Beruflich Dokumente
Kultur Dokumente
and resource requirement. When the resource is money and time objective. Talbot (1982), Sprecher et al (1997) and
it is non-renewable, the problem is referred to as the time/ Demeulemeester et al (2000) propose branch and bound
cost trade-off problem. In time/cost trade-off problems, the algorithms to find exact solutions. There are many
trade-off is between the project completion time and total approximation studies, some noteworthy of which are due
cost. When the trade-off is between the project completion to Alcaraz et al (2003), Ranjbar and Kianfar (2007), Mori
time and renewable resource usage, the problem is referred and Tseng (1997), Van Peteghem and Vanhoucke (2009),
to as the time/resource trade-off problem. Jozefowska et al (2001) and Ranjbar et al (2009).
In this study we consider the RIP with time/resource Mohring (1984) shows that the single mode RIP with
trade-offs. We assume a single renewable resource and a a single renewable resource is strongly NP-hard. Drexl and
given project deadline. The objective is to minimize the Kimms (2001) develop two lower bounds and heuristic
maximum resource usage over all periods. When all units procedures that are byproducts of these lower bounds.
of the resource are made available at the beginning of the Demeulemeester (1995) proposes a branch and bound
project and are freed altogether when the project is com- algorithm for the non-decreasing resource cost function.
plete, the maximum usage defines the amount of resources Gather et al (2008) propose a spanning tree-based enu-
consumed by the project, hence, the total cost of buying meration method for finding exact solutions. Neumann
resources. The number of workers hired for a construction and Zimmermann (1999), Yamashita et al (2007), Shadrokh
project may be defined by the period that gives the maxi- and Kianfar (2007), and Ranjbar et al (2008) propose
mum worker requirements. In such a case the total labour approximation algorithms. Shahsavar et al (2010) consider
investment is an increasing function of the number of the the RIP with discounted cash flows and propose a genetic
workers. The number of available workers allocated to a algorithm.
project may be reduced by selecting an appropriate mode To the best of our knowledge there are only two studies
for each activity. Our main experiments take the project on the RIP with time/resource trade-offs. Yamashita and
deadline as the minimum project completion time that can Morabito (2009) propose an exact procedure to generate
be achieved by consuming the highest resource level for each all efficient schedules with respect to the resource avail-
activity. By offering such a tight deadline, we emphasize ability cost and project completion time criteria. Their
time-based competition. We also look for the effect of the computational results reveal that their procedure is capable
higher deadlines on the performance of our procedures. of solving instances with up to 15 activities. Hsu and Kim
Despite its practical importance, the research on the RIP (2005) propose a priority rule-based heuristic procedure.
with time/resource trade-offs is quite scarce. To the best of They calculate a priority value for each unscheduled
our knowledge, there is one study by Hsu and Kim (2005) activity, its mode, and starting time and select the decision
that offers a heuristic solution and one study by Yamashita leading to the smallest priority value. Their function con-
and Morabito (2009) that offers a trade-off curve for the siders the impact of the decision on the unscheduled activi-
resource availability cost and project completion time ties and resource usages. They show that their heuristic
criteria. Recognizing this gap in the literature, we develop outperforms some other priority rule-based heuristics.
powerful problem size reduction mechanisms, lower
bounds and a heuristic procedure.
The rest of the paper is organized as follows: In the next Problem definition and the model
section, we review the related literature. Next we define our
problem, give its mathematical model and discuss the mode We consider a project with a single renewable resource and
elimination properties. In the succeeding sections, we N non-preemptive activities, that is, once an activity starts
discuss our lower bounds and heuristic procedure and it should be processed until it is completed. A dummy
report on the computational experiment. Then, we present activity 0 stands for the starting activity and dummy
the extensions of our results to the multi-resource envi- activity N þ 1 stands for the ending activity. Activity i has
ronments. Finally, we give a brief conclusion of our study. mi modes. Mode j of activity i is characterized by the
following parameters.
The precedence network is defined by the immediate unique optimal solution. Properties 3 and 4 are used to
predecessor sets. Ei is the set of immediate predecessors of eliminate the modes and start times that cannot lead to
activity i. unique optimal solutions.
