Manuskript 344
Manuskript 344
A Service of
zbw
Leibniz-Informationszentrum
Wirtschaft
Kolisch, Rainer
Manuskripte aus den Instituten für Betriebswirtschaftslehre der Universität Kiel, No. 344
Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Documents in EconStor may be saved and copied for your
Zwecken und zum Privatgebrauch gespeichert und kopiert werden. personal and scholarly purposes.
Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle You are not to copy documents for public or commercial
Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich purposes, to exhibit the documents publicly, to make them
machen, vertreiben oder anderweitig nutzen. publicly available on the internet, or to distribute or otherwise
use the documents in public.
Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen
(insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, If the documents have been made available under an Open
gelten abweichend von diesen Nutzungsbedingungen die in der dort Content Licence (especially Creative Commons Licences), you
genannten Lizenz gewährten Nutzungsrechte. may exercise further usage rights as specified in the indicated
licence.
www.econstor.eu
Nr. 344
Rainer Kolisch
May 1994
Rainer Kolisch, Institut für Betriebswirtschaftslehre, Lehrstuhl für Produktion und Logistik,
Christian-Albrechts-Universität zu Kiel, Ohlshausenstr. 40, D-24118 Kiel,
Fax +49-431-880-2072, E-Mail [email protected]
-1-
1. Introduction
2. Problem Description
In order to model the RCPSP we make use of the following additional notation: Let Pj define
the set of immediate predecessors of activity j. For ease of notation the activities are
topologically ordered, i.e. each predecessor of activity j has a smaller number than j.
Furthermore, activity j=1 (j-J) is defined to be the unique dummy source (sink) and T denotes
an Upper bound on the project's makespan. Now, a conceptual model of the RCPSP can be
formulated as follows [cf. Talbot / Patterson (1978)]:
subject to
)=2,.../,/e^ (2)
The variable FTj denotes the (integer valued) finish times of activity j, and At, the
set of activities being in progress in period t, is defined as At := {j \ FTj-dj+l <t<
FTj). The objective fiinction (1) minimises the completion time of the unique sink and thus the
makespan of the project. Constraints (2) take into consideration the precedence relations
between each pair of activities (/, j), where i immediately precedes j. Finally, constraint set (3)
limits the total resource usage within each period to the available amount. Note that (1) to (4)
provide no mechanism in order to identify At and hence the problem formulation is not
amenable to Solution via integer programming techniques. To overcome this deficiency, the
RCPSP has to be modelled with 0-1 variables as outlined in Pritsker et al. (1969).
The RCPSP is a generalisation of the static Job shop problem and hence belongs to the class
of TVP-complete problems [cf. Blazewicz et al. (1983)]. By relaxing the resource-constraints
(4), the RCPSP reduces to the CPM-case [cf. Davis (1966)] which can be solved by forward
recursion in polynomial time [cf. Elmaghraby (1977), p. 18 ff.]. On account of the inherent
intractability of the RCPSP, a multitude of exact and heuristic approaches were proposed.
Optimal procedures are dynamic programming [cf. Carruthers / Battersby (1966)], zero-
one programming [cf. Bowman (1959), Pritsker et al. (1969), Patterson / Huber (1974),
Patterson / Roth (1976)], as well as implicit enumeration with branch and bound [cf. Balas
(1971), Davis / Heidorn (1971), Hastings (1972), Radermacher (1985 / 86), Stinson et al.
(1978), Talbot / Patterson (1978), Christofides et al. (1987), Bell / Park (1990), Carlier /
Latapie (1991), Demeulemeester / Herroelen (1992)]. Currently, the branch and bound
approach of Demeulemeester and Herroelen (1992) seems to be the most powerful optimal
procedure available.
Heuristic approaches for the RCPSP basically involve five different Solution methodologies:
Single- and multi-pass priority rule based scheduling (cf. Section 3), truncated branch and
-3 -
Although belonging to the oldest Solution methodology to solve the RCPSP, priority rule
based scheduling is still the most important (heuristic) Solution technique. This is due to several
reasons: (/) The method is intuitive and easy to use which makes it highly suitable to be
employed within commercial packages, (//) the method is fast in terms of the computational
effort which recommends it to be integrated within local search approaches from Artificial
Intelligence [cf. Storer et al. (1992) and Balakrishnan / Leon (1993)]. Finally, (iii) multi-pass
implementations of the method show the best results obtainable by heuristics today [cf. Kolisch
(1994)].
Generally, a priority rule based scheduling heuristic is made up of two components, a
schedule generation scheme and a priority rule. Two different schemes can be distinguished.
The so-called serial and the parallel method. Both generale a feasible schedule by extending
a partial schedule (i.e. a schedule where only a subset of the activities has been assigned a
finish time) in a stage-wise fashion. In each stage the generation scheme forms the set of all
schedulable activities, the so-called decision set. A specific priority rule is then employed in
order to choose one or more activities from the decision set which then is scheduled. Note that
- within a single pass- each activity is only scheduled once. Both scheduling schemes are
presented in detail. For a conceptual comparison with optimal branch and bound based
procedures compare Kolisch (1994).
