Extending GENET with lazy arc consistency

Constraint satisfaction problems (CSP's) naturally occur in a number of important industrial applications, such as planning, scheduling and resource allocation. GENET is a neural network simulator to solve binary constraint satisfaction problems. GENET uses a convergence procedure based on a relaxed form of local consistency to nd assignments which are locally minimal in terms of constraint violation. It uses heuristic learning to escape local minima which do not represent solutions. We describe a lazy arc consistency technique which is suitable for integration into the convergence procedure of GENET. We compare the eeciency of the GENET using lazy arc consistency against GENET, both alone and using a full arc consistency preprocessing step, on a number of hard or large instances of binary CSP's. GENET with lazy arc consistency betters the original GENET on instances of binary CSP's which are not arc consistent in their original formulation, and does not suuer the overhead of full arc consistency for problems whose original formulation is arc consistent.