We let T denote the deadline of the project. ESi and LSi
are the earliest and latest possible start time of activity i, Property 1 If LCiESiopij, then modes j, j þ 1, . . . , mi of
respectively. ECi and LCi are the earliest and latest activity i cannot lead to a feasible solution.
completion time of activity i, respectively. The ESi, LSi,
ECi and LCi values are found by the CPM using the Proof LCiESi is the maximum duration allowed for
shortest duration modes and specified deadline T. activity i to maintain feasibility. Hence, a solution in which
Our main decision variable xijt gives the start time and activity i is assigned mode j such that pij4LCiESi can
selected mode of each activity and is defined as never lead to a feasible solution. &
1; if activity i starts at time t and is executed with mode j Activity i is said to be critical if its earliest start time is
xijt :
0; otherwise equal to its latest start time, that is, ESi ¼ LSi, otherwise it
8i; j ¼ 1; . . . ; mi ; t 2 ½ESi ; LSi is non-critical. Critical activities exist only when the
deadline is set to the earliest possible completion time.
The decision variable R is defined as the maximum Property 2 defines the start times and modes of the critical
amount of resource used. activities.
The objective is to minimize the maximum resource
usage and is expressed as: Property 2 In all feasible solutions, a critical activity i is
processed between ESi and ECi at its shortest duration
Minimize R ð1Þ mode.
Our constraints are explained next.
The proof of Property 2 is omitted, as it directly follows
that of Property 1.
XX
LSi
xijt ¼ 1 8i ð2Þ We use the following additional notation to state
j2mi t¼ESi Properties 3 and 4.
N X
X X
t LB ¼ A lower bound on the optimal objective function
rij xijk pR 8t ð3Þ value
i¼1 j2mi k¼tpij þ1 LtE ¼ Amount of resource used at period t when all
activities are at their shortest duration modes and
xNþ1;1;T ¼ 1 ð4Þ
start at their earliest start times
XX
LSi X X
LSk LtL ¼ Amount of resource used at period t when all
t xijt X ðt þ pkj Þ xkjt activities are at their shortest duration modes and
j2mi t¼ESi j2mk t¼ESk start at their latest start times
i ¼ 1; 2; . . . ; N þ 1; 8k 2 Ei ð5Þ
Property 3 There exists an optimal schedule in which
each activity i satisfying LtEoLB for every tA[0, ECi] starts
xijt ¼ f0; 1g 8i; j ¼ 1; . . . ; mi t 2 ½ESi ; LSi ð6Þ at ESi and is processed at its shortest duration mode.
Constraint set (2) ensures that each activity is assigned to Proof Assume an optimal schedule A1 that does not
exactly one mode and one start time between its earliest satisfy the condition of the property. That is in A1, there
and latest start times. The resource usage at any time exists an activity i with LtEoLB at every tA[0, ECi] that
cannot exceed the available number of resources as stated either starts later than ECi or assigned to a longer mode.
by constraint set (3). Constraint set (4) is the deadline Shift activity i, all its predecessors and all activities with
constraint, that is, the dummy activity starts at time T. ECjpECi and their predecessors to their earliest start times
According to constraint set (5), an activity can start only and assign them to their shortest duration modes, and get
after all of its predecessors are completed. schedule A2. Note that A2 obeys the conditions of the
property. With this shift and mode assignments, the
resource usages in some periods tA[0, ECi] may increase,
Properties of the optimal solution
however they cannot exceed LB as LtEoLB holds. It
In this section, we develop some properties that help to follows that the maximum resource usage of A2 is equal to
reduce the problem size. Properties 1, 2, 5 and 6 are used to that of the optimal schedule A1. Hence, the schedule A2 that
eliminate the modes that cannot lead to a feasible or obeys the conditions of the property is also optimal. &
780 Journal of the Operational Research Society Vol. 65, No. 5
Property 4 There exists an optimal schedule in which to its lower resource consuming modes will never improve
each activity i satisfying LtLoLB for every tA[LSi, T ] the objective function value, which is certainly not smaller
starts at LSi and is processed at its shortest duration mode. than LB. &
P
Property 6 If rik þ j2FSi rj1 pLB then there exists an
Lower bound 2
optimal schedule in which activity i is assigned to modes
1, . . . , k. If the earliest completion time of an activity is later than its
latest start time, then the activity has to be processed
Proof Note that FSi contains every activity thatP can be between its latest start and earliest completion times.
