Distributed Unidirectional and Bidirectional Heuristic Search: Algorithm Design and Empirical Assessment

Abstract

Since its introduction three decades ago, bidirectional heuristic search did not deliver the expected performance improvement over unidirectional search methods. The problem of search frontiers passing each other is a widely accepted conjecture led to amendments to steer the search using computationally demanding heuristics. The computation cost associated with front-to-front evaluations crippled further investigation and hence bidirectional search was long neglected. However, recent findings demonstrate that the initial conjecture is wrong since the major search effort is spent after the frontiers have already met [7]. In this paper we reconsider bidirectional search by proposing a new generic approach based on cluster computing. The proposed approach is then evaluated and compared with its unidirectional counterparts. The obtained results reveal that cluster computing is a viable approach for distributed heuristic search. Particularly, clustered bidirectional search is capable of solving problems beyond unidirectional search capabilities and in the same time outperforms unidirectional approaches in terms of memory space and execution time.