0% found this document useful (0 votes)
12 views14 pages

Kolisch 1996

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
12 views14 pages

Kolisch 1996

Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 14

EUROPEAN

JOURNAL
OF OPERATIONAL
RESEARCH
ELSEVIER European Journal of Operational Research 90 (1996) 320-333

Serial and parallel resource-constrained project scheduling


methods revisited: Theory and computation
Rainer Kolisch *
lnstitut fftr Betriebswirtschaftslehre, Lehrstuhl ffir Produktion und Logistik, Christian -A lbrechts- Universit~t zu Kiel,
Olshausenstr. 40, D-24098 Kiel, Germany
Received August 1994; revised April 1995

Abstract

We consider the so-called parallel and serial scheduling method for the classical resource-constrained project scheduling
problem. Theoretical results on the class of schedules generated by each method are provided. Furthermore, an in-depth
computational study is undertaken to investigate the relationship of single-pass scheduling and sampling for both methods. It
is shown that the performance-ranking of priority rules does not differ for single-pass scheduling and sampling, that
sampling improves the performance of single-pass scheduling significantly, and that the parallel method cannot be generally
considered as superior.

Keywords: Resource-constrained project scheduling; Serial and parallel scheduling method; Single-pass scheduling and sampling; Active
and non-delay schedules; Priority rules; Experimental investigation

1. Introduction r ~ R during every period of its duration. Since


resource r is only available with the constant period
The classical resource-constrained project sched- availability of K r units for each period, activities
uling problem (RCPSP) can be stated as follows: We might not be scheduled at their earliest (precedence
consider a single project which consists of j = feasible) start time but later. The objective of the
1 . . . . , J activities with a non-preemptable duration RCPSP is to schedule the activities such that prece-
of d i periods, respectively. The activities are interre- dence and resource constraints are obeyed and the
lated by two kinds of constraints: Precedence con- makespan of the project is minimised.
straints - as known from traditional CPM-analysis - Two of the oldest and best known heuristics for
force an activity not to be started before all its the RCPSP are the serial and the parallel scheduling
predecessors have been finished. Additionally, re- scheme, respectively (cf. the literature cited in Sec-
source constraints arise as follows: In order to be tion 3). Both can be applied as a deterministic
processed, activity j requires kit units of resource single-pass or a probabilistic multi-pass (sampling)
method.
The majority of publications dealing with schedul-
* Fax +49-431-880-2072, E-maih [email protected] ing schemes for the RCPSP report on the perfor-

0377-2217/95/$09.50 © 1995 Elsevier Science B.V. All rights reserved


(95)00357-6
R. Kolisch / European Journal of Operational Research 90 (1996) 320-333 321

mance of one scheme when applied as a single-pass finish times of activity j, j = 1. . . . . J, and At, the
approach only. Solely Vails et al. (1992) provide a set of activities being in progress in period t, is
direct comparison of the deterministic serial and defined as A t : = { j l j = l . . . . . J, F T j - d j + I < t <
parallel scheduling scheme. Investigations on sam- FTfl. The objective function (1) minimizes the com-
pling applications to solve the RCPSP were reported pletion time of the unique sink and thus the makespan
by Levy et al. (1962) and Wiest (1967). Finally, of the project. Constraints (2) take into consideration
Alvarez-Valdes and Tamarit (1989b) and Cooper the precedence relations between each pair of activi-
(1976) studied deterministic single-pass and proba- ties (i, j), where i immediately precedes j. Finally,
bilistic sampling approaches. constraint set (3) limits the total resource usage
The purpose of this paper is to provide an exten- within each period to the available amount. Note that
sive comparison of the parallel and the serial (1) to (4) provide no mechanism in order to identify
scheduling scheme. The remainder is organised as A, and hence the problem formulation is not
follows: Section 2 gives a formal model of the amenable to solution via integer programming tech-
RCPSP as well as an overview of the available niques. To overcome this deficiency, the RCPSP has
solution procedures. Section 3 presents the determin- to be modelled with 0-1 variables as outlined in
istic algorithm of the serial and the parallel schedul- Pritsker et al. (1969).
ing schemes, respectively, as well as their probabilis- The RCPSP is a generalisation of the static job
tic extensions. The ability of the scheduling schemes shop problem and hence belongs to the class of
to derive feasible and optimal solutions is covered by NP-hard problems (Blazewicz et al. 1983). By relax-
observations made in Section 4. Section 5 is devoted ing the resource-constraints (3), the RCPSP reduces
to an in-depth computational study based on a set of to the CPM-case (Davis 1966) which can be solved
360 systematically generated instances. Finally, Sec- by forward recursion in polynomial time (Elmagh-
tion 6 comes up with a summary of the derived raby, 1977, p. 18 ft.). On account of the inherent
results. intractability of the RCPSP, a multitude of exact and
heuristic approaches were proposed.
Optimal procedures are dynamic programming
2. Problem description (Carruthers and Battersby 1966), zero-one program-
ming (Bowman, 1959; Pritsker et al., 1969; Patterson
In order to model the RCPSP we make use of the and Huber, 1974; Patterson and Roth, 1976), as well
following additional notation: Let Pj define the set as implicit enumeration with branch and bound
of immediate predecessors of activity j. For ease of (Balas, 1971; Davis and Heidom, 1971; Hastings,
notation the activities are topologically ordered, i.e. 1972; Radermacher, 1985/86; Stinson et al., 1978;
each predecessor of activity j has a smaller number Talbot and Patterson, 1978; Christofides et al., 1987;
than j. Furthermore, activity j = 1 ( j = J ) is defined Bell and Park, 1990; Carlier and Latapie, 1991;
to be the unique dummy source (sink) and T denotes Demeulemeester and Herroelen, 1992). Currently,
an upper bound on the project's makespan. Now, a the branch and bound approach of Demeulemeester
conceptual model of the RCPSP can be formulated and Herroelen (1992) seems to be the most powerful
as follows [cf. Talbot/Patterson (1978)]: optimal procedure available.
Heuristic approaches for the RCPSP basically
Min FI'j ( 1)
involve five different solution methodologies: Sin-
subject to gle- and multi-pass priority rule based scheduling
(cf. Section 3), truncated branch and bound proce-
FTi<FTi-di j=2 ..... J,i~Pg, (2)
dures (Alvarez-Valdes and Tamarit, 1989a), integer
kj~<_K r rER, t = I . . . . . T, (3) programming based heuristics (Oguz and Bala,
jEAt 1994), disjunctive arc concepts (Shaffer et al., 1965;
Alvarez-Valdes and Tamarit, 1989a; Bell and Han,
FTj>O j=l . . . . . J. (4)
1991), and local search techniques (Sampson and
The variable FTj denotes the (integer valued) Weiss, 1993; Leon and Balakrishnan, 1995).
322 R. Kolisch / European Journal of Operational Research 90 (1996) 320-333