processed together with activity i. Hence, rik þ j2FSi rjm1 Formally, for activity i if ECiLSiX0 then it is certainly
is an upper bound on the resource usage for the time processed in interval [LSi, ECi].
periods at which activity i is processed at mode k. If this Recall that SSt is the set of activities that have to be
upper bound is not bigger than LB, then setting activity i processed at time t. Activity i is in SSt if ECiLSiX0. The
Erdem Colak and Meral Azizoglu—A RIP with time/resource trade-offs 781
minimum resource usage in SSt is found by setting all minimization problem. To strengthen the relaxation we
activities to their minimum resource consuming modes. develop some valid cuts. Our cuts explain the relations that
The maximum of the minimum resource usages over all are satisfied by the original model however, not by its LPR.
periods gives our second lower bound. Formally, The first cut is an extension of Property 1 for two
activities i and k where kAEi, at respective modes a and b.
( )
X If LCiESkopia þ pkb then modes a, a þ 1, . . . , mi of
LB2 ¼ Max rimi activity i and b, b þ 1, . . . , mk of activity k cannot lead to
t
i2St
a feasible solution, as stated below.
T X
X mi T X
X mk
Lower bound 3 xijt þ xkjt p1 8ði; kÞ 2 FC ð8Þ
t¼1 j¼a t¼1 j¼b
The workload of activity i, at mode j is pij rij . The
minimum workload for activity i over all its modes is
Minj fpij rij g. where FC ¼ {(i, k)|kAEi4(LCiESkopia þ pkb)}
Activity i should complete before its latest completion Akkan et al (2005) use the same cut for their discrete
time, LCi, hence, it consumes at least Minj fpij rij g units time/cost trade-off model.
of the resource before LCi. The minimum possible total The second cut is an extension of Property 5 for two
resource consumed by the end of LCi is the sum of the activities (i, k) in set FR, at respective modes a and b. If
minimum workloads of the activitiesP whose latest comple- their total resource usage exceeds a known upper bound
tion is no later than LCi, that is, k2Finii Minj fpij rij g
(UB) then modes 1, 2, . . . , a and 1, 2, . . . , b cannot lead to
where Fini:{k|LCkpLCi}. an optimal solution. The cut is as stated below.
For activity i, we distribute the total workload evenly in
[0,PLCi] and round up to the smallest integer, that is, X
T X
a T X
X b
k2Finii
Minj fpij rij g xijt þ xkjt p1 8ði; kÞ 2 OC ð9Þ
LCi . To find a lower bound, we consider all
t¼1 j¼1 t¼1 j¼1
P
Minj fpij rij g
k2Finii
activities and select the maximum of the LCi
where OC ¼ {(i, k)|kAFR4((a, b) such that ria þ ribXUB)}.
values. P
Therefore, a valid lower bound becomes Our next two cuts use the results of Properties 3 and 4.
Minj fpij rij g
k2Finii
Maxi LCi .
xi1ESi ¼ 1 8i 2 ESO ð10Þ
Note that this bound considers all activities that
complete no later than LCi. However, there may exist an
activity such that LCk4LCi, but LSkoLCj. For such an ESO ¼ Set of activities at their first mode and earliest
activity, the minimum processing before LCi is LCiLSk. start times by Property 3.
We include this portion with minimum weight rkmk and get
the following expression for our third lower bound. xi1LSi ¼ 1 8i 2 LSO ð11Þ
maximum resource period may shift to another time period, activity, hence, a higher CNC implies a more connected
which in turn may reduce the maximum resource usage. network. Alvarez-Valdes and Tamarit (1989) show that as
For all iASmaxt and activity i starts earlier than its latest the CNC values increase, that is, the networks become
start time more connected, the problems become easier to solve.
Let R4max(i) be the maximum resource usage In our main experiment, we generate 10 problem instances
found by shifting activity i to its latest start time for each N, hence, a total of 100 problem instances. The
and adjusting the early start, late start and start CNC values of those instances are 1.5 and their deadlines are
times of its predecessors and successors. If such a set to the minimum possible project completion times. With
shift leads to a violation in the project deadline, such tight deadlines we value time-based competition and
we set R4max(i) to infinity. with such small CNC values we consider hard-to-solve
Note that a solution in the neighbourhood is found problem instances. We analyse the effect of the deadline
by either changing the start time or mode of an activity. values by solving 20 additional problems.
