Explanation-Based Repair Techniques For Solving Dynamic Scheduling Problems

Explanation-based repair techniques for solving dynamic scheduling problems
Abdallah Elkhyari
1
and Christelle Gu eret
1,2
and Narendra Jussien
1
1

Ecole des Mines de Nantes BP 20722 F-44307 Nantes Cedex 3 France
2
IRCCyN Institut de Recherche en Communications et Cybern etique de Nantes
email: {aelkhyar, gueret, jussien}@emn.fr
Keywords: Dynamic scheduling, repair-based techniques, explanation-based constraint programming,
Resource-Constrained Project Scheduling Problem.
Introduction
Scheduling problems have been studied a lot over the last
decade. Due to the complexity and the variety of such
problems, most works consider static problems in which
activities are known in advance and constraints are xed.
However, every scheduling problem is subject to unexpected
events (consider for example a new activity to schedule, or
a machine breakdown). In these cases, a new solution is
needed in a preferably short time taking these events into
account and as close as possible to the current solution.
In this paper, we present an exact approach for solv-
ing dynamic scheduling problems. This approach uses
explanation-based constraint programming and operational
research techniques. Our tools have been designed for a gen-
eral scheduling problem: the Resource-Constrained Project
Scheduling Problem (RCPSP).
Problem description
The RCPSP can be dened as follows: let A = {1, . . . , n}
be a set of activities, and R = {1, . . . , r} a set of renew-
able resources. Each resource k is available in a constant
amount R
k
. Each activity i has a duration p
i
and requires
a constant amount r
ik
of the resource k during its execu-
tion. Preemption is not allowed. Activities are related by
precedence constraints, and resource constraints require that
for each period of time and for each resource, the total de-
mand of resource does not exceed the resource capacity. The
objective considered here is to nd a solution for which
the end of the schedule is minimized. This problem, de-
noted by PS/prec/C
max
(Brucker et al. 1999), is NP-hard
(Blazewicz, Lenstra, & Rinnoy Kan 1983).
The static RCPSP has been extensively studied (Brucker et
al. 1998). A major difculty in this problem is to maintain
the resource limitation over the horizon time. Several deduc-
tion rules exist: core-times (Klein &Scholl 1999), energetic-
reasoning (Erschler &Lopez 1990), task-interval (Caseau &
Laburthe 1996), etc.
The dynamic RCPSP is seldom studied. Two classical
methods are used to solve it:
recomputing a new schedule each time an event occurs.
This is quite time consuming and may lead to a solution
quite different from the previous one.
building a partial schedule and completing it progres-
sively as time goes by (like in on-line scheduling prob-
lems). In this case the schedule cannot be constructed in
advance.
Recently, (Artigues & Roubellat 2000) introduced a for-
mulation of the RCPSP based on a ow network model. They
developed a polynomial algorithm based on this model in or-
der to be able to insert an unexpected activity.
Explanation-based constraint programming
Constraint programming techniques have been widely used
to solve scheduling problems (Klein 1999). Constraint pro-
gramming is based upon the notion of constraint satisfaction
problems.
A constraint satisfaction problem (CSP) consists in a set
V of variables dened by a corresponding set of possible
values (the domains D) and a set C of constraints. A solu-
tion for the CSP is an assignment of the variable such that all
the constraints are satised.
Explanation-based constraint programming (e-
constraints) has already proved its interest in many
applications (Jussien 2001). This section recalls what is an
explanation and how it can be used.
Explanations
In the following, we consider a CSP (V, D, C). Decisions
(variable assignments) made during the enumeration phase
of the resolution of this problem correspond to adding or
removing constraints fromthe current constraint system(eg.,
upon backtracking).
A conict set (a.k.a. nogood) is a subset of the cur-
rent constraints system of the problem that, left alone, leads
to a contradiction (no feasible solution contains a con-
ict set). A conict set divides into two parts: a subset
of the original set of constraints (C
C) and a sub-
set of decision constraints introduced so far in the search:
(C
v
1
= a
1
v
k
= a
k
).
In a conict set composed of at least one decision con-
straint, a variable v
j
is selected and the previous formula
can be rewritten as
1
: C
i[1..k]\j
(v
i
= a
i
) = v
j
= a
j
1
A conict set that does not contain such a constraint denotes
an over-constrained problem.
The left hand side of the implication constitutes an elim-
inating explanation for the removal of value a
j
from the
domain of variable v
j
and is noted expl(v
j
= a
j
).
Classical CSP solvers use domain-reduction techniques
(removal of values). Recording eliminating explanations
is sufcient to compute conict sets. Indeed, a contradic-
tion is identied when the domain of a variable v
j
is emp-
tied. A conict set can easily be computed from the elim-
inating explanations associated with each removed value:
aD
v
j
expl(v
j
= a)
There exist generally several eliminating explanations for