3. Priority rule based scheduling is removed from the decision set and put into the
scheduled set. This, in turn, may place a number of
Although belonging to the oldest solution method- activities into the decision set, since all their prede-
ology to solve the RCPSP, priority role based cessors are now scheduled. The algorithm terminates
scheduling is still the most important (heuristic) solu- at stage number n = J, when all activities are in the
tion technique. This is due to several reasons: (i) The partial schedule, i.e. the scheduled set.
method is intuitive and easy to use, which makes it To give a formal description of the serial schedul-
highly suitable to be employed within commercial ing scheme some additional notation has to be intro-
packages. (ii) The method is fast in terms of the duced. Let 7rKrt , the left over capacity of the renew-
computational effort which recommends it to be able resource r in period t, and D,, the decision set,
integrated within local search approaches from artifi- be defined as follows:
cial intelligence (Storer et al., 1992; Leon and Bal-
rrK~, := K~ - Y'. kj•,
akrishnan, 1993). Finally, (iii) multi-pass implemen- jeA,
tations of the method show the best results obtain-
able by heuristics today (Kolisch, 1995).
Generally, a priority rule based scheduling heuris- Further, let EFTj denote the earliest precedence
tic is made up of two components, a schedule gener- feasible finish time of activity j within the current
ation scheme and a priority rule. Two different partial schedule and let LFTj denote the latestprece-
schemes can be distinguished: The so-called serial dence feasible finish time of activity j as determined
and the parallel method. Both generate a feasible by backward recursion from the upper bound of the
schedule by extending a partial schedule (i.e. a project's makespan T. Finally, let v ( j ) be a priority
schedule where only a subset of the activities has value of activity j, j ~ / 9 , . Now, the serial schedul-
been assigned a finish time) in a stage-wise fashion. ing scheme (SSS) can be formally described as
In each stage the generation scheme forms the set of follows:
all schedulable activities, the so-called decision set.
A specific priority rule is then employed in order to [sssl
choose one or more activities from the decision set Initialisation: n := 1, Sn :-- ~;
which then is scheduled. Note that - within a single WHILE I anl < J DO Stage n
pass - each activity is only scheduled once. Both BEGIN
scheduling schemes are presented in detail. For a COMPUTE D~ and 7 r g r t , t = 1. . . . . T, r E R;
conceptual comparison with optimal branch and j* := minj~ o~{Jl v ( j ) = infi~ o {v(i)}};
bound based procedures, see Kolisch (1995). EFTj. := max{FTi I i ~ Pj. } + d f ;
FTj. := min{t l E F T : ~ t < LFTj., kj. • < 7rK•~,
3.1. The serial method r=t-dj. + l . . . . . t, r E R } ;
Sn+ l := S. tO {j*};
The serial method was proposed by Kelley (1963). n : = n + l;
It consists of n = 1. . . . . J stages, in each of which END,
one activity is selected and scheduled. Associated Stop
with each stage are two disjoint activity-sets: In the
scheduled set S, are the activities which were al- Utilising the serial method in a single-pass envi-
ready scheduled and thus belong to the partial sched- ronment, results were published by Pascoe (1966),
ule. The decision set D n contains the unscheduled Mtiller-Merbach (1967), Gonguet (1969), Fehler
activities with every predecessor being in the sched- (1969), Cooper (1976, 1977), i Boctor (1990), and
uled set. In each stage one activity from the decision
set is selected with a priority rule (in case of ties the
t Misleadingly, Cooper terms his scheduling scheme to be
activity with the smallest activity number is selected)
parallel (serial) when using priority rules in a dynamic (static)
and scheduled at its earliest precedence and resource fashion. But, as already pointed out by Valls et al. (1992), he
feasible start time. Afterwards, the selected activity clearly employed a serial scheduling scheme.
R. Kolisch/ European Journal of Operational Research 90 (1996) 320--333 323

Vails et al. (1992). Boctor performed a computa- time. Afterwards, this activity is removed from the
tional study on the basis of 36 small instances from decision set and put into the active set. Step (2) is
the literature. The two best priority rules employed repeated until the decision set is empty, i.e. activities
in his study selected the activity with the smallest were scheduled or are not longer available for
final node number in an activity-on-arrow network. scheduling w.r.t, resource constraints. The parallel
Ties were resolved by scheduling the activity with method terminates when all activities are in the
the smallest initial node and with the maximum total complete or active set.
resource requirement, respectively. Both rules ob- Given A,, the active set, and C,, the complete
tained an average increase of 9.13% above the opti- set, respectively, rrK r, the left over period capacity
mal objective function. of the renewable resource r at the schedule time, and
D,, the decision set, are defined as follows:
3.2. The parallel method
7rKr:=Kr- E kjr,
j~A.
Today, two algorithms are associated with the
so-called parallel method: The algorithm of Kelley D.:= {jI jC~ {C. U A.}, P j C C . ,
(1963) and the one of Brooks (Bedworth and Bailey,
kj~ <_ ¢rK r Vr ~ R}.
1982), which is also termed "Brooks algorithm"
(BAG). Like in the majority of publications, the Now, a formal description of the parallel scheduling
scheduling scheme as proposed by Brooks is em- scheme (PSS) arises to:
ployed herein and referred to as parallel method.
The parallel method consists of at most J stages [PSSI
in each of which a set of activities (which might be Initialisation: n := l, t. := 0, D. := {1}, A. := C.
empty) is scheduled. A unique feature of the parallel := ¢, 7rK, := K~ V r ~ R, GOTO Step (2);
method is that each stage n is associated with a WHILE I A. U C. I < J DO Stage n
schedule time t,, where t~ < t, for m < n holds. On BEGIN
account of this schedule time, the set of scheduled (1) t, := min{FTjl j ~ A,_ l};
activities is now divided into the following two A,:= A,_I \ { j [ j ~ A , _ j . FTj= t.};
subsets: Activities which were scheduled and are C, := C,_1U {jl j ~ A , _ I . F T j = t,};
completed up to the schedule time are in the com- COMPUTE 7rK~ V r ~ R and D,;
plete set C,, while activities which were scheduled, (2) j* := minj~ o{Jl u ( j ) = infi~ o{u(i)}};
but which are at the schedule time still active, are in FTj. := t. + dj. ;
the active set A,. Finally, we have the decision set A . : = A . O{j*};
D,. In contrast to the serial method it contains all yet COMPUTE 7rK r Vr ~ R and D.;
unscheduled activities which are available for IF D. 4=~J THENGOTO Step (2) ELSE n := n + 1;
scheduling w.r.t, precedence and resource con- END;
straints. The partial schedule of each stage is made Stop
up by the activities in the complete set and the active
set. The schedule time of a stage equals the earliest Computational experiments conducted with the
completion time of activities in the active set of the single-pass version of the parallel method are more
ancestral stage. Each stage is made up of two steps: frequent than those with the serial method and are
(1) The new schedule time is determined and activi- reported by Alvarez-Valdes and Tamarit (1989a,b),
ties with a finish time equal to the (new) schedule Boctor (1990), Davis and Patterson (1975), Elsayed
time are removed from the active set and put into the (1982), Lawrence (1985), Pascoe (1966), Patterson
complete set. This, in turn, may place a number of (1973, 1976), Thesen (1976), Ulusoy and ()zdamar
activities into the decision set. (2) One activity from (1989), Valls et al. (1992), and Whitehouse and
the decision set is selected with a priority rule (again, Brown (1979). Additionally, Arora and Sachdeva
in case of ties the activity with the smallest label is (1989) report about an implementation of the parallel
chosen) and scheduled to start at the current schedule method on parallel processors. Davis and Patterson
324 R. Kolisch / European Journal of Operational Research 90 (1996) 320-333