We select the move that defines Mini(R1max(i), R2max(i), We solve our mathematical model and its LP relaxation
R3max(i), R4max(i)). The selected move to a neighbour with GAMS using CPLEX 12 solver. The algorithms are
solution is said to be improving if its resulting maximum coded in the C++ programming language. We conduct
resource usage is smaller than the current maximum our experiments on an Intel Core(2) Duo 2.33 GHz, 1 GB
resource usage. RAM computer.
Our procedure starts with the best of the three con- Our initial experiments have revealed that, in 1 h,
struction schedules. It generates the neighbour solutions CPLEX cannot return optimal solutions to our MILP
and each time moves to a best solution in the neighbour- model for instances with more than 20 activities.
hood. If no improving moves are found in the last five
iterations or the number of iterations reaches to 1000,
Main experiment
we continue with another construction schedule. In our
preliminary runs we try many values for iteration limits In our main runs we take the shortest possible project
(for non-improving moves from 2 to 10 and for number of completion time as the deadline and 1.5 as the CNC value.
iterations from 100 to 2000). We find that five non- We investigate the power of the mode elimination rules and
improving moves and 1000 iterations best catch the the performances of the lower bounds and heuristic
trade-off between solution speed and solution quality. procedures.
We terminate when all three construction schedules are
evaluated.
Performance of the optimality properties
We apply Properties 2, 3 and 4 with the hope of settling the
Computational experience mode decisions, and Properties 1, 5 and 6 with the hope of
eliminating some modes of the activities whose mode
Our data set includes projects with 10, 20, 30, 40, 50, 60, decisions could not be settled. Accordingly, first we define
70, 80, 90 and 100 activities. We adapt Ranjbar and the critical activities and set them to their shortest modes
Kianfar’s (2007) scheme to generate the durations and using the result of Property 2 and apply Properties 3 and 4
resource requirements. In doing so, we generate a uniform to fix the non-critical activities to their shortest modes.
random number, wi, in [10, 100] for each activity i and Then, the modes of the unfixed activities that would lead to
generate the first mode
of the activity with resource infeasible solutions are eliminated by Property 1 and that
pffiffiffiffiffi
requirement ri1 ¼ wi and processing time pi1 ¼ would lead to non-optimal solutions are eliminated by
dwi =ri1 e: The procedure then generates the second mode Properties 5 and 6.
of activity i with resource requirement ri2 ¼ ri11 and its Table 1 includes the number of activities, number of
corresponding processing requirement pi2 ¼ dwi =ri2 e: This non-critical activities found by Property 2, the number of
new mode is accepted as the second mode only if its activities fixed by Properties 3 and 4. The table also
processing requirement is higher than that of the first includes the number of modes for the unfixed activities and
mode. This mode generation scheme continues until the the number of modes reduced by Property 1.
resource requirement reduces to 1. Note that when N ¼ 10, about half of the activities are
The p maximum
ffiffiffiffiffiffiffiffi
value of wi is 100, hence, the ri1 can be at set critical by Property 2 and taken out of consideration.
most 100 ¼ 10: With unit reductions of the ri1 value, a On average, one out of five non-critical activities are fixed,
maximum of 10 modes can be found for each activity. that is, their start times are set to their early or late start
We use Project Scheduling Instance Generator (ProGen) times at their shortest modes, by Properties 3 and 4. When
of Kolisch and Sprecher (1996) to generate the networks of N ¼ 100 the average number of non-critical activities is
specified complexity. The coefficient of network complexity found to be 87 by Property 2, about six of which are fixed
(CNC ) is defined as the number of precedence relations per by Properties 3 and 4. The power of Property 2 deteriorates
784 Journal of the Operational Research Society Vol. 65, No. 5
Max
48
7
24
11
19
30
17
17
45
30
N ¼ 10, 40 and 100, respectively. Property 1 is effective in
Number of modes
eliminated by
23.3
2.1
6.5
6.8
7.6
11.4
10.4
10.2
21.4
18.1
through 4 are all effective in reducing the problem size.
However, we find that Properties 5 and 6 are not as
effective in eliminating modes, they eliminate only
20 modes over 100 problem instances.
Max
560
27
67
123
185
240
317
389
427
488
modes for unfixed
Total number of
503.9
20.1
106.0
159.5
212.7
278.6
340.5
373.9
440.5
optimal solution.