The serial method was proposed by Kelley (1963). It consists of stages, in each of
which one activity is selected and scheduled. Associated with each stage are two disjoint
activity-sets: In the complete set C„ are the activities which already were scheduled and thus
belong to the partial schedule. The decision set D„ contains the unscheduled activities with
every predecessor being in the complete set. In each stage one activity from the decision set is
selected with a priority rule (in case of ties the activity with the smallest activity number is
selected) and scheduled at its earliest precedence and resource feasible Start time. Afterwards,
the selected activity is removed from the decision set and put into the complete set. This, in
tum, may place a number of activities into the decision set, since all their predecessors are now
completed. The algorithm terminales at stage number n=J, when all activities are in the partial
schedule, i.e. the complete set.
To give a formal description of the serial scheduling scheme some additional notation has to
be introduced. Let nKrt, the left over capacity of the renewable resource r in period t, and D„,
the decision set, be defined as follows:
-4-
Further, let EFTj denote the earliest precedence feasible finish time of activity j within the
current partial schedule and let LFTj denote the latest precedence feasible finish time of activity
j as determined by backward recursion from the Upper bound of the project's makespan T.
Finally, let v(j) be a priority value of activity j, j e Dn. Now, the serial scheduling scheme
(SSS) can be formally described as follows:
SSS
Utilising the serial method in a single-pass environment, results were published by Pascoe
(1966), Müller-Merbach (1967), Gonguet (1969), Fehler (1969), Cooper (1976 and 1977),1
Boctor (1990), and Valls et al. (1992). Boctor performed a computational study on the basis of
36 small instances from the literature. For the two best priority rules employed, he reported an
average increase of 9.13% above the optimal objective fiinction.
Today, two algorithms are associated with the so-called parallel method: The algorithm of
Kelley (1963) and the one of Brooks [cf. Bedworth / Bailey (1982)] which is also termed
"Brooks algorithm" (BAG). Like in the majority of publications, the scheduling scheme as
proposed by Brooks is employed herein and referred to as parallel method.
The parallel method consists of at most J stages in each of which a set of activities (which
might be empty) is scheduled. Associated with each stage n is a schedule time tn and three
disjoint activity-sets: Activities which are completed up to the schedule time are in the
complete set Cn. Activities which are already scheduled, but during the schedule time still
1 Misleading, Cooper terms his scheduling scheme to be parallel (serial) when using priority rules in a dy-
namic (static) f ashion. But, as already pointed out by Valls et al. (1992) he clearly employed a serial
scheduling scheme.
-5-
active are in the active set A„. Finally, activities which are available for scheduling w.r.t.
precedence and resource constraints, but yet unscheduled are in the decision set Dn. The
partial schedule of each stage is made up by the activities in the complete set and the active set.
The schedule time of a stage equals the earliest completion time of activities in the active set of
the ancestral stage. Each stage is made up of two steps: (1) The new schedule time is
determined and activities with a finish time equal to the (new) schedule time are removed from
the active set and put into the complete set. This, in turn, may place a number of activities into
the decision set. (2) One activity from the decision set is selected with a priority rule (again, in
case of ties the activity with the smallest label is chosen) and scheduled, respectively started, at
the current schedule time. Afterwards, this activity is removed from the decision set and put
into the active set. Step (2) is repeated until the decision set is empty, i.e. activities were
scheduled or are not longer available for scheduling w.r.t. resource constraints. The parallel
method terminales when all activities are in the complete or active set.
Given A„, the active set, and C«, the complete set, respectively, nKr, the left over period
capacity of the renewable resource r at the schedule time, and D„, the decision set, are defined
as follows:
nKr\— Kr - Zkjr
jeAn
Dn '•= {j I j g { Cn U An }, Pj C Cn, kjr < 7tKr V r € Ä}
Now, a formal description of the parallel scheduling scheme (PSS) arises to:
PSS
Initialisation: n:=1, tn:=0, D„:={ 1}, An:=Cn:= 0, nKr~Kr V r e R, GOTO Step (2);
Computational experiments conducted with the single-pass version of the parallel method are
more frequent than those with the serial method and are reported by Alvarez-Valdes / Tamarit
(1989a and 1989b), Boctor (1990), Davis / Patterson (1975), Elsayed (1982), Lawrence
(1985), Pascoe (1966), Patterson (1973 and 1976), Thesen (1976), Ulusoy / Özdamar (1989),
-6-
Valls et al. (1992), and Whitehouse / Brown (1979). Additionally, Arora / Sachdeva (1989)
report about an Implementation of the parallel method on parallel processors. Davis and
Patterson (1975) document an average increase above the optimum of 5.6% and 6.7%,
respectively, for the two best priority rules applied to 83 of the instances employed in
Patterson (1984). On the basis of their 48 test-instances with 27 activities each, Alvarez-
Valdes and Tamarit (1989a) come up with an average increase above the Optimum of 2.89%
and 3.09% for the two best rules, respectively.
The only comparison of the serial and the parallel scheduling scheme when applied as
deterministic single-pass heuristic is reported by Valls et al. (1992). They concluded that none
of the schemes is dominant which contradicts the assumption made by Alvarez-Valdes /
Tamarit (1989a).