(1975) document an average increase above the opti- a more sensible choice (Baker, 1974, p. 72). In the
mum of 5.6% for the minimum slack priority rule context of the job shop problem, biased random
and 6.7% for the latest finish time priority rule, when sampling is usually referred to as probabilistic dis-
applied to 83 of the instances employed in Patterson patching (Conway et al., 1967, p. 124; Baker, 1974,
(1984). On the basis of their 48 test instances with pp. 202-206). (iii) A special case of biased random
27 activities each, Alvarez-Valdes and Tamarit sampling is the utilisation of regret measures for
(1989a) come up with an average increase above the determining the selection probabilities. This was in-
optimum of 2.89% for the greatest positional rank troduced by Drexl (1991) and Drexl and Grfinewald
priority rule and 3.09% for the latest finish time (1993) and is referred to as regret based biased
priority rule. random sampling. Let a priority rule be defined by
The only comparison of the serial and the parallel the mapping v : j ~ D, ~ R > 0 which assigns to
scheduling scheme when applied as deterministic each activity j in the decision set D n a priority value
single-pass heuristic is reported by Vails et al. (1992). v(j) and an objective 0 stating whether the activity
They concluded that none of the schemes is domi- of the decision set with the minimum ( 0 = min) or
nant which contradicts the assumption made by A1- maximum ( 0 = max) priority value is selected. Then,
varez-Valdes and Tamarit (1989a) that parallel algo- the regret pj compares the priority value of activity j
rithms "seem to work better than the serial ones". with the worst consequence in the decision set as
follows:
3.3. Sampling /maxu(i)-u(j), ifO=min,
~ iED n
The way the serial and the parallel scheduling PJ := [ . ( j) -- miDn . ( i ), ifO=max. (6)
method have been described so far is termed as
single-pass approach, i.e. one single pass and one
Therewith, the parameterised probability mapping
priority rule are employed to derive one feasible
arises to
solution. Multi-pass procedures, on the contrary,
perform Z single passes in order to generate a ( p / + 1)"
sample of at most Z unique feasible solutions, where $(J):= E (pi+l)"" (7)
the best one is chosen. Basically, two different kinds
i~D n
of multi-pass methods can be distinguished: The
multi-priority rule approach (Lawrence, 1985; Boc- Adding the constant " 1 " to the regret value pj
tor, 1990; Li and Willis, 1992) employs one schedul- assures that the selection probability for each activity
ing scheme and different priority rules while sam- in the decision set is greater than zero and thus every
pling (Levy et al., 1962; Wiest 1967; Cooper, 1976; schedule of the population may be generated. By
Alvarez-Valdes and Tamarit, 1989b) makes use of choice of the parameter a, the amount of bias can be
one scheduling scheme and one priority rule. Differ- controlled. Associated with an arbitrary large a will
ent schedules are obtained by biasing the selection of be no bias and thus deterministic activity selection
the priority rule through a random device. The use of on the basis of the employed priority rule (with
a random device can be interpreted as a mapping random selection as a tie breaker) while an a of 0
will give way for random activity selection.
0 : j ~ D.---) [0, 1] (5)
Sampling applications of the serial method are
which at stage n assigns to each activity in the documented by Cooper (1976) while sampling ef-
decision set D. a probability ¢ ( j ) of being selected forts on the basis of the parallel method are reported
(where Ej~ D.~(J) = 1 holds). Three different meth- by Wiest (1967) and Alvarez-Valdes and Tamarit
ods can be distinguished. (i) Random sampling as- (1989b). Employing a sample size of 100, Cooper
signs each activity in the decision set the same (1976) compared deterministic scheduling and biased
probability. (ii) Biased random sampling biases the random sampling with nine different priority rules on
probabilities dependent on the priority values of the one benchmark instance. He concluded that sampling
activities to favour those activities which seem to be produces results which are (at the 99% level of
R. Kolisch / European Journal of Operational Research 90 (1996) 320-333 325

confidence) at least 7% better than the solutions Starting with a feasible schedule S = (FT~ . . . . . FFj,
derived by the deterministic approach. Alvarez- .... FT s) we ceteris paribus assign activity j the ear-
Valdes and Tamarit (1989b) compared a single-pass lier finish time FF~', i.e. we are "left shifting"
and a sampling approach on a set of 48 instances activity j from FTj to FT] with FF~'< FTj. We
with 103 activities each. The sampling approach evaluate the r e s u l t i n g schedules S" =
generated for every instance 100 solutions. For the (FFj . . . . . FT7 . . . . . FTj) and all intermediate sched-
best (second best) priority rule an average increase ules S' = (FF l . . . . . FF~ . . . . . FTj) with FT~' < FF~ <
above an upper bound of 3.23% (3.45%) when used FFj. If the resulting schedule and all intermediate
in the single-pass procedure and 2.31% (1.65%) schedules are feasible we have performed a local left
when used in the sampling procedure is reported. shift of activity j, if the resulting schedule is feasible
To the best of our knowledge, no comparison of and at least one intermediate schedule is infeasible
the serial and the parallel scheduling scheme when we have performed a global left shift of activity j.
applied as sampling procedures is reported in the Now, Sprecher et al. (1994) define active and
literature. non-delay schedules as follows. An active schedule
is defined as a feasible schedule where none of the
3.4. Theoretical results activities can be locally or globally left shifted.
Contrary, a non-delay schedule is defined as a feasi-
With the following two observations we will now ble schedule where none of the sub-activities of the
show that both scheduling methods generate feasible corresponding unit-time-duration schedule (a sched-
schedules which are optimal for the resource-uncon- ule where each activity j is split into dj sub-activi-
strained case. Furthermore, we prove that the parallel ties with duration " 1 " ) can be locally or globally
method generates non-delay-schedules while the se- left shifted. Note that by definition the set of non-de-
rial method constructs active schedules. lay-schedules is a non-proper subset of the set of
active schedules.
Observation 1. For any (feasible) instance of the (a) To prove that any schedule generated by the
RCPSP a feasible schedule is derived by each of the serial scheduling scheme belongs to the set of active
two scheduling schemes, because both methods take schedules we proceed as follows. First we show that
into account precedence and resource constraints. the schedule is (at least) active and then we verify
that it is not a non-delay schedule. Consider activity
Observation 2. For instances which are resource-un- j has been selected at stage n. Then, its time window
constrained, both methods reduce to a simple for- of precedence feasible finish times is restricted from
ward recursion which yields optimal solutions for the the earliest finish time, i.e. the maximum finish time
polynomial solvable problem (1), (2), and (4). of its immediate predecessors plus its duration, to its
latest finish time. Activity j is now scheduled at the
Theorem. (a) A schedule S generated with the earliest contiguous resource feasible interval of length
serial scheduling scheme and any priority rule be- dj within its precedence feasible time window.
longs to the set of active schedules while (b) a Therefore, a left shift of any activity is not possible
schedule S generated with the parallel scheduling and the schedule has to be at least active. Further-
scheme and any priority rule belongs to the set of more, it has to be verified that the schedule is not a
non-delay schedules. non-delay schedule. This can be achieved by show-
ing that in the corresponding unit-time-duration
Proof. In order to show that the Theorem holds, we schedule at least one of the sub-activities can be
need to introduce the notion of local and global left locally or globally left shifted. Again, consider that
shifts for active and non-delay schedules. For a activity j with a duration of dj > 1 has been selected
general classification of schedules for the RCPSP we at stage n and, additionally, that there are two con-
refer to Sprecher et al. (1994). tiguous resource feasible intervals in the precedence
Following Sprecher et al. (1994) we can define feasible time window of activity j. The "earlier"
local and global left shifts along the following line. one with less than dj and the "latter" one with dj
326 R. Kolisch / European Journal of Operational Research 90 (1996) 320-333