the removal of a given value. Recording all of them leads to
an exponential space complexity. Another technique relies
on forgetting (erasing) eliminating explanations that are no
longer relevant
2
in the current variable assignment. By do-
ing so, the space complexity remains polynomial. We keep
only one explanation at a time for a value removal.
Using explanations
Explanations can be used in several ways (Jussien 2001).
For example, when debugging, explanations can be used
to: clearly explain failures, explain differences between in-
tended and observed behavior for a given problem (why is
value n not assigned to variable x ?).
Explanations can also be used to determine direct or in-
direct effects of a given constraint on the domains of the
variables of the problem, and for dynamic constraint re-
moval. This is the case with the justication system used
in (Bessi` ere 1991) for solving dynamic CSP. This justica-
tion system is actually a partial explanation system. More-
over, being able to explain failure and to dynamically re-
move a constraint facilitates the building of dynamic over-
constrained problem solvers.
We added explanation handling within a branch and
bound algorithm in order to provide a dynamic RCPSP prob-
lem solver.
Solving dynamic RCPSP
We developed an environment for solving dynamic RCPSP
that is based upon:
a branch and bound algorithm (inspired from (Brucker et
al. 1998)) within a constraint programming solver: at
each node, deduction rules are applied in order to deter-
mine redundant information. In the following, we call
constraint each initial constraint of the problem (prece-
dence and resource constraints), but also each decision
taken by the branching scheme, and each deduction made
(thanks to propagation rules) during the search.
an extensive use of explanations. Explanations are
recorded during the search, and improved thanks to prop-
agation rules (namely core-times and task-interval) which
have been upgraded in order to provide a precise explana-
tion for every deduction made.
2
An explanation is said to be relevant if all the decision con-
straints in it are still valid in the current search state.
Our environment can efciently handle dynamic events.
Indeed, an unexpected event leads to add, modify or remove
a constraint in the system. In the rst two cases, if the cur-
rent solution is no more valid, then the explanations tell us
which are the constraints responsible of the contradiction.
Repairing is done by removing at least one constraint from
the explanation. The resulting solution is generally quite
similar to the previous one, and is found faster than if we
have had to solve the problem from scratch.
Notice that, as we saw before, if the constraints of the
explanation are only initial constraints of the problem, then
the problem is over-constrained. In this case, our system
explains to the user why the problem has no solution. Re-
sponsible constraints can therefore be relaxed.
Conclusion
We developed a dynamic environment for solving general
scheduling problems which can let the user interact with
his/her problem: dene a new time-window for an activ-
ity, add/remove precedence constraints, add/remove must-
overlap constraints (i.e. stating that two activities must over-
lap), dene/reduce/remove a required time-lag between two
activities, etc.
We are currently experimenting our tools in order to show
the interest of a real dynamic approach compared to re-
execution from scratch. We do expect promising results.
References
Artigues, C., and Roubellat, F. 2000. A polynomial activity insertion algorithm in a
multi-resource schedule with cumulative constraints and multiple modes. European
Journal of Operational Research 127(2):297316.
Bessi` ere, C. 1991. Arc consistency in dynamic constraint satisfaction problems. In
Proceedings AAAI91.
Blazewicz, J.; Lenstra, J.; and Rinnoy Kan, A. 1983. Scheduling projects subject to
resource constraints: classication and complexity. Discrete Applied Mathematics
5:1124.
Brucker, P.; Knust, S.; Schoo, A.; and Thiele, O. 1998. A branch and bound algo-
rithm for the Resource-Constrained Project Scheduling Problem. European Journal
of Operational Research 107:272288.
Brucker, P.; Drexl, A.; M oring, R.; Neumann; and Pesch, E. 1999. Resource-
constrained project scheduling: notation, classication, models and methods. Euro-
pean Journal of Operational Research 112:341.
Caseau, Y., and Laburthe, F. 1996. Cumulative scheduling with task-intervals. In
Joint International Conference and Symposium on Logic Programming (JICSLP).
Erschler, J., and Lopez, P. 1990. Energy-based approach for task scheduling under
time and resource-constraints. In Second international workshop on Project Man-
agement and Scheduling (PMS), 115121.
Jussien, N. 2001. e-constraints: explanation-based constraint programming. In
CP01 Workshop on User-Interaction in Constraint Satisfaction.
Klein, R., and Scholl, A. 1999. Computing lower bounds by destructive improve-
ment: an application to Resource-Constrained Project Scheduling Problem. Euro-
pean Journal of Operational Research 112:322345.
Klein, R. 1999. Scheduling of Resource Constraints Projects. Boston: Kluwer
Academic.

Explanation-Based Repair Techniques For Solving Dynamic Scheduling Problems

Uploaded by

Copyright:

Available Formats

Explanation-Based Repair Techniques For Solving Dynamic Scheduling Problems

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Explanation-Based Repair Techniques For Solving Dynamic Scheduling Problems

Uploaded by

Copyright:

Available Formats

Explanation-based repair techniques for solving dynamic scheduling problems

There exist generally several eliminating explanations for

You might also like