3.3 Sampling
The way the serial and the parallel scheduling method have been described so far is termed as
single-pass approach, i.e. one Single pass and one priority rule are employed to derive one
feasible Solution. Contrary, multi-pass procedures perform Z Single passes in order to generale
a sample of at most Z unique feasible solutions, where the best one is chosen. Basically, two
different kinds of multi-pass methods can be distinguished: The multi-priority rule approach
[cf. Lawrence (1985), Boctor (1990), and Li / Willis (1992)] employs one scheduling scheme
and different priority rules while sampling [cf. Levy et al. (1962), Wiest (1967), Cooper
(1976), and Alvarez-Valdes / Tamarit (1989b)] makes use of one scheduling scheme and one
priority rule. Different schedules are obtained by biasing the selection of the priority rule
through a random device. The use of a random device can be interpreted as a mapping
y/ J &Dn-+[0,1] (5)
which at stage n assigns to each activity in the decision set Dn a probability yAj) of being se-
lected (where Xyeo„ yAj) =1 holds). Three different methods can be distinguished: (/) Random
sampling assigns each activity in the decision set the same probability. (ii) Biased random
sampling biases the probabilities dependent on the priority values of the activities to favour
those activities which seem to be a more sensible choice [cf. Baker (1974, p. 72)]. In the
context of the job shop problem, biased random sampling is usually referred to as probabilistic
dispatching [cf. Conway et al. (1967, p. 124) and Baker (1974, pp. 202-206)]. (iii) A special
case of biased random sampling is the utilisation of regret measures for determining the
selection probabilities. This was introduced by Drexl (1991) and Drexl / Grunewald (1993) and
is referred to as regret based biased random sampling. Let a priority rule be defined by the
mapping v : j e Dn —> R> o which assigns to each activity j in the decision set Dn a priority
value v(J) and an objective O stating whether the activity of the decision set with the minimum
(0=min) or maximum (<9=max) priority value is selected. Then, the regret pj compares the
priority value of activity j with the worst consequence in the decision set as follows:
-7-
(6)
pj:= 1 rfft"
v(j) - min v(z), if 0=max
isD„
(?)
i^Dn
Adding the constant "1" to the regret value pj assures that the selection probability for each
activity in the decision set is greater zero and thus every schedule of the population may be
generated. By choice of the parameter a, the amount of bias can be controlled. Associated
with an arbitrary large a will be no bias and thus deterministic activity selection on the basis of
the employed priority rule (with random selection as a tie breaker) while an a of 0 will give
way for random activity selection.
Sampling applications ofthe serial method are documented by Cooper (1976) while sampling
efforts on the basis of the parallel method are reported by Wiest (1967) and Alvarez-Valdes /
Tamarit (1989b). Employing a sample size of 100, Cooper (1976) compared deterministic
scheduling and biased random sampling with 9 different priority rules on one benchmark-
instance. He concluded that sampling produces results which are (at the 99% level of
confidence) at least 7% better than the solutions derived by the deterministic approach.
Alvarez-Valdes / Tamarit (1989b) compared a single-pass and a sampling approach on a set of
48 instances with 103 activities each. The sampling approach generated for every instance 100
solutions. For the best (second best) priority rule an average increase above an Upper bound of
3.23% (3.45%) when used in the single-pass procedure and 2.31% (1.65%) when used in the
sampling procedure is reported.
To the best of our knowledge, no comparison of the serial and the parallel scheduling scheme
when applied as sampling procedures is reported in the literature.
4. Theoretical Results
Three Important questions associated with each heuristic are: (/) Is a feasible Solution always
granted, (/'/') w ill the heuristic produce optimal solutions for particular instances which are no
more longer in NP, and (iii) on what kind of Solution space is the heuristic operating, i.e. will
the heuristic generally be capable to produce optimal solutions. In order to provide deeper in-
sight into the serial and the parallel scheduling scheme these three questions will be answered
in the sequel.
Theorem 1: For any (feasible) instance of the RCPSP a feasible schedule is always generated
by each of the two scheduling schemes, respectively.
Proof: Whenever an activity is scheduled, both scheduling schemes take into account
precedence and resource constraints. More precisely: Within the serial scheduling scheme, line
3 of stage n meets the precedence constraints, while the resource constraints are preserved by
-8-
line 4 of stage n for each activity For the parallel scheduling scheme, both constraints
are insured by definition of the decision set.•
Theorem 2 . For any resource-unconstrained instance of the RCPSP both scheduling schemes
always derive the optimal Solution.
Proof: For the CPM-case resource constraints are not binding anymore and activities are
scheduled as early as possible w.r.t. precedence constraints only by each of the scheduling
methods, respectively. More detailed: The serial scheduling scheme starts the chosen activity
as early as possible (line 4 of stage n). Within the parallel scheduling scheme an activity, say j,
will enter the decision set at the schedule time /«=max {FTj | i e Pj}, i.e. as soon as its
predecessors are finished. On account of abundant resources j is scheduled in step (2) of the
same stage, i.e. without an increment of the schedule time.•
Theorem 3: a) A schedule S generated with the serial scheduling scheme and any priority
rule belongs to the set of active schedules while b) a schedule S generated with the parallel
scheduling scheme and any priority rule belongs to the set of non-delay schedules..