units in length. Then, activity j will be scheduled at scheduling scheme searches in a smaller solution
the earliest contiguous resource feasible interval of space than the serial scheduling scheme, but with the
length dj in its precedence feasible time window. severe drawback that, when considering a regular
Hence, within the corresponding unit-time-duration performance measure, the solution space might not
schedule at least the first sub-activity emanating contain the optimal solution.
from activity j can be globally left shifted. Therefore
the schedule is not a non-delay schedule and hence
has to be an active schedule. 4. Experimental investigation
(b) To prove that any schedule generated by the
parallel scheduling scheme belongs to the set of 4.1. Statistical model
non-delay schedules, it must be shown that in the
corresponding unit-time-duration schedule none of In order to study the performance of the schedul-
the sub-activities can be locally or globally left ing schemes, the following statistical model (with
shifted. Assume that in the unit-time-duration sched- five factors) was employed (Kurtulus and Davis,
ule the first sub-activity emanating from activity j 1982; Kurtulus and Narula, 1985):
can be globally or locally left shifted to period t,.
DEVab ....
Hence, at stage n of the parallel method, activity j
has been in the decision set because as a prerequisite = 8(PR a, SS b, Z c, NC,,, RFn, RSo) + 4~ab. . . . .
to left shift the first sub-activity of j all predecessors (8)
of j had to be finished and each resource had to
provide enough left over capacity to process activity CPUab . . . .
j. In addition, stage n has been finished without = 0(PRa, SSb, Zc, NCm, RFn, RSo) +eab . . . . .
scheduling activity j, leaving enough left over capac-
(9)
ity in period t n to accommodate the first sub-activity
of j. Therefore, activity j still has been in the where DEV~b. . . . (CPUabcmno) denotes the average
decision set when stage n had been finished. But this deviation from the optimum solution (the average
is not possible because the algorithm terminates a running time in CPU seconds) when the instances
stage only when the decision set is empty. This with the m-th, n-th, and o-th level of the problem
assures that the first sub-activity of j cannot be left parameters NC, RF, and RS are solved with the
shifted at all which limits the resulting schedule to a-th, b-th, and c-th level of the procedure parame-
be a non-delay schedule. [] ters PR, SS, and Z, respectively. For the errors ~ it
is assumed that they are mutually independent and
At this juncture, it has to be recalled that the set that each e is drawn from the same continuous
of non-delay schedules might not contain a schedule population.
which optimises a regular measure of performance The procedure parameters are characterised as
(Sprecher et al., 1994). In other words: The parallel follows: PR denotes the priority rule, SS stands for

Table 1
Good priority roles presented in the literature
Priority rule Paper O ~,( j )

Most total successors (MTS) Alvarez-Valdes and Tamarit max


Latest start time (LST) Alvarez-Valdes and Tamarit min LF'Tj - dj
Greatest rank positional weight (GRPW) Alvarez-Valdes and Tamarit max dj + ErE s di
Weighted resource utilisation ratio and Ulusoy and 0zdamar max 0.7 I Sj I -}" 0.3 E r E R k j r / g r
precedence (WRUP)
Latest finish time (LFF) Davis and Patterson min
Minimum slack (MSLK) Davis and Patterson min L rj - EFr
R. Kolisch / European Journal t~"Operational Research 90 (1996) 320-333 327

