Boosting Local Search with Artificial Ants

Abstract

Incomplete approaches for solving CSPs are usually based on local search — or neighborhood search—techniques [4]:the idea is to start from an inconsistent complete assignment of values to the variables, and then gradually and iteratively repair it by hanging some variable-value assignments, preferably towards better ones. One of the main problems with local search is that it may get stuck in local optima,i.e., complete assignments that cannot be locally improved by hanging one conflicting variable/value assignment, and that are not globally optimal. Therefore, local search has been combined with different meta-heuristics in order to help it escape from local optima, e.g., simulated annealing or tabu search [2]. Local search has proved to be efective and efficient to solve very large CSPs. However,like complete search, it often has more dificulties in solving problems that are within the phase transition region —where the solvable probability is around 50%. Indeed,before the phase transition region, problems are weakly constrained and have many solutions so that local search can usually easily find one. On the other side, beyond the phase transition region, problems are hardly constrained and only have few solutions, but they also have mu h less local optima so that local search can more easily reach a solution without being trapped in local optima [7]. Between these two -“easy” regions, search space landscapes of problems contain more local minima so that local search is more often trapped in these local minima and, even when using some meta-heuristics for escaping from them, local search often “walks” from a local minimum to another without finding a global minimum.