Proof: As a prerequisite to prove Theorem 3, we have to introduce the notion of local and
global left shifts for active and non-delay schedules (for a general Classification of schedules for
the RCPSP cf. to Sprecher et al. (1994)):
Local and global left shifts can be briefly explained as follows: Starting with a feasible schedule
S=z(FT1,...,FTJ,...JrTj) we ceteris paribus assign activity j the earlier finish time FT"P i.e. we are
"left shifting" activity j from FTj to FT') with FT') < FTj. We evaluate the resulting schedule
S"=(FTh...,FT"„...,FTj) and all intermediate schedules S'={FTh...,FT'p...,FTj) with FT" < FT)
< FTj. If the resulting schedule and all intermediate schedules are feasible we have performed a
local left shift of activity j, if the resulting schedule is feasible and at least one intermediate
schedule is infeasible we have performed a global left shift of activity j. Now, an active
schedule is defined as a feasible schedule where none of the activities can be locally or globally
left shifted. Contrary, a non-delay schedule is defined as a feasible schedule where none of the
sub-activities of the corresponding unit-time-duration schedule (a schedule where each activity
j is split into dj sub-activities with duration one) can be locally or globally left shifted. Note
that by definition the set of non-delay-schedules is a non-proper subset of the set of active
schedules.
a) To prove that any schedule generated by the serial scheduling scheme belongs to the set of
active schedules we proceed as follows: First we show that the schedule is (at least) active and
then we proof that it is not a non-delay schedule: Consider activity j has been selected at stage
n. Then, its time window of precedence feasible finish times is restricted from the earliest finish
time, i.e. the maximum finish time of its immediate predecessors plus its duration, to its latest
finish time. Activity j is now scheduled at the earliest contiguous resource feasible interval of dj
units length within its precedence feasible time window. Therefore, a left shift of any activity is
not possible and the schedule has to be at least active. Furthermore, it has to be proven that the
schedule is not a non-delay schedule. This can be achieved by showing that in the
-9-
5. Experimental Investigation
In order to study the Performance of the scheduling schemes, the following Statistical model
(with five factors) was employed [cf. Kurtulus / Davis (1982) and Kurtulus / Narula (1985)]:
where DEVabcmno (CPUabcmno) denotes the average deviation from the Optimum Solution (the
average running time in CPU-seconds) when the instances with the m-th, %-th, and o-th level of
- 10 -
the problem parameters NC, RF, and RS are solved with the a-th, b-th, and c-th level of the
procedure parameters PR, SS, and Z, respectively. For the errors s it is assumed that they are
mutually independent and that each s is drawn from the same continuous population.
The procedure parameters are characterised as follows: PR denotes the priority rule, SS
stands for the scheduling scheme while Z denotes the sample size. Priority rules were chosen
according to the studies of Davis / Patterson (1975), Alvarez-Valdes / Tamarit (1989a), Valls
et al. (1992), Uluzoy / Özdamar (1989), and Boctor (1990). Table 1 provides an overview of
the six priority rules which rank among the top three rules in at least one of these studies. Note
that only the studies by Boctor (1990) and Valls et al. (1992) employed the (parallel and the)
serial scheduling scheme and that none of the three best rules were applied within the serial
scheduling scheme. Additional to the notation already introduced, Sj (Sj) denotes the set of all
(immediate) successors of activity j and EFT'j denotes the earliest precedence and resource
feasible finish time of activity j. Note that in the parallel scheduling scheme EFTequals t„ + dj
for each activity in the decision set. Since within step (2) of the parallel scheduling scheme tn is
constant for all j e D„, v is equal for MSLK and LST which was proven by Davis / Patterson
(1975). Further note that LST and MSLK are listed separately because the priority rules are
employed in both schemes.
Two levels of the scheduling schemes, i.e. the parallel scheduling scheme (PSS) and the serial
scheduling scheme (SSS), were considered. The levels of the sample size are provided in the
subsequent sections.
The problem parameters are characterised as follows [for details cf. Kolisch et al. (1992)]:
The network complexity NC is the ratio of non-redundant precedence relations to the number
of activities. The resource factor RF reflects the density of the two dimensional array kjr,
j=2,...,J-\ and r=l,...,|^|. Finally, the resource strength RS measures the degree of resource-
constrainedness in the interval [0,1]. The resource strength is computed as follows: RS=(Kr
Krmirt)/(Krmax-Krmm), where Krmin is the minimal availability of resource type r in order to
assure feasibility of the RCPSP, i.e. Krmin = max {kjr[j=l,...,J}, and Krmax is the peak demand
of resource type r in a CPM-schedule.
-11-
All other problem parameters were adjusted as follows (where intervals consist of uniformly
distributed integers): The number of non-dummy activities was set to 30, i.e. J= 32, the number
of resource types was set to 4, i.e. |/?|=4, the activity duration was drawn out of the interval
[1,10]. In case of a positive resource demand, i.e. kjr > 0, the latter was drawn out of the
interval [1,10]. The number of different resource types requested by one activity was -
depending on the resource factor RF - in the ränge [1,4]. Finally, the precedence network was
generated with the following constraints: The number of immediate successors (predecessors)
of the dummy-source (dummy-sink) was set to 3, respectively, and the number of successors
(predecessors) of each non-dummy activity was drawn out of the interval [1,3].