the scheduling scheme while Z denotes the sample Table 2


size. Priority rules were chosen according to the Levels of variable problem parameters
studies of Davis and Patterson (1975), Alvarez- NC m RF,, RS o
Valdes and Tamarit (1989a), Vails et al. (1992), {1.5, 1.8, 2.1} {0.25, 0.5, 0.75, 1} {0.2, 0.5, 0.7}
Uluzoy and tDzdamar (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 a positive resource demand, i.e. kjr > 0, the latter
of these studies. Note that only the studies by Boctor was drawn from the interval [1, 10]. The number of
(1990) and Vails et al. (1992) employed the (parallel different resource types requested by one activity
and the) serial scheduling scheme and that none of was - depending on the resource factor RF - in the
the three best rules were applied within the serial range [1, 4]. Finally, the precedence network was
scheduling scheme. Additional to the notation al- generated with the following constraints: The num-
ready introduced, Sj(Sj.) denotes the set of all (im- ber of immediate successors (predecessors) of the
mediate) successors of activity j and EFT~ denotes dummy-source (dummy-sink) was set to 3, respec-
the earliest precedence and resource feasible finish tively, and the number of successors (predecessors)
time of activity j. Note that in the parallel scheduling of each non-dummy activity was drawn from the
scheme EFT~ equals t n + d/ for each activity in the interval [1, 3].
decision set. Since within step (2) of the parallel In order to generate instances with these problems
scheduling scheme t, is constant for all j ~ Dn, v ( j ) parameters, ProGen - an instance generator for a
is equal for MSLK and LST which was proven by broad class of precedence and resource-constrained
Davis and Patterson (1975). Further note that LST (project) scheduling problems (Kolisch et al., 1995)
and MSLK are listed separately because the priority - was employed as follows: As shown in Table 2, 10
rules are employed in both schemes, instances for each combination of NC, RF, and RS
Two levels of the scheduling schemes, i.e. the were generated which equalled a total of 3 • 4 . 3 • 10
parallel scheduling scheme (PSS) and the serial -- 360 problems. Imposing a time limit of 3600 CPU
scheduling scheme (SSS), were considered. The lev- seconds, for 308 of these instances the optimal solu-
els of the sample size are provided in the subsequent tion was obtained with the exact procedure of De-
sections. meulemeester and Herroelen (1992) on a personal
The problem parameters are characterised as fol- computer with 80386sx processor, mathematical co-
lows (for details cf. Kolisch et al., 1995): The net- processor and 15 MHz clockpulse. Hence, each of
work complexity NC is the ratio of non-redundant these 308 problems was treated by every level com-
precedence relations to the number of activities. The bination of the procedure parameters, scheduling
resource factor RF reflects the density of the two scheme, priority rule, and sample size.
dimensional array k~r, j = 2 . . . . . J - 1 and r = Since it is not confirmed that DEV is normally
1 . . . . . I RI. Finally, the resource strength RS mea- distributed, only nonparametric tests, namely the
sures the degree of resource-constrainedness in the Wilcoxon signed rank test and the Friedman test,
interval [0, 1]. The resource strength is computed as were employed on a one-way layout without replica-
follows: RS = ( K r - K , . rain ) / ( K rmax - K , . min ), where tions (Alvarez-Valdes and Tamarit, 1989a; Golden
K min is the minimal availability of resource type r and Steward, 1985). Confidence levels are denoted
in order to assure feasibility of the RCPSP, i.e. with s. Observations with confidence levels of less
K mi" = max{kit I j = 1. . . . . J}, and K rmax is the peak than or equal to 1%, i.e. s < 0.01, will be judged as
demand of resource type r in a CPM schedule. significant. The one-way layout was derived by sim-
All other problem parameters were adjusted as ply averaging over all factors except the one under
follows (where intervals consist of uniformly dis- consideration (Kurtulus and Davis, 1982).
tributed integers): The number of non-dummy activi- All algorithms were coded in PASCAL and im-
ties was set to 30, i.e. J = 32, the number of re- plemented on an IBM compatible personal computer
source types was set to 4, i.e. I RI = 4, the activity with 80386dx processor and 40 MHz clockpulse at
duration was drawn from the interval [1, 10]. In case the laboratory of the Christian-Albrechts-Universif~it
328 R. Kolisch / European Journal of Operational Research 90 (1996) 320-333

Table 3 DEV
Performance of priority rules
PR~ LST LFT MTS MSLK GRPW WRUP
11
DEVa 5.06 5.32 6.65 7.53 10.78 11.66
CPU a 0.02 0.02 0.03 0.03 0.02 0.03 10

zu Kiel. Random numbers were drawn with the


generator proposed by Schrage (1979).
I-- 2 3 4 m,n,o
4.2. S i n g l e - p a s s a n a l y s i s
Fig. 1. Effect of the problem parameters on the overall perfor-
mance.
For the single-pass analysis the sample size Z c
was set to " 1 " and the selection of activities was
performed deterministically. Table 3 gives a compar- 1989a). Each priority rule performs better w.r.t, the
ison of the priority rules. The Friedman test reveals a quality of solutions when applied in the parallel
significant different performance w.r.t, the average scheme. Even more, every priority rule with the
deviation from the optimal solution (s = 0.0000). By exception of the lower bound based rules LST and
pairwise application of the Wilcoxon test the follow- LFT is significantly better within the parallel
ing ranking is observed (where " >- " denotes better scheduling scheme (s < 0.0013). Nevertheless, since
and " >->- " denotes significant better): LST >- LFT LST and LF'I" belong to the best rules in both
>->- MTS ~- MSLK >->- GRPW >->- WRUP. Thus, schemes, the serial scheduling scheme cannot be
four groups can be distinguished significantly (s < excluded a priori. This conclusion was already drawn
0.0051): The (lower bound based) rules LST and in the study by Vails et al. (1992).
LFT which perform quite good, MTS and MSLK The running time of the parallel scheduling
ranging in the middle, as well as GRPW and finally scheme is slightly less than the one of the serial
WRUP, the two latter revealing a poor performance, scheme. Whereas the parallel scheme uses most of
respectively. The computational effort for all rules is the time to update the decision set, the serial scheme
very modest. Within both scheduling schemes, the requires the majority of the CPU time for setting up
rules MTS and WRUP require slightly more CPU and managing the array 7rKrt , r ~ R , t = 1 . . . . . T ,
time, whereas the MSLK rule demands only within which is (especially in the case of a poor upper
the serial scheduling scheme more CPU time. This is bound for the makespan T) very time consuming.
because the earliest precedence and resource feasible The effect of the problem parameters on the
finish time has to be determined for every activity in overall performance is as follows (cf. Fig. 1 with
the decision set. monotonically increasing parameter levels on the
Table 4 demonstrates the performance of the x-axis): The network complexity NC does not reveal
scheduling schemes. The ranking reveals to be PSS a significant influence (s = 0.2929), whereas the ef-
>->- SSS and hence confirms the conjecture that the fect of the resource factor RF and resource strength
parallel scheduling scheme is significantly (s = RS, respectively, is highly significant (s = 0.0000).
0.0000) superior to the serial one when used as Based on the benchmark-instances of Kolisch et al.
single-pass heuristic (Alvarez-Valdes and Tamarit, (1992), we can conclude that the employed problem
parameters influence single-pass scheduling schemes
in the same manner as optimal branch-and-bound-
Table 4
based procedures. That is, a high resource factor and
Performance of scheduling schemes
a low resource strength will generally induce a poor
SS b PSS SSS
performance.
DEVb 6.46 9.21 A bell-shaped performance w.r.t, the resource pa-
CPU b 0.02 0.03
rameters as originally proposed in Elmaghraby and
R. Kolisch / European Journal of Operational Research 90 (1996) 320-333 329

