Learning Precedences for Scheduling Problems with Graph Neural Networks

Hélène Verhaeghe, Quentin Cappart, Gilles Pesant, Claude-Guy Quimper

Abstract:

The resource constrained project scheduling problem (RCPSP) consists of scheduling a finite set of
resource-consuming tasks within a temporal horizon subject to resource capacities and precedence
relations between pairs of tasks. It is NP-hard and many techniques have been introduced to
improve the efficiency of CP solvers to solve it. The problem is naturally represented as a directed
graph, commonly referred to as the precedence graph, by linking pairs of tasks subject to a precedence.
In this paper, we propose to leverage the ability of graph neural networks to extract knowledge from
precedence graphs. This is carried out by learning new precedences that can be used either to add
new constraints or to design a dedicated variable-selection heuristic. Experiments carried out on
RCPSP instances from PSPLIB show the potential of learning to predict precedences and how they
can help speed up the search for solutions by a CP solver.