A Study of Adaptive Restarting Strategies for Solving Constraint Satisfaction Problems

In this paper we present a study of four generic strategies for solving constraint satisfaction problems (CSPs) of any kind, be they soluble or insoluble. All four methods combine learning with restarting, two use a fixed cutoff restarting strategy followed by a run to completion, the other two use universal restarting strategies where the cutoff varies from run to run. Learning takes the form of weighting constraints which repeatedly cause failure. This information is then used by a variable ordering heuristic to identify bottleneck variables in search. We evaluate these restarting approaches on scheduling problems and radio link frequency assignment problems (RLFAPs). We show that this form of learning is suited for combining with a restarting approach, and we provide insight into how each approach works.