On Metaheuristic “Failure Modes”: A Case Study in Tabu Search for Job-Shop Scheduling

It is widely recognized that long-term memory mechanisms such as reintensification and path relinking can significantly improve the performance of ”core” metaheuristics such as tabu search or iterated local search. Case studies abound, e.g., see [2]. These algorithms are commonly designed around a set or pool of elite, high-quality solutions from which intensification and diversification schemes are initiated. However, it is equally well known that the details of how the elite pool is managed has a major impact on performance [3]. Considering that the relationships between elite pool maintenance, long-term memory, and search space structure are poorly understood, the fact that the development of effective long-term memory and pool maintenance mechanisms is more art than science is not surprising. In this paper, we analyze the relationship between pool maintenance schemes, long-term memory mechanisms, and search space structure, with the goal of placing metaheuristic design on a more concrete foundation.