NCm RFn
For the single-pass analysis the sample size Zc was set to " 1" and the selection of activities
was performed deterministically. Table 3 gives a comparison of the priority rules. The
Friedman test reveals a significant different Performance w.r.t. the average deviation from the
optimal Solution (5=0.0000). By pairwise application of the Wilcoxon test the following
ranking is observed (where 'V denotes better and denotes significant better):
- 12-
Table 4 demonstrates the Performance of the scheduling schemes. The ranking reveals to be
PSSxxSSS and hence confirms the conjecture that the parallel scheduling scheme is signifi
cantly (5=0.0000) superior to the serial one when used as single-pass heuristic [cf.
Alvarez-Valdes / Tamarit (1989a)]. Each priority rule performs better w.r.t. the quality of so-
lutions when applied in the parallel scheme. Even more, every priority rule with the exception
of the lower bound based rules LST and LFT is significantly better within the parallel schedul
ing scheme (s<0.0013). Nevertheless, since LST and LFT belong to the best rules in both
schemes, the serial scheduling scheme cannot be excluded a priori. This conclusion was already
drawn in the study by Valls et al. (1992).
The running time of the parallel scheduling scheme is slightly less than the one of the serial
scheme. Whereas the parallel scheme uses most of the time to update the decision set, the se
rial scheme requires the majority of the CPU-time for setting up and managing the array 1tKrt, r
e R, 1=1,...,T, which is (especially in the case of a poor upper bound for the makespan 7) very
time consuming.
% PSS SSS
6.46 9.21
.02 .03
The efFect of the problem parameters on the overall Performance is as follows (cf. Figure 1
with monotonically increasing parameter levels on the x-axis): The network complexity NC
does not reveal a significant influence (5=0.2929), whereas the efFect of the resource factor RF
and resource strength RS, respectively, is highly significant (5=0.0000). It thus can be
concluded that for single-pass scheduling schemes the influence of the problem parameters on
the Performance is w.r.t. the tendency similar as for optimal procedures [cf. Kolisch et al.
- 13 -
(1992)]: Generally, a high resource factor and a low resource strength will induce a poor
Performance. Precisely, for an ascending resource factor, the Performance of optimal
procedures is monotonically decreasing, whereas the single-pass priority rule based heuristics
reveal the lowest Performance for RF2 = 0.5.
Neither the ranking of priority rules nor scheduling schemes is significantly influenced by the
problem parameters. Regarding priority rules, this outcome confirms the conclusion made in
the studies by Cooper (1976) and Alvarez-Valdes / Tamarit (1989a). Furthermore, the
computational effort is not influenced by any of the problem parameters.
DEV
n "
11
10
9
8
7
6
5
1 2 3 4
In the following it is investigated if the conclusions which were drawn for the single-pass
case are still valid in the case of sampling. CPU-times are not provided any more because - as
was shown in the last section - they are not influenced by any of the parameters. According to
preliminary computational results [cf. Kolisch (1994)], the levels of the sample size Zc were
chosen to be {10, 40, 70, 100} and the bias parameter a was exclusively set to "1
Table 5 reveals the Performance of the priority rules. As for the single-pass case, a significant
difference between rules can be detected (5=0.0024). The (lower bound based) rules LST and
LFT perform best, MTS and MSLK ränge in the middle while WRUP and GRPW have the
worst Performance. The results of MSLK have to be interpreted with care because - like for
the single-pass approach - it shows a quite different Performance within each of the two
scheduling schemes, respectively: For the parallel scheme it performs like the "good" LST rule,
within the serial scheme MSLK gives rather poor results.
well as WRUP and GRPW) the ranking obtained is the same as for the single-pass case. (/'/)
While the difference between groups (i.e. LFT, LST vs. MTS, MSLK vs. WRUP vs. GRPW)
has about the same level of significance (s<0.0036), the difference between rules is slightly less
significant. Thus, it can be stated that, in general, priority rules which are good for single-pass
approaches are good for biased random sampling approaches and vice versa. Although not
explicitly pointed out, Alvarez-Valdes / Tamarit (1989b) attained similar results in their study.
Their rankings obtained when utilising 6 priority rules within the deterministic- and the
sampling-based parallel scheduling scheme differed only w.r.t. one rule. The contrary
Observation of Cooper (1976) is a consequence of his (not regret based) probability mapping
which in conjunction with the LFT priority rule tends to perform (pure) random sampling. Of
course, the interrelation between single-pass and sampling heuristics depends on the amount of
bias. Whereas with 0% bias the significance equals the one of the (deterministic) single-pass
case, a bias of 100% (i.e. random sampling) results in no (significant) difference between the
priority rules.
Figure 2 gives insight into the Performance w.r.t. the sample size. As expected, it is
demonstrated that increasing the sample size continuously produces better solutions. Depend-
ing on the sample size, the average Performance of the single-pass approach is thus improved
between 50% (Zc=10) and 73% (Zc=100). This is up to ten times more than observed by
Cooper (1976). Hence, it can be stated that sampling significantly outperforms the single-pass
approach (5=0.0000). This contradicts the conclusions drawn by Conway et al. (1967, p. 128)
for the job shop problem, stating that sampling reveals only modest improvement over single-
pass procedures. But it has to be noted that the marginal improvement diminishes. This implies
a growing computational effort in order to produce better solutions.