Herroelen (1980) and later experimentally detected form best, MTS and MSLK range in the middle
in the elaborate study of De Reyck and Herroelen while WRUP and GRPW have the worst perfor-
(1993) was not perceived. De Reyck and Herroelen mance. The results of MSLK have to be interpreted
too generated instances with ProGen. But for the with care because - like for the single-pass approach
analyse of the results they employed two different - it shows a quite different performance within each
parameters, namely the complexity index (CI) and of the two scheduling schemes, respectively: For the
the resource constraidness (RC). The CI was intro- parallel scheme it performs like the " g o o d " LST
duced by Bein et al. (1992). Essentially, it measures rule, within the serial scheme MSLK gives rather
how nearly series-parallel an activity-on-arc network poor results. For sampling the computational require-
is. The RC was proposed by Patterson (1976) as ment of all priority rules increases linearly to an
follows: RC r = Dr/K r where D r denotes the aver- average of 1.4 CPU seconds. This is very moderate
age demand of resource type r, i.e. D r = E k j J E { 1 , compared to the average time of 45.24 CPU seconds
if k jr ~> 0; 0 else}. De Reyck and Herroelen found needed by the Demeulemeester and Herroelen algo-
out that keeping CI constant the computation time of rithm in order to solve all problems to optimality.
the exact procedure of Demeulemeester and Herroe- But using the latter procedure as truncated branch-
len (1992) reveals a bell-shaped performance for and-bound method with a maximum of 1 CPU sec-
ascending RC values. Hence, it is conjectured that ond, it achieves a deviation of 1.21% and thus
the absence of the bell-shaped curve in this study is outperforms each of the priority rules. Hence, when
caused by not keeping CI constant a n d / o r not sys- employing sampling in a competitive and efficient
tematically varying RC. manner, one has to make use of bounding rules in
Neither the ranking of priority rules nor schedul- order to cut down the computational effort (Cooper,
ing schemes is significantly influenced by the prob- 1974; Kolisch, 1995).
lem parameters. Regarding priority rules, this out- By comparing the ranking (LFT >- LST >- >- MTS
come confirms the conclusion made in the studies by >- MSLK >- >- WRUP >- >- GRPW) with the one of
Cooper (1976) and Alvarez-Valdes and Tamarit the single-pass approach the following can be stated:
(1989a). Furthermore, the computational effort is not (i) With two exceptions (LST and LFT as well as
influenced by any of the problem parameters. WRUP and GRPW) the ranking obtained is the same
as for the single-pass case. (ii) While the difference
4,3. Sampling analysis between groups (i.e. LFF, LST vs. MTS, MSLK vs.
WRUP vs. GRPW) has about the same level of
In the following it is investigated if the conclu- significance (s < 0.0036), the difference between
sions which were drawn for the single-pass case are rules is slightly less significant. Thus, it can be stated
still valid in the case of sampling. According to that, in general, priority rules which are good for
preliminary computational results (Kolisch, 1995), single-pass approaches are good for biased random
the levels of the sample size Z c were chosen to be sampling approaches and vice versa. Although not
{10, 40, 70, 100} and the bias parameter a was ex- explicitly pointed out, Alvarez-Valdes and Tamarit
clusively set to " l " . (1989b) attained similar results in their study. Their
Table 5 reveals the performance of the priority rankings obtained when utilising six priority rules
rules. As for the single-pass case, a significant differ- within the deterministic- and the sampling-based par-
ence between rules can be detected (s = 0.0024). allel scheduling scheme differed only w.r.t, one rule.
The (lower bound based) rules LST and LFT per- The opposite observation by Cooper (1976) is a
consequence of his (not regret based) probability
mapping which in conjunction with the LFT priority
Table 5
Performance of traditional priority rules for sampling
rule tends to perform (pure) random sampling. Of
course, the interrelation between single-pass and
PR ,~ LFT LST MTS MSLK WRUP GRPW
sampling heuristics depends on the amount of bias:
DEVa 2.08 2.11 2.48 2.56 3.36 3.79
Whereas with 0% bias the significance equals the
CPU a 1.12 1.11 1.68 1.63 1.17 1.65
one of the (deterministic) single-pass case, a bias of
330 R. Kolisch / European Journal of Operational Research 90 (1996) 320-333

DEVc D E Vc

7 8 SS

6 7
6 PS
5
5
4 4
3 3
2 -z c 2 , . ,~ ._.7c
I0 40 70 100 10 40 70 100
Fig. 2. Performance as a function of the sample size. Fig. 3. Impact of the sample size on the performance of the
scheduling schemes.
100% (i.e. random sampling) results in no (signifi-
cant) difference between the priority rules. ules, as announced in Conway et al. (1967, pp.
Fig. 2 gives insight into the performance w.r.t, the 121-124) for the job shop problem when minimising
sample size. As expected, it is demonstrated that the average flow time, does not hold true. The
increasing the sample size continuously produces rationale of the observation is as follows: For small
better solutions. Depending on the sample size, the sample sizes the superiority of the parallel scheme
average performance of the single-pass approach is w.r.t, the single-pass approach is dominant. With
thus improved between 50% (Z c = 10) and 73% increasing sample size, this effect diminishes and at
(Z c = 100). This is up to ten times more than ob- the same time the parallel scheme suffers from the
served by Cooper (1976). Hence, it can be stated that fact that the sample space is the set of non-delay
sampling significantly outperforms the single-pass schedules which not necessarily contains the optimal
approach (s = 0.0000). This contradicts the conclu- solution. We conjecture that this "critical sample
sions drawn by Conway et al. (1967, p. 128) for the size", i.e. 40 for the instances tested, increases when
job shop problem, stating that sampling reveals only the problem size, expressed in the number of activi-
modest improvement over single-pass procedures. ties, is enlarged.
But it has to be noted that the marginal improvement In order to investigate how much the parallel
diminishes. This implies a growing computational scheduling scheme is restricted by operating on the
effort in order to produce better solutions. set of non-delay schedules, a full enumeration of all
Finally, Fig. 3 demonstrates the effect of the schedules in this set was performed as follows: In
scheduling schemes w.r.t, the sample size Zc. The step (2) of PSS a branch was created for each
overall performance of both schemes is with 2.75% non-dominated set of activities out of D n which
for the parallel and 2.71% for the serial scheduling could be jointly started at t n. Imposing a time limit
scheme almost identical. But a second glance reveals of 24 CPU hours, the best non-delay schedule was
a (not significant ( s > 0 . 1 9 9 4 ) ) different perfor- obtained for 298 of the 308 instances. Table 6 shows
mance w.r.t, the sample size. While the parallel the frequency distribution of the deviation for the
scheme is clearly superior f o r small sample sizes best non-delay schedules. For 123 problems, i.e.
(i.e. less than 40 generated schedules), the serial 41.27%, the set of non-delay schedules did not con-
scheme shows better results for large samples. Con- tain the optimal solution.
sequently, for sampling procedures solving the The effect of the three problem parameters net-
RCPSP the general superiority of non-delay sched- work complexity, resource factor, and resource

Table 6
Frequency distribution of the deviation for the best non-delay schedules
DEV range 0 (0, 1] (1, 2] (2, 3] (3, 4] (4, 5] (5, 10] > 10
Number of instances 175 1 33 25 19 14 26 5
R. Kolisch / European Journal of Operational Research 90 (1996) 320-333 331

