Characterizing Search Spaces for Tabu Search

A large number of heuristic search algorithms are available for function optimization. Each of these heuristics, e.g., simulated annealing, genetic algorithms, tabu search, etc., has been shown to be effective at finding good solutions efficiently. However, little work has been directed at determining what are the important problem characteristics for which one algorithm is more efficient than the others. By examining two problems, the location–allocation problem and the quadratic assignment problem, characteristics of successful tabu search are illustrated. A tabu search for the location–allocation problem is described and implemented. The results of this tabu search are compared against a genetic algorithm. For the quadratic assignment problem, tabu search has been shown more effective than genetic algorithms; however, for the location–allocation problem, the genetic algorithm finds better solutions more efficiently than tabu search. To investigate what characteristics of the location–allocation problem makes it less amenable to tabu search, a comparison between the location–allocation problem and the quadratic assignment problem is performed. A comparison of the problem characteristics reveals that the location–allocation problem has very large basins of attraction around a few local optima. For tabu search to escape these minima requires a large number of iterations. Finally, a combination of both tabu search and genetic algorithms is presented for the location-allocation problem,where regions around genetically determined sample points are marked as tabu. This combination compares favorably to the genetic algorithm in terms of increased computational efficiency.