1
Properties of lower bounds for the RCPSP
Evgeny R. Gafarov1, Alexander A. Lazarev1 and Frank Werner2
1
2
Institute of Control Sciences of the Russian Academy of Sciences, Russia
axel73@mail.ru, jobmath@mail.ru
Fakultät für Mathematik, Otto-von-Guericke-Universität Magdeburg, Germany
frank.werner@mathematik.uni-magdeburg.de
Keywords: Project scheduling, Makespan, Lower bounds
1
Abstract
We show that the calculation of the well-known lower bound of Mingozzi for the RCPSP
is an NP-hard problem and that the relative error of this lower bound can be equal to
O(log n), where n is the number of jobs.
2
Introduction
Problem RCPSP may be formulated as follows. Given a set N = {1, . . . , n} of jobs. A
constant amount of Qk > 0 units of resource k, k = 1, . . . , K, is available at any time. Job
j ∈ N has be processed for pj ≥ 0 time units without preemption. During this period,
a constant amount of qjk ≥ 0 units of resource k is occupied. Furthermore, finish-start
precedence relations i → j are defined between the jobs according to an acyclic directed
graph G. The objective is to determine the starting times Sj for each job j = 1, . . . , n,
in such a way that: at each time t, the total resource demand is less than or equal to the
resource availability for each resource type; the given precedence constraints are fulfilled;
the makespan Cmax = maxnj=1 Cj , where Cj = Sj + pj , is minimized.
∗
Let Cmax
be the optimal value of the objective function for the problem when pre∗
(pmtn) be the optimal value when preemptions are
emptions are not allowed and Cmax
allowed.
3
Lower Bound of Mingozzi et al.
We consider a linear programming formulation that partially relaxes the precedence
constraints and allows preemption. The columns of this LP correspond to the so-called
non-dominated feasible subsets. A feasible set X is a set of jobs that may be processed
simultaneously, i.e., there are no precedence
relations between any pair i, j ∈ X and all
qik ≤ Qk for k = 1, . . . , K). Such a set is called
resource constraints are satisfied (i.e.,
i∈X
non-dominated if it is not a proper subset X of another feasible set Y . We consider all
non-dominated feasible sets and additionally the one-element sets {i} for all i = 1, . . . , n.
We denote all these sets by X1 , X2 , . . . , Xf , where f is the number of such sets, and
associate each set Xj with an incidence vector aj ∈ {0, 1}n defined by aji = 1 if i ∈ Xj ,
and aji = 0 otherwise, j = 1, . . . , f .
Furthermore, let xj be a variable denoting the number of time units over which all the
jobs in Xj are processed simultaneously. Then the following linear programming problem
provides a lower bound LBM (Mingozzi A. et. al. 1998) for the RCP SP by relaxing the
precedence constraints and allowing preemption:
2
f
xj −→ min
(1)
j=1
⎧ f
⎨
aji xj ≥ pi , i = 1, . . . , n;
⎩ j=1
xj ≥ 0, j = 1, . . . , f.
(2)
It is known that the calculation of LBM is an N P -hard problem (by a reduction from
the N P -hard Bin packing problem) (Lazarev A.A. and Gafarov E.R. 2008), and there are
∗
Cmax
≈ 2.
instances for which LB
M
4
Relative errors of well-known lower bounds for the problem
∗
In the paper (Lazarev A.A. and Gafarov E.R. 2008), there is a conjecture that Cmax
<
∗
2 · Cmax (pmtn). This conjecture is true for the special case of problem P m|prec|Cmax
(Lawler E.L. et. al. 1998), for the special case of RCP SP with a constant amount of
Q1 > 0 units of a single resource and without precedence constraints, and for the special
case for which there are only one or two preempted job in an optimal schedule for RCPSP
with preemptions. However, the conjecture is false for the general case.
Theorem 1. There exists an instance of RCP SP for which
∗
Cmax
= O(log n).
∗ (pmtn)
Cmax
Proof.