Number of activities
10
3
11
5
6
7
4
4
2.3
0.9
5.0
5.5
2.1
2.2
3.1
2.5
2.5
4.3
70.0
70.0
87.5
89.0
76.7
80.0
82.0
86.7
85.7
87.8
non-critical activities
left by Property 2
53.0
87.4
65.0
71.7
75.2
77.8
80.0
82.4
84.0
84.0
89
23
32
41
52
60
79
14
70
5.3
13.0
87.4
21.5
30.1
38.9
48.0
57.7
67.2
75.6
Table 2 The relative deviations of the lower bounds and best lower bound
N % dev for LB1 % dev for LB2 % dev for LB3 % dev for LB4 % dev for best LB
Aver Max Freq Aver Max Freq Aver Max Freq Aver Max Freq Aver Max Freq
Moreover we observe that the optimal objective function Table 3 The relative deviations and gaps of the construction
values are very low (about 22, 25 and 30 units for N ¼ 10, heuristics
20 and 30, respectively); therefore, a small difference N % dev for ESS % dev for LSS % dev for AELSS
between the optimal objective function value and best
lower bound may lead to a high relative deviation. We find Aver Max Aver Max Aver Max
that the difference between the optimal objective function
value and best lower bound is at most 3 units over all 30 10 31.8 60 28.5 85 22.0 37
20 56.5 120 51.2 136 45.6 105
instances. These observations show that the best lower 30 60.8 104 54.1 81 50.4 85
bound is quite powerful and can be used to estimate the
unknown optimal objective function values. Given this % gap for ESS % gap for LSS % gap for AELSS
performance, we use the best lower bound in evaluating the 10 40.1 78 36.9 106 29.6 53
heuristic procedures. 20 64.1 132 59.2 160 53.0 116
30 72.1 130 64.9 104 61.0 109
40 80.7 126 82.5 137 64.7 107
Performance of the heuristic procedures 50 83.8 119 89.3 134 63.6 75
60 99.3 129 95.3 128 60.9 87
Recall that our heuristic has two main phases: construction 70 94.9 159 124.5 163 86.5 116
and improvement. In the construction phase, three con- 80 80.1 117 105.4 143 72.0 104
struction heuristics, namely Early Start Schedule (ESS), 90 107.4 163 98.8 136 77.2 102
Late Start Schedule (LSS) and Alternating Early-Late 100 156.6 195 193 250 145.3 176
Start Schedule (AELSS), are used. We measure their
performance relative to the optimal solutions for small-
Table 4 The relative deviations and gaps of the construction
sized problem instances up to 30 activities and relative to
heuristics using fastest modes
the best lower bound over all problem set.
We define %dev ¼ ((CSiOPT)/OPT) 100 and %gap ¼ N % dev for ESS % dev for LSS % dev for AELSS
((CSiLB)/LB) 100 where
Aver Max Aver Max Aver Max
CSi ¼ objective function value of construction heuristic I 10 53.3 100 44.4 90 45.5 100
i¼ ESS, LSS, AELSS 20 99.0 205 78.0 164 76.8 175
30 91.3 128 87.7 128 74.2 128
Table 3 reports the relative deviations from the optimal
% gap for ESS % gap for LSS % gap for AELSS
solution and gaps from the best lower bound.
Note that even when N ¼ 10, the construction heuristics 10 63.0 122 53.6 111 54.8 122
return solutions that are about 20–30% away from the 20 108.5 221 87.3 190 85.6 189
optimal solution. This is not surprising as they consider 30 114.2 183 109.8 154 93.8 135
40 131.4 204 131.0 241 110.6 185
only the earliest or latest start times and a single specified 50 139.2 206 144.3 203 108.0 150
mode. Table 3 also reveals that when the problem size gets 60 156.7 194 154.6 233 109.7 154
larger, the performances of the construction heuristics 70 158.8 284 183.1 261 143.2 222
deteriorate. The table also shows that AELSS is by far the 80 135.3 190 175.3 267 136.5 252
best construction heuristic as it performs better than the 90 170.3 278 152.2 213 133.2 191
100 157.9 251 157.5 205 120.9 175
others over all problem sizes.