Table 7 when considering a regular performance measure -


Effect o f R F and RS on the r a n k i n g o f scheduling s c h e m e s
the solution space might not contain an optimal
RS I = 0.2 RS 2 = 0.5 RS 3 = 0.7 schedule.
R F t = 0.25 SS SS SS * SS * An in-depth computational study with the instance-set
R F 2 = 0.5 PS SS SS * SS of Kolisch et al. (1995) produced the following
R F 3 = 0.75 PS PS * PS PS *
results when using both methods as a frame for
RF4=1 PS * PS * PS PS *
single-pass scheduling and sampling: Priority rules
PS PS * SS * which derive good results for single-pass scheduling
do so for sampling. When performed properly, i.e.
when pure random sampling is avoided, sampling
improves the results of single-pass scheduling up to
70%. Deviating from the general belief, we showed
strength on the average performance turned out to be that the parallel method does not generally perform
as for the single-pass case. That is, no significant better than the serial method. Rather, it gives only
influence can be observed for the network complex- better results for single-pass scheduling and small
ity (s = 0.6485) while resource strength and resource sample sizes as well as for " h a r d " (that is highly
factor, in the order mentioned, turned out to be resource-constrained) problems. Hence, the serial
highly significant (s = 0.0000). The ranking of the method is superior for large sample sizes and for
priority rules is not significantly effected by the instances which are only moderately resource-con-
problem parameters but, deviating from the single- strained. This insight should be of importance when
pass case, resource factor and resource strength in- deriving fast problem-specific parameter-guided
fluence the ranking of scheduling schemes signifi- heuristics.
cantly.
Table 7 shows the ranking of the scheduling
schemes w.r.t, both resource parameters as well as
their combination (where * denotes a 1-tailed signif- Acknowledgement
icance at the 1% level of confidence and the bold
entries signal superiority of the serial method). The author is indebted to Erik Demeulemeester,
Roughly, it can be stated that the parallel scheduling Katholieke Universiteit Leuven, for providing the
scheme performs better for "hard" problems (with code of his algorithm, to Andreas Drexl, Universit~it
a high resource factor a n d / o r a low resource Kiel, for his help in this research, and to two anony-
strength) while the serial scheduling scheme is better mous referees for their helpful comments and sug-
for "easy" problems (with a low resource factor gestions on an earlier version of this paper.
a n d / o r a high resource strength).

References
5. S u m m a r y
Alvarez-Valdes, R., a n d Tamarit, J.M. (1989a), " H e u r i s t i c algo-
We reviewed two scheduling schemes, the serial rithms for resource-constrained project scheduling: A review
a n d an empirical a n a l y s i s " , in: R. Slowinski a n d J. W e g l a r z
and the parallel method, by giving a formal descrip-
(Eds.), Advances in Project Scheduling, Elsevier, A m s t e r d a m ,
tion and an extensive literature review. Both methods 113-134.
generate feasible schedules which are optimal in the Alvarez-Valdes, R., a n d Tamarit, J.M. (1989b), " A l g o r i t m o s
absence of resource restrictions. It was proven that heuristicos deterministas y aleatorios en secuenciacion de
the serial method generates active schedules while proyectos c o n recursos l i m i t a d o s " , Q~estiio 13, 1 7 3 - 1 9 1 .
Arora, R.K., and S a c h d e v a , (1989), " D i s t r i b u t e d simulation o f
the parallel scheduling scheme creates non-delay
resource constrained project s c h e d u l i n g " , Computers & Oper-
schedules. Hence, the parallel scheduling scheme ations Research 16, 2 9 5 - 3 0 4 .
searches in a smaller solution space than the serial Baker, K.R. (1974), Introduction to Sequencing and Scheduling,
scheduling scheme but with the drawback that - Wiley, N e w York.
332 R. Kolisch / European Journal of Operational Research 90 (1996) 320-333

Balas, E. (1971), "Project scheduling with resource constraints", project scheduling problem", Management Science 38, 1803-
in: E.M.L. Beale (Ed.), Applications of Mathematical Pro- 1818.
gramming techniques, English University Press, London, 187- Drexl, A. (1991), "'Scheduling of project networks by job assign-
200. ment", Management Science 37, 1590-1602.
Bedworth, D.D., and Bailey, J.E. (1982), Integrated Production Drexl, A., and Grilnewald J. (1993), "Nonpreemptive multi-mode
Control Systems - Management, Analysis, Design, Wiley, resource-constrained project scheduling", liE Transactions
New York. 25/5, 74-81.
Bein, W.W., Kamburowski, J. and Stallmann, M.F.M. (1992), Elmaghraby, S.E. (1977), Activity Networl~s': Project Planning
"Optimal reduction of two-terminal directed acyclic graphs", and Control by Network Models, Wiley, New York.
SlAM Journal on Computing 21, 1112-1129. Elmaghraby, S.E., and Herroelen, W.S. (1980), "'On the measure-
Bell, C.E., and Han, S. (1991), " A new heuristic solution method ment of complexity in activity networks", European Journal
in resource-constrained project scheduling", Naval Research of Operational Research 5, 223-234.
Logistics 38, 315-331. Elsayed, E.A. (1982), "Algorithms for project scheduling with
Bell, C.E., and Park, K. (1990), "Solving resource-constrained resource constraints", International Journal of Production
project scheduling problems by A* search", Naval Research Research 20, 95-103.
Logistics 37, 61-84. Fehler, D.W. (1969), "Die Variationen-Enumeration - Ein
Blazewicz, J., Lenstra, J.K., and Rinnooy Kan, A.H.G. (1983), NSherungsverfahren zur Planung des optimalen Betriebsmit-
"Scheduling subject to resource constraints: Classification and teleinsatzes bei der Terminierung von Projekten", Elektroni-
complexity", Discrete Applied Mathematics 5, 11-24. sche Datenverarbeitung 10, 479-483.
Boctor, F.F. (1990), "Some efficient multi-heuristic procedures Golden, B.L., and Steward, W.R. (1985), "Empirical analysis of
for resource-constrained project scheduling", European Jour- heuristics", in: E.L. Lawler, J.K. Lenstra, A.H.G. Rinnooy
nal of Operational Research 49, 3-13. Karl and D.B. Shmoys (Eds.), The Traveling Salesman Prob-
Bowman, E.H. (1959), "The schedule-sequencing problem", Op- lem, Wiley, New York, 207-249.
erations Research 7, 621-624. Gonguet, L. (1969), "Comparison of three heuristic procedures
Carlier, J., and Latapie, B. (1991), "Une methode arborescente for allocating resources and producing schedules", in: H.J.M.
pour resoudre les problemes cumulatifs", Recherche opera- Lombaers (Ed.), Project Planning by Network Analysis,
tionnelle 25, 311-340. North-Holland, Amsterdam, 249-255.
Carruthers, J.A., and Battersby, A. (1966), "Advances in critical Hastings, N.A.J. (1972), "On resource allocation in project net-
path methods", Operational Research Quarterly 17, 359-380. works", Operational Research Quarterly 23, 217-221.
Christofides, N., Alvarez-Valdes, R., and Tamarit, J.M. (1987), Kelley, J.E., Jr. (1963), "The critical-path method: Resources
"Project scheduling with resource constraints: A branch and planning and scheduling", in: J.F. Muth and G.L. Thompson
bound approach", European Journal of Operational Research (Eds.), Industrial Scheduling, Prentice-Hall, New Jersey, pp.
29, 262-273. 347-365.
Conway, R.W., Maxwell, W.L., and Miller, L.W. (1967), Theory Kolisch, R. (1995), Project Scheduling under Resource Con-
of Scheduling, Addison-Wesley, Reading, MA. straints - Efficient Heuristics for Several Problem Classes,
Cooper, D.F. (1976), "Heuristics for scheduling resource-con- Physica, Heidelberg.
strained projects: An experimental investigation", Manage- Kolisch, R., Sprecher, A., and Drexl, A. 0995), "Characteriza-
ment Science 22, I 186- l 194. tion and generation of a general class of resource-constrained
Cooper, D.F. (1977) " A note on serial and parallel heuristics for project scheduling problems", Management Science 41,
resource-constrained project scheduling", Foundations of 1693-1703.
Control Engineering 2, 131-134. Kurtulus, I.S., and Davis E.W. (1982), "Multi-project scheduling:
Davis, E.W. (1966), "Resource allocation in project network Categorization of heuristic rules performance", Management
models - A survey", The Journal of Industrial Engineering Science 28, 161-172.
17, 177-188. Kurtulus, I.S., and Narula, S.C. (1985), "Multi-project schedul-
Davis, E.W., and Heidorn, G.E. (1971), "An algorithm for opti- ing: Analysis of project performance", liE Transactions 17,
mal project scheduling under multiple resource constraints", 58-66.
Management Science 17, 803-816. Lawrence, S.R. (1985), "Resource-constrained project scheduling
Davis, E.W., and Patterson, J.H. (1975), " A comparison of - A computational comparison of heuristic scheduling tech-
heuristic and optimum solutions in resource-constrained pro- niques", Working Paper, Graduate School of Industrial Ad-
ject scheduling", Management Science 21,944-955. ministration, Caruegie-Mellon University, Pittsburgh, PA.
De Reyck, B., and Herroelen W.S. (1996), "On the use of the Leon, V.J., and Balakrishnan, R. (1995), "Strength and adaptabil-
complexity index as a measure of complexity in activity ity of problem-space based neighborhoods for resource con-
networks If", European Journal of Operational Research, to strained scheduling", OR Spektrum 17, 173-182.
appear. Levy, F.K., Thompson, G.L., and Wiest, J.D. (1962), "Multiship,
Demeulemeester, E., and Herroelen, W.S. (1992), "'A branch- multishop, workload-smoothing program", Naval Research
and-bound procedure for the multiple resource-constrained Logistics Quarterly 9, 37-44.
R. Kolisch / European Journal of Operational Research 90 (1996) 320-333 333

