congurations, the use of buer allocation enhances
arrangement stability when compared with the execution of a nominal schedule with-
out time buers. The larger the importance of stable execution of individual tasks
(the larger the value of ), the larger the importance of time buers.
For the robustness metrics proposed, the insertion of buers after activities proves
more eective than their insertion before activities. When buers are inserted before
activities, the metric R
1
is most eective. When buers are inserted after activities,
the metric R
2
is most eective. For the majority of the objective functions analysed,
the most eective algorithm is the algorithm BufR optimising the metric R
2
with
= 0.25, inserting buers after activities. This means that it is eective to rst insert
buers protecting the schedule against the lengthening of most time-consuming tasks.
The parameter has major eect on the eectiveness of the BufR algorithm.
Schedules generated with = 0.25 prove most stable. It is advisable to identify
70 M. Klimek, P. ebkowski
the most eective value of parameter for each project, as this value has an enormous
eect on the robustness of a proactive schedule.
A more detailed analysis of the eect of parameter on production stability was
performed for a test task J30 1 3 with F
n
= 392. One hundred duration disturbance
scenarios were generated from the beta distribution. Table 4 and Fig. 5 present results
of experiments run for the BufR algorithm with the following parameters: metric R
2
and buer insertion after activities with small task duration variability. In simulations
was analyzed the objective function F
2
with = 0.25.
Table 4. The average value of F
2
for selected values of parameter with small task
duration variability
F
2
F
2
F
2
F
2
0.00 16.50 0.25 5.9 0.50 14.5 0.75 23.1
0.05 16.50 0.30 7.1 0.55 18.8 0.80 28.3
0.10 5.34 0.35 8.2 0.60 19.5 0.85 36.5
0.15 5.34 0.40 8.2 0.65 19.5 0.90 40.4
0.20 5.76 0.45 11.6 0.70 21.6 0.95 50.5
0 0.2 0.4 0.6 0.8 1
0
10
20
30
40
50
60

F
2
1
Fig. 4. The instability cost F
2
as a function of parameter describing what proportion of
the time reserve rez
i
has been applied towards the time buer bm
i
The parameter exerts an enormous impact on production stability. Its optimum
value depends on numerous factors, including the specic nature of the test task
involved and the variability of task durations.
Robust Buer Allocation for Scheduling . . . 71
7. SUMMARY
The paper discusses a scheduling model for resource-constrained projects with pre-
dened deadlines for certain tasks connected with project milestones. The model
proposed may prove very useful for the execution of large production, construction or
development orders, because progress inspection times (points) are commonly dened
for large projects. Timely execution of milestones mitigates the risk of unsuccessful
execution of the entire project.
For the model developed, the authors also develop an algorithm for robust al-
location of time buers. The results of computational experiments conrm the eec-
tiveness of the algorithm proposed.
The subject of further research will include the development of eective algo-
rithms of reactive scheduling for the problem dened.
REFERENCES
Al-Fawzan M., Haouari, M. (2005). A bi-objective problem for robust resource-constrained
project scheduling. International Journal of Production Economics, 96, pp. 175187.
Aytug, H., Lawley, M., McKay, K., Mohan, S., Uzsoy, R. (2005). Executing production sched-
ules in the face of uncertainties: A review and some future directions. European Journal
of Operational Research, 161(1), pp. 86-110.
Goldratt, E.M. (1997). Critical chain. Great Barrington: The North River Press.
Hall, N.G., Posner, M.E. (2004). Sensitivity Analysis for Scheduling Problems. Journal of
Scheduling, 7(1), pp. 4983.
Herroelen, W., Leus R. (2004). Robust and reactive project scheduling: a review and classica-
tion of procedures. International Journal of Production Research, 42(8), pp. 15991620.
Jensen, M.T. (2001). Improving robustness and exibility of tardiness and total ow-time job
shops using robustness measures. Applied Soft Computing, 1, pp. 3552.
Klimek, M., ebkowski P. (2007). Predictive-Reactive Project Scheduling. in Innovations
technologies in economics and innovative management (ed. J. Duda), Uczelniane
Wydawnictwa Naukowo-Dydaktyczne Akademii Grniczo-Hutniczej, Krakw, pp. 198
206.
Klimek, M., ebkowski P. (2008a). Miary odpornoci harmonogramw [Schedule Robust-
ness Metrics, in Polish]. in Komputerowo Zintegrowane Zarzdzanie (ed. R. Knosala),
Ocyna Wydawnicza Polskiego Towarzystwa Zarzdzania Produkcj, Opole, Vol. I,
pp. 569577.
Klimek, M., ebkowski P. (2008b). Algorytmy metaheurystyczne dla problemu harmono-
gramowania projektu z kamieniami milowymi [Metaheuristics Algorithms for Schedul-
ing Problem for Projects with Milestones, in Polish], Zeszyty Naukowe Politechniki
lskiej, Series: Automatyka, Fasc. 150, pp. 6372.
Kobylaski, P., Kuchta D. (2007). A note on the paper by M. A. Al-Fawzan and M. Haouari
about a bi-objective problem for robust resource-constrained project scheduling. Interna-
tional Journal of Production Economics, 107, pp. 496501.
Kolisch, R., Sprecher A. (1997). PSPLIB a project scheduling library, European Journal
of Operational Research, 96, pp. 205216.
Lambrechts, O., Demeulemeester, E. , Herroelen, W. (2006). Proactive and reactive strategies
for resource-constrained project scheduling with uncertain resource availabilities. Report
KBI 0606, K. U. Leuven.
72 M. Klimek, P. ebkowski
Leus, R. (2003). The generation of stable project plans, PhD thesis at K. U. Leuven, Belgium.
Policella, N., Oddi, A., Smith, S., Cesta, A. (2004). Generating robust partial order schedules.
in Proceedings of CP2004, Toronto, Canada.
Policella, N. (2005). Scheduling with Uncertainty A Proactive Approach using Partial Order
Schedules. PhD thesis at La Sapienza Universita, Rome.
Van de Vonder, S., Demeulemeester, E., Herroelen, W., Leus, R. (2005). The use of buers
in project management: The trade-o between stability and makespan. International
Journal of Production Economics, 97, pp. 227240.
Van de Vonder, S., Demeulemeester, E., Herroelen, W., Leus, R. (2006). The trade-o between
stability and makespan in resource-constrained project scheduling, International Journal
of Production Research, 44(2), pp. 215236.
Van De Vonder, S. (2006). Proactive-reactive procedures for robust project scheduling, PhD
thesis at K. U. Leuven, Belgium.
Vieira, G.E., Herrmann, J.W., Lin, E. (2003). Rescheduling manufacturing systems: a frame-
work of strategies, policies and methods, Journal of Scheduling, 6(1), pp. 3558.