To see the effect of the mode decisions on the
performance of the construction heuristics, we set all Note that when the fastest modes are used instead of
activities to their fastest modes and report the relative the modes returned by the strengthened LPR, the
deviations from the optimal solution and gaps from the performances of the construction heuristics significantly
best lower bound in Table 4. deteriorate. The increases in the deviations and gaps are
786 Journal of the Operational Research Society Vol. 65, No. 5
CPU times by IH and HKH are also quite small. Note from heuristic procedures. Table 8 compares the power of the
the table that when N ¼ 80 the maximum times used by IH properties for the deadline value of the shortest project
and HKH are less than 90 and 70 s, respectively, and when completion time and 1.5 times the shortest project
N ¼ 100 the maximum times used by IH and HKH are less completion time for N ¼ 20 and 80. Each problem instance
than 140 and 115 s, respectively. For all problem instances is solved using two deadline values.
HKH produces slightly quicker solutions than IH, however, As can be observed from the table, all activities are
at an expense of significantly lower quality. non-critical when the deadline is above the shortest
We computeP the PtotalPamountPof resource idle time left project completion time. We can conclude that our
as T R Tt¼1 N i¼1 r
j2mi ij
t
k¼tpij þ1 xijk : Note that properties are effective for all deadline values and their
the smaller the total idle time the better the resource effect increases as the deadline values decrease. Note
levelling. We report the total idle time values for IH and that for N ¼ 80, when the deadline is the shortest
HKH in Table 7. possible completion time, Properties 2, 3 and 4 leave
As can be observed from the tables, our heuristic an average of 374 modes, 21 of which are further
algorithm IH levels the resource usages significantly better reduced by Property 1. When the deadline increases by a
than HKH. Note that the idle times of IH are about two factor of 1.5, the average number of modes left by
times smaller than those of HKH, for all problem sizes. Properties 2, 3 and 4 is 559, 15 of which are removed by
These results are in line with our expectations that IH gives Property 1. This is due to the fact that low deadline
smaller maximum loads that lead to smaller idle times and values lead to small activity slack times. Small slack time
better resource levelling. values directly affect the elimination power of the
properties.
Table 9 reports on the performances and solution times
Effect of the deadline values of the heuristic procedures for two different deadline
We analyse the effect of the deadline values on the values. We observe that in nine out of 10 instances LPR
performance of the reduction properties, lower bounds and produce optimal solutions and in one solution the
deviation is about 10%. The optimal LPR solutions
become closer to the optimal Integer Program (IP)
Table 7 The total resource idle time kept solutions for high deadline values, as the activities have
higher slack times, hence, higher choices for mode assign-
N IH HKH ments, thereby, a higher chance of integer mode assign-
ments. The CPU times of the strengthened LP relaxation
Average Maximum Average Maximum
are also given in Table 9.
10 21.08 32.81 36.59 49.38 Note that the performances of the heuristic procedure
20 19.85 32.42 37.87 52.13 deteriorate as the deadline values increase. The average
30 17.73 25.54 39.40 46.95 and maximum percentage gaps of IH are about 20
40 15.58 20.14 37.58 44.45
50 16.79 23.33 38.45 54.73 and 30%, respectively, when the deadline is set to
60 16.81 25.60 37.34 46.23 1.5 shortest project completion time. These gaps for
70 15.52 20.61 37.89 46.15 HKH are about 60 and 90% when N ¼ 20 and 45 and
80 19.59 35.99 37.69 42.39 75% when N ¼ 80.
90 17.49 25.32 37.70 45.57 As can be observed from the table, the CPLEX times
100 19.71 27.45 37.51 46.82
significantly increase as the deadline values increase. This is
Aver Max Aver Max Aver Max Aver Max Aver Max
Table 9 The deviations and gaps of the heuristics and CPU times for different deadline values
N Deadline % dev for IH % dev for HKH % gap for IH % gap for HKH
CPLEX time CPU time of LB4 CPU time of IH CPU time of HKH
due to the fact that the activities have higher slack times, Properties of the optimal solution
thereby, a higher number of choices for their start times.
Properties 1 and 2 hold for the multi-renewable resources,
The CPU times of the strengthened LPRs increase about
as they are related to feasibility.
two times as the deadline values increase by 1.5 times.
We extend the other properties using the following
There is no significant effect of the deadline values on the
additional notation.
solution times of the heuristic procedures. All CPU times
are below 1.5 min.