7
6
5
4
3
2L , , i —* „ 7
1 10 40 70 100 c
Figure 2: Performance as a Function of the Sample Size
Finally, Figure 3 demonstrates the efifect of the scheduling schemes w.r.t. the sample size Zc.
The overall Performance of both schemes is with 2.75% for the parallel and 2.71% for the
serial scheduling scheme almost identical. But a second glance reveals a (not significant
(S>0.1994)) different Performance w.r.t. the sample size. While the parallel scheme is clearly
superior for small sample sizes (i.e. less than 40 generated schedules), the serial scheme shows
better results for large samples. Consequently, for sampling procedures solving the RCPSP the
general superiority of non-delay schedules - as announced in Conway et al. (1967, pp. 121-
- 15 -
124) for the job shop problem when minimising the average flow time - does not hold true.
The rationale of the Observation is as follows: For small sample sizes the superiority of the
parallel scheme w.r.t. the single-pass approach is dominant. With increasing sample size, this
effect diminishes and at the same time the parallel scheme suffers from the fact that the sample
space is the set of non-delay schedules which not necessarily contains the optimal Solution. We
conjecture that this "critical sample size", i.e. 40 for the instances tested, increases when the
problem size - expressed in the number of activities - is enlarged.
Figure 3: Impact of the Sample Size on the Performance of the Scheduling Schemes
The efFect of the three problem parameters network complexity, resource factor, and re
source strength on the average Performance tumed out to be as for the single-pass case. That
is, no significant influence can be observed for the network complexity (5=0.6485) while
resource strength and resource factor, in the order mentioned, tumed out to be highly
significant (5=0.0000). The ranking of the priority rules is not significantly efFected by the
problem parameters but - deviating from the single-pass case - resource factor and resource
strength influence the ranking of scheduling schemes significantly.
Table 6 shows the ranking of the scheduling schemes w.r.t. to both resource parameters as
well as their combination (where * denotes a 1-tailed significance at the 1% level of confidence
and the hatched areas signal superiority of the serial method). Roughly, it can be stated that the
parallel scheduling scheme performs better for "hard" problems (with a high resource factor
and / or a low resource strength) while the serial scheduling scheme is better for "easy"
problems (with a low resource factor and / or a high resource strength).
- 16 -
6. Summary
On account of a detailed description of the serial and the parallel scheduling scheme the
following was proven:
(/) Both scheduling schemes generale feasible schedules which are optimal in the absence of
resource restrictions.
(«) The serial scheduling scheme generates active schedules while the parallel scheduling
scheme creates non-delay schedules. Hence, the parallel scheduling scheme searches in a
smaller Solution space than the serial scheduling scheme but with the drawback that - when
considering a regulär Performance measure - the Solution space might not contain an optimal
schedule.
Additionally, an in-depth computational study brought forth the following results w.r.t. (/)
the comparison of single-pass scheduling and biased random sampling, the Performance of (//')
priority rules, (/'«) the Performance of scheduling schemes, and finally (iv) the impact of
problem parameters:
Ad (zz): Four groups of priority rules are distinguished significantly, i.e. {LST, LFT}, {MTS,
MSLK}, {GRPW}, and {WRUP}. The ranking of priority rules is not significantly influenced
by any of the problem parameters. Furthermore, priority rules which are good for single-pass
approaches are also good for biased random sampling approaches. This relationship diminishes
with increasing bias.
Ad (///): The parallel scheduling scheme turns out to be better for the deterministic single-pass
case and sampling up to a sample size of 40. It is conjectured that this "critical sample size",
i.e. 40 for the instances tested, increases when the problem size - expressed in the number of
activities - is enlarged. The resource based parameters resource factor and resource strength
significantly effect the ranking of the scheduling schemes when employed within a sampling
approach. More precisely, the parallel scheduling scheme performs better for "hard" problems
(with a high resource factor and / or a low resource strength) while the serial scheduling
scheme is better for "easy" problems (with a low resource factor and / or a high resource
strength).
Ad (zv): The resource based problem parameters resource factor and resource strength do have
a significant influence on the solution-quality of single-pass and sampling heuristics: Generally,
a high resource factor and a low resource strength will induce a poor Performance.
- 17-
References
Alvarez-Valdes, R. and J.M. Tamarit (1989a): Heuristic algorithms for resource-constrained project schedul
ing: A review a nd an empirical analysis, in: Slowinski, R. and J. Weglarz (Eds.): Advances in project
scheduling, Elsevier, Amsterdam, pp. 113-134.
Alvarez-Valdes, R. and J.M. Tamarit (1989b): Algoritmos heuristicos deterministas y aleatorios en secuencia-
cion de proyectos con recursos limitados, Qüestiio, Vol. 13, pp. 173-191.
Arora, R.K. and R.K. Sachdeva (19 89): Distributed Simulation of resource constrained project scheduling,
Computers & Operations Research, Vol. 16, pp. 295-304.
