IDWalk: A candidate list strategy with a simple diversification device

Abstract

This paper presents a new optimization metaheuristic called IDWalk (Intensification/Diversification Walk) that offers advantages for combining simplicity with effectiveness. In addition to the number S of moves, IDWalk uses only one parameter Max which is the maximum number of candidate neighbors studied in every move. This candidate list strategy manages the Max candidates so as to obtain a good tradeoff between intensification and diversification. A procedure has also been designed to tune the parameters automatically. We made experiments on several hard combinatorial optimization problems, and IDWalk compares favorably with correspondingly simple instances of leading metaheuristics, notably tabu search, simulated annealing and Metropolis. Thus, among algorithmic variants that are designed to be easy to program and implement, IDWalk has the potential to become an interesting alternative to such recognized approaches. Our automatic tuning tool has also allowed us to compare several variants of IDWalk and tabu search to analyze which devices (parameters) have the greatest impact on the computation time. A surprising result shows that the specific diversification mechanism embedded in IDWalk is very significant, which motivates examination of additional instances in this new class of "dynamic" candidate list strategies. © Springer-Verlag Berlin Heidelberg 2004.