LBs(UBs) ¼ A lower (an upper) bound on the usage of
resource s
LstE(LstL) ¼ Amount of resource s used at period t when
Extensions to multi-resource environments all activities are at their shortest duration
In this section we discuss some possible extensions of our modes and start at their earliest (latest) start
results, in particular our reduction procedures and lower times
bounds, to multi-renewable resource environments. We
assume there are S renewable resource types and mode j of We omit the proofs of the properties as they directly
activity i for resource s is characterized by the following follow their single resource counterparts.
parameters.
Property 3 (extended) There exists an optimal schedule in
rijs ¼ amount of resource s required by activity i when which activity i satisfying LstEoLBs for all s and every
executed in mode j tA[0, ECi], starts at ESi and is processed at its shortest
pij ¼ duration (processing time) of activity i when executed duration mode.
in mode j
Property 4 (extended) There exists an optimal schedule in
We assume pijopij þ 1 for all i and j, accordingly, the first which activity i satisfying LstLoLBs for all s and every
mode of an activity is its shortest duration mode and tA[LSi, T] starts at LSi and is processed at its shortest
rijsoriks implies pij4pik for at least one s. duration mode.
The objective is to minimize the total resource cost,
that is,
Minimize SSs ¼ 1csRs where cs is the cost of buying P 5 (extended) If rijsoriks implies pij4pik and
Property
resource s and Rs is the maximum amount of resource s riks þ j2FRi rjmj s 4UBs for all s then modes 1, . . . , k of
used. activity i cannot lead to an optimal solution.
Note that when S ¼ 1, the objective function reduces to
minimize R, that is, the maximum amount of the single Property P 6 (extended) If rijsoriks implies pij4pik
resource used. and riks þ j2FSi rj1s pLBs for all s then there exists an
PTheN P
multi-resource
PMinftþpij 1;LSi g constraint becomes optimal schedule in which activity i is assigned to modes
i¼1 r
j2mi ijs k¼Maxft;ESi g xijk pRs 8s and t. 1, . . . , k.
Erdem Colak and Meral Azizoglu—A RIP with time/resource trade-offs 789
Lower bounds: A valid lower bound on the optimal problem. European Journal of Operational Research 165(2):
SSs ¼ 1csRs value is SSs ¼ 1csLB(Rs) where LB(Rs) is a valid 339–358.
lower bound on the optimal Rs value. LB(Rs) can be found Alcaraz J, Maroto C and Ruiz R (2003). Solving the multi-mode
resource-constrained project scheduling problem with genetic
by our lower bounding procedures for each resource s algorithms. Journal of the Operational Research Society 54(6):
independently. In doing so, the modes with rijs4riks and 614–626.
pij4pik are eliminated and remaining tasks are indexed Alvarez-Valdes R and Tamarit JM (1989). Heuristic algorithms for
such that rijsoriks implies pij4pik. resource constrained project scheduling: A review and an
empirical analysis. In: Slowinski R and Weglarz J (eds). Advances
in project scheduling. Elsevier: Amsterdam, pp 113–134.
Demeulemeester E (1995). Minimizing resource availability costs
Conclusions in time-limited project networks. Management Science 41(10):
1590–1598.
This study considers an RIP with time/resource trade-offs in Demeulemeester E, De Reyck B and Herroelen W (2000). The
project networks. We assume that there is a single renewable discrete time/resource trade-off problem in project networks:
resource and the processing requirement of an activity can A branch and bound approach. IIE Transactions 32(11):
be reduced by additional resources. Our problem is to find 1059–1069.
Drexl A and Kimms A (2001). Optimization guided lower and
activity durations in order to minimize the maximum upper bounds for the resource investment problem. Journal of
resource usage while meeting the pre-specified deadline. The the Operational Research Society 52(3): 340–351.
problem is observed to be strongly NP-hard. Gather T, Zimmermann J and Bartels J (2008). Tree based methods
We propose several problem size reduction techniques, for resource investment and resource levelling problems. Procee-
lower bounding and heuristic procedures. We observe that dings of POMS 2008, April 28–30, Istanbul, Turkey, pp 94–98.