Baker, K R. (1974): Introduction to sequencing and scheduling, Wiley, New York.
Balakrishnan, R. and V. J. Leon (1993 ): Quality and adaptability of problem-space base d neighborhoods for
resource constrained scheduling, Working Paper, Department of Industrial Engineering, Texas A&M
University, USA.
Balas, E. (1971): Project scheduling with resource constraints, in: Beale, E.M.L. (Ed.): Applications of mathe-
matical programming techniques, Engüsh University Press, London, pp. 187-200.
Bedworth, D.D. and J E. Bailey (1982): Integrated production control systems - Management, analysis, design,
Wiley, New York.
Bell, C.E. and J. Han (1991): A new heuristic Solution me thod in resource-constrained project scheduling, Na-
val Research Logistics, Vol. 38, pp. 315-331.
Bell, C.E. and K. Park (1990): Solving resource-con strained project scheduling problems by A * search, Naval
Research Logistics, Vol. 37, pp. 61-84.
Blazewicz, J., J.K. Lenstra, and A.H.G. Rinnooy Kan (1983): Scheduling subject to resource constraints:
Classification and complexity, Discrete Applied Mathematics, Vol. 5, pp. 11-24.
Boctor, F.F. (1990): Some efficient multi-heuristic procedures for resource-constrained project scheduling,
European Journal of Operational Research, Vol. 49, pp. 3-13.
Bowman, E.H. (1959): The schedule-sequencing problem, Operations Research, Vol. 7, pp. 621-624.
Carlier, J. and B. Latapie (1991): Une methode arborescente pour resoudre les problemes cumulatifs, Recherche
operationnelle, Vol. 25, pp. 311-340.
Carruthers, J.A. and A. Battersby (1966): Advances in critical path methods, Operational Research Quarterly,
Vol. 17, pp. 359-380.
Christofides, N., R. Alvarez-Valdes, and J.M. Tamarit (1987): Project scheduling with resource constraints: A
branch and bound approach, European Journal of Operational Research, Vol. 29, pp. 262-273.
Conway, R.W., W.L. Maxwell, and L.W. Miller (1967): Theory of scheduling, Addison-Wesley, Reading, Mas
sachusetts.
Cooper, D.F. (1976): Heuristics for s cheduling resource-constrained projects: An experimental investigation,
Management Science, Vol. 22, pp. 1186-1194.
Cooper, D.F. (1977): A note on serial an d parallel heuristics for resource-constrained project scheduling, Foun-
dations of Control Engineering, Vol. 2, pp. 131-134.
Davis, E.W. (196 6): Resource allocation in project network models - A survey, The Journal of Industrial Engi
neering, Vol. 17, pp. 177-188.
Davis, E.W. and G.E. Heidorn (1971): An algorithm for optimal project s cheduling under multiple resource
constraints, Management Science, Vol. 17, pp. 803-816.
Davis, E.W. and JH. Patterson (1975): A comparison of heuristic and optimum solutions in resource-con
strained project scheduling, Management Science, Vol. 21, pp. 944-955.
- 18-
Demeulemeester, E. and W.S. Herroelen (1992): A branch-and-bound procedure for the multiple resource-
constrained project scheduling problem, Management Science, Vol. 38, pp. 1803-1818.
Drexl, A. (1991): S cheduling of project networks by Job assignment, Management Science, Vol. 37, pp. 1590-
1602.
Drexl, A. and J. Grunewald (1993): Nonpreemptive multi-mode resource-constrained project scheduli ng, IIE
Transactions, Vol. 25 No. 5, pp . 74-81.
Elmaghraby, S.E. (1977): Project planning and control by network mo dels, Wiley, New York.
Elsayed, E.A. (1982): Algorithms for project scheduling with resource c onstraints, International Journal of
Production Research, Vol. 20, pp. 95-103.
Fehler, D.W. (1969): Die Variationen-Enumeration - Ein Näherungsverfahren zur Planung des optimalen Be
triebsmitteleinsatzes bei der Terminierung von Projekten, Elektronische D atenverarbeitung, Vol. 10, pp.
479-483.
Golden, B.L. and W.R. Steward (1985): Empirical analysis of heuristics, in: Lawle r, E.L, J.K. Lenstra, A.H.G.
Rinnooy Kan, and D.B. Shmoys (Eds.): The traveling salesman problem, Wiley, New York, pp. 207-249.
Gonguet, L. (1969): Comparison of three heuristic procedures for allocating resources and producing schedules,
in: Lombaers, H.J.M. (Ed.): Project planning by network analysis, North-Holland, Amsterdam, pp. 249-255.
Hastings, N.A.J. (1972): O n resource allocation in project networks, Operational Research Quarterly, Vol. 23,
pp. 217-221.
Kelley, J.E., Jr. (1963): The critical-path method: Resources planning and sc heduling, in: Muth, J.F. and G L.
Thompson (Eds.): Industrial scheduling, Prentice-Hall, New Jersey, pp. 347-365.
Kolisch, R. (1994): Project scheduling under resource constraints - Efficient heuristics for several problem
classes, PhD Dissertation, Kiel.