We consider an instance of the type given in Fig. 1 (a). For this instance, we have
m + 1 levels of jobs. At the highest level, we have a job j11 with processing time M p, at
the second highest level, we have two jobs j12 and j22 with processing times M
2 p, and so
m+1
→ e2 →
on. At the lowest level, we have a chain of short and very short jobs e1 → j1
m+1
→ eM+1 , where pei = ε, qei = m + 1, i = 1, . . . , M + 1,
j2m+1 → . . . → eM → jM
and pj m+1 = p, qj m+1 = 1, i = 1, . . . , M . For each job jik from level k, k = 2, . . . , m,
i
i
M
p, qjik = 1. At each level k, k = 2, . . . , m, we have a chain of jobs
we have pjik = 2(k−1)
k
e1 → j1 → e M +1 → j2k → e2 M +1 → . . . → j2kk−1 → eM+1 . Some of these precedence
2(k−1)
2(k−1)
relations are illustrated in Fig. 1 (a). For this instance, we have (M + 1)ε ≪ p.
By dotted lines, we mark all jobs which will be processed in parallel in an optimal
schedule for the non-preemptive problem.
In Fig. 1 (b), for an instance with only m + 1 = 6 levels, we give the resulting optimal
∗
∗
= 112p + 33ε.
(pmtn) = 32p + 33ε and Cmax
solution. For this instance, we obtain Cmax
∗
∗
For the general case, we have Cmax (pmtn) = M p + (M + 1)ε and Cmax = M p + m
2 Mp+
(M + 1)ε. Hence, we obtain
∗
m+2
Cmax
≈
.
∗
Cmax (pmtn)
2
Let us now express M by means of m. We have 2m = M . Then
n = 2M + 1 + 1 + 2 + . . . + 2m−1 = 2M + 1 + 2m − 1 = 2 · 2m + 2m = 3 · 2m
from which we obtain
m = log
n
.
3
3
Therefore, we get
∗
Cmax
∗
Cmax (pmtn)
≈
m+2
log n − log 3 + 2
=
.
2
2
As a consequence, there exists an instance of RCP SP for which
we obtain the following result:
∗
Cmax
LBM
= O(log n), and
Theorem 2. There exists a type of instances of RCP SP for which
∗
Cmax
= O(log n),
LBM
and the calculation of LBM is an NP-hard problem.
The idea of constructing such instances is not difficult. The instance contains two subsets of jobs N1 and N2 . The jobs from the first subset correspond to the instance illustrated
in Fig. 1, where Q1
= m+1. In the set N2 , we have n independent jobs with unit processing
times pj = 1 and j∈N2 qj = 2m + 2. Additionally, we have a dummy job o1 such that
j → o1 → l for all j ∈ N1 , l ∈ N2 . It is obvious that we can give a reduction from the
partition problem to the problem of calculating LBM for this type of instances.
Additionally, let us consider the relaxation of the problem in which we do not take
into consideration non-preemptive jobs, or different processing times, or different values qi ,
′
the optimal value of the
or we do not consider the precedence relations. Denote by Cmax
objective function for the relaxed instance. Then there exist instances for which
∗
Cmax
= O(log n),
′
Cmax
′
i.e., we have bad approximation ratio for the lower bound Cmax
.
Acknowledgements
Partially supported by DAAD (Deutscher Akademischer Austauschdienst;
A/08/80442/Ref. 325).
References
Mingozzi A., Maniezzo V., Ricciardelli S. and Bianco L., 1998, “An exact algorithm for project
scheduling with resource constraints based on new mathematical formulation", Management
Science, Vol. 44, pp. 714–729.
Lawler E.L., Lenstra J.K., Rinnooy Kan A.H.G. and Shmoys D.B., 1998, “Sequencing and Scheduling: Algorithms and Complexity.", Report BS-R8909, Centre for Mathematics and Computer
Science, Amsterdam, The Netherlands.
Lazarev A.A. and Gafarov E.R. , 2008, “On project scheduling problem.", Automation and Remote
Control. Vol. 69, N 12, pp. 2070-2087
Contribution to C max
{
+MP
+1|2 Mp
+1|2 Mp
pi = ½ Mp
p = 1/4 Mp
i
pi = 1/8 Mp
qi1 = m+1, p =
+1|2 Mp
+1|2 Mp
precedence relations
i
(a)
qi1 = 1, p = p
i
+1|2 Mp
2M+1 jobs, total processing time Mp+(M+1)
(b)
Fig. 1. An instance for illustrating Theorem 1
m levels
pi = Mp
pi = ½ Mp
32p
p
2p
p
4p
p
2p
p
8p
p
2p
p
4p
p
2p
p
2p
p
4p
p
2p
p
8p
p
2p
p
4p
p
2p
p
16p
p
112p+33
4