Hartmann S and Briskorn D (2010). A survey of variants and
our procedures significantly reduce the problem size either extensions of the resource-constrained project scheduling pro-
by eliminating the activity modes or by setting the start blem. European Journal of Operational Research 207(1): 1–14.
times. We find that our lower bounds produce solutions Herroelen W, De Reyck B and Demeulemeester E (1998).
that are very close to the optimal ones. Resource-constrained project scheduling: A survey of recent
We also observe the satisfactory behaviour of our developments. Computers and Operations Research 25(4):
279–302.
heuristic procedure and its superiority over the heuristic Hsu CC and Kim DS (2005). A new heuristic for the multi-mode
reported in the literature. The experiments have revealed resource investment problem. Journal of the Operational
that the performance of our heuristic procedure deterio- Research Society 56(4): 406–413.
rates with the increases in the number of activities and Jozefowska J, Mika M, Rozycki R, Waligora G and Weglarz J
deadline values. (2001). Simulated annealing for multi-mode resource-constrained
project scheduling. Annals of Operations Research 102(1–4):
We discuss some possible extensions of our optimality 137–155.
properties and lower bounds to the multi-resource Kolisch R and Sprecher A (1996). PSPLIB—A project scheduling
environments. Future research may consider more thor- library. European Journal of Operational Research 96(1):
ough analysis of the multi-resources like proposing more 205–216.
powerful reduction mechanisms and lower bounds and Mohring R (1984). Minimizing costs of resource requirements in
project networks subject to a fixed completion time. Operations
developing efficient solution approaches. Research 32(1): 89–120.
We hope that our study fills an important gap in the Mori M and Tseng CC (1997). A genetic algorithm for multi-mode
project scheduling literature. The future research around resource constrained project scheduling problem. European
our problem may consider the following issues. Our Journal of Operational Research 100(1): 134–141.
reduction mechanisms and lower bounds can be embedded Neumann K and Zimmermann J (1999). Resource levelling for
projects with sequence-dependent time windows. European
into an optimization procedure. Our experiments reveal Journal of Operational Research 117(3): 591–605.
that the performances of our procedures deteriorate as the Ranjbar M and Kianfar F (2007). Solving the discrete time/resource
deadline values increase. Future research may point out trade-off problem in project scheduling with genetic algorithms.
developing more efficient solution procedures to handle Applied Mathematics and Computation 191(2): 451–456.
longer deadlines. Other resource levelling objectives like Ranjbar M, Kianfar F and Shadrokh S (2008). Solving the resource
availability cost problem in project scheduling by path relinking
minimizing the range of resource usages might also be and genetic algorithm. Applied Mathematics and Computation
worth studying. 196(2): 879–888.
Ranjbar M, De Reyck B and Kianfar F (2009). A hybrid scatter
search for the discrete time/resource trade-off problem in project
Acknowledgements —We are grateful to two anonymous referees for scheduling. European Journal of Operational Research 193(1):
their valuable comments. 35–48.
Shadrokh S and Kianfar F (2007). A genetic algorithm for resource
investment project scheduling problem, tardiness permitted with
References penalty. European Journal of Operational Research 181(1): 86–101.
Shahsavar M, Niaki STA and Najafi AA (2010). An efficient
Akkan C, Drexl A and Kimms A (2005). Network decomposition- genetic algorithm to maximize net present value of project
based benchmark results for the discrete time-cost tradeoff payments under inflation and bonus-penalty policy in resource
790 Journal of the Operational Research Society Vol. 65, No. 5
investment problem. Advances in Engineering Software 41(7): Yamashita DS and Morabito R (2009). A note on time/cost trade-
1023–1030. off curve generation for project scheduling with multi-mode
Sprecher A, Hartmann S and Drexl A (1997). An exact algorithm resource availability costs. International Journal of Operational
for project scheduling with multiple modes. OR Spektrum 19(3): Research 5(4): 429–444.
195–203. Yamashita DS, Armentano VA and Laguna M (2007). Robust
Talbot FB (1982). Resource-constrained project scheduling with optimization models for project scheduling with resource
time-resource tradeoffs: The nonpreemptive case. Management availability cost. Journal of Scheduling 10(1): 67–76.
Science 28(10): 1197–1210.
Van Peteghem V and Vanhoucke M (2009). A genetic algorithm for
the preemptive and non-preemptive multi-mode resource-con-
strained project scheduling problem. European Journal of Received October 2011;
Operational Research 201(2): 409–418. accepted March 2013 after three revisions
Reproduced with permission of the copyright owner. Further reproduction prohibited without
permission.