Kolisch, R, A. Sprecher, and A. Drexl (1992): C haracterization and generation of a general class of resource-
constrained project scheduling problems: Easy and hard instances, Research Report No. 301, In stitut für Be
triebswirtschaftslehre, Christian-Albrechts-Universität zu Kiel (under review for Management Science).
Kurtulus, I.S. and E.W. Davis (1982): Multi-project scheduling: Categorization of heuristic rules Performance,
Management Science, Vol. 28, pp. 161-172.
Kurtulus, I.S. and S.C. Narula (1985): Multi-project scheduling: Analysis of project Performance, IIE Transac
tions, Vol. 17, pp.58-66.
Lawrence, S R. (1985): Resource-constrained project scheduling - A computational comparison of heuristic
scheduling techniques, Working Paper, Graduate School of Industrial Administration, Camegie-Mellon
University, Pittsburgh, USA.
Levy, F.K., G L. Thompson, and J.D. Wiest (1962): Multiship, multishop, workload-smoothing program, Naval
Research Logistics Quarterly, Vol. 9, pp. 37-44.
Li, R.K.-Y. and J. Willis (1992): An iterative scheduling technique for resource-constrained project scheduling,
European Journal of Operational Research, Vol. 56, pp. 370-379.
Müller-Merbach, H. (1967): Ein Verfahren zur Planung des optimalen Betrie bsmitteleinsatzes bei der Ter
minierung von Großprojekten, Zeitschrift für wirtschaftliche Fertigung, Vol. 62, pp. 83-88, 135-140.
Oguz, O. and H. Bala (1994): A comparative study of computational procedures for the resource constrained
project scheduling problem, European Journal of Operational Research, Vol. 72, pp. 406-416.
Pascoe, TL. (1966): Allocation of resources C.P.M., Revue Francaise Recherche Operationelle, No. 38,
pp. 31-38.
Patterson, J.H. (1973): Alt emate methods of project s cheduling with limited resources, Naval Research Logis
tics Quarterly, Vol. 20, pp. 767-784.
Patterson, J.H. (1976): Project scheduling: The effects of problem structure on heuristic Performance, Naval
Research Logistics Quarterly, Vol. 23, pp. 95-123.
Patterson, J.H. (1984): A co mparison of exact approaches for solving the multiple constrained resource, project
scheduling problem, Management Science, Vol. 30, pp. 854-867.
Patterson, J.H. and W.D. Huber (1974): A horizon-v arying, zero-one approach to project scheduling, Manage
ment Science, Vol. 20, pp. 990-998.
- 19-
Patterson, JH. and G.W. Roth (1976): Scheduling a project under multiple resource constraints: A zero-one
programming approach, AHE Transactions, Vol. 8, pp. 449-455.
Pritsker, A.A.B., L.J. Watters, and P.M. Wolfe (1969): Multiproject scheduling with limited resources: A zero-
one programming approach, Management Science, Vol. 16, pp. 93-107.
Radermacher, F.J. (1985 / 86): Scheduling of project networks, Annais of Op erations Research, Vol. 4, pp.
227-252.
Sampson, S.E. and E.N. Weiss (1993): Local search techniques for the generalized resource constrained project
scheduling problem, Naval Research Logistics, Vol. 40, pp. 3 65-375.
Schräge, L. (1979): A more portable Fortran random number generator, ACM Transactions u Ma thematical
Software, Vol. 5, pp. 132-138.
Shaffer, L.R., J.B. Ritte r, and W.L. Meyer (1965): The critical-path method, McGraw-Hill, New York.
Sprecher, A., R. Kolisch, and A. Drexl (1994): Semi-active, active, and non-delay schedules for the resource-
constrained project scheduling problem, European Journal of Operational Research, to appear.
Stinson, J.P., E.W. Davis, and B.M. Khumawala (1978): Multiple resource-constrained scheduling using
branch and bound, AIIE Transactions, Vol. 10, pp. 252-259.
Storer, R.H., S.D. Wu, and R. Vaccari (1992): New search spaces for sequencing problems with application to
job shop scheduling, Management Science, Vol. 38, pp. 1495-1509.
Talbot, B. and J H. Patterson (1978): An efficient integer programming algorithm with network cuts for solving
resource-constrained scheduling problems, Management Science, Vol. 24, pp. 1163-1174.
Thesen, A. (1976): Heuristic scheduling of activites under resource and precedence restrictions, Management
Science, Vol. 23, pp. 412-422.
Ulusoy, G. and L. Özdamar (1989): Heuristic Performance and network / resource characteristics in resource-
constrained project scheduling, Journal of the Operational Research Society, Vol. 40, pp. 1145 -1152.
Valls, V , M.A. Perez, and M.S. Quintanilla (1992): Heuristic Performance in large resource-constrained pro-
jects, Working Paper, Departament D'Estadistica I Investigacio Operativa, Universität De Valencia, Spain.
Whitehouse, G.E. and J.R. Brown (1979): Genres: An extension of Brooks algorithm for project scheduling
with resource constraints, Computers & Industrial Engineering, Vol. 3, pp. 261-268.
Wiest, J.D. (1967): A heuristic model for scheduling large projects with limited resources, M anagement Sci
ence, Vol. 13, pp. B359-B377.