Li, R.K.-Y., and Willis, J. (1992), "'An iterative scheduling Schrage, L. (1979), "A more portable Fortran random number
technique for resource-constrained project scheduling", Euro- generator", ACM Transactions on Mathematical Software 5,
pean Journal of Operational Research 56, 370-379. 132-138.
Miiller-Merbach, H. (1967), "Ein Verfahren zur Planung des Shaffer, L.R., Ritter, J.B., and Meyer, W.L. (1965), The Critical-
optimalen Betriebsmitteleinsatzes bei der Terminierung von Path Method, McGraw-Hill, New York.
Grofiprojekten", Zeitschrift fiJr wirtschaftliche Fertigung 62, Sprecher, A., Kolisch, R., and Drexl, A. (1995), -Semi-active,
83-88, 135-140. active, and non-delay schedules for the resource-constrained
Oguz, O., and Bala, H. (1994), "'A comparative study of computa- project scheduling problem", European Journal of Opera-
tional procedures for the resource constrained project schedul- tional Research 80, 94-102.
ing problem", European Journal of Operational Research 72, Stinson, J.P., Davis, E.W., and Khumawala, B.M. (1978), "' Multi-
406-416. ple resource-constrained scheduling using branch and bound",
Pascoe, T.L. (1966), "Allocation of resources C.P.M.", Revue AIIE Transactions 10, 252-259.
Francaise Recherche Operationelle 38, 31-38. Storer, R.H., Wu, S.D., and Vaccari R. (1992), "'New search
Patterson, J.H. (1973), "Alternate methods of project scheduling spaces for sequencing problems with application to job shop
with limited resources", Naval Research Logistics Quarterly scheduling", Management Science 38, 1495-1509.
20, 767-784. Talbot, B., and Patterson, J.H. (1978), "An efficient integer
Patterson, J.H. (1976), "'Project scheduling: The effects of prob- programming algorithm with network cuts for solving re-
lem structure on heuristic performance", Naval Research source-constrained scheduling problems", Management Sci-
Logistics Quarterly 23, 95-123. ence 24, 1163-1174.
Patterson, J.H. (1984), "A comparison of exact approaches for Thesen, A. (1976), "Heuristic scheduling of activities under
solving the multiple constrained resource, project scheduling resource and precedence restrictions", Management Science
problem", Management Science 30, 854-867. 23, 412-422.
Patterson, J.H., and Huber, W.D. (1974), "A horizon-varying, Ulusoy, G., and t~zdamar, L. (1989), "Heuristic performance and
zero-one approach to project scheduling", Management Sci- network/resource characteristics in resource-constrained pro-
ence 20, 990-998. ject scheduling", Journal of the Operational Research Society
Patterson, J.H., and Roth, G.W. (1976), "Scheduling a project 40, 1145-1152.
under multiple resource constraints: A zero-one programming Vails, V., Perez, M.A., and Quintanilla, M.S. (1992), "'Heuristic
approach", AIIE Transactions 8, 449-455. performance in large resource-constrained projects", Working
Pritsker, A.A.B., Watters, L.J., and Wolfe, P.M. (1969), "Multi- Paper, Departament D'Estadistica I investigacio Operativa,
project scheduling with limited resources: A zero-one pro- Universitat De Valencia, Spain.
gramming approach", Management Science 16, 93-107. Whitehouse, G.E., and Brown, J.R. (1979), "Genres: An exten-
Radermacher, F.J. (1985/86), "Scheduling of project networks", sion of Brooks algorithm for project scheduling with resource
Annals of Operations Research 4, 227-252. constraints", Computers" & Industrial Engineering 3,261-268.
Sampson, S.E., and Weiss, E.N. (1993), "Local search techniques Wiest, J.D. (1967), "A heuristic model for scheduling large
for the generalized resource constrained project scheduling projects with limited resources", Management Science 13,
problem", Naval Research Logistics 40, 365-375. B359-B377.

You might also like