Bidirectional A* search with additive approximation bounds

Abstract

In this paper, we present new theoretical and experimental results for bidirectional A* search. Unlike most previous research on this topic, our results do not require assumptions of either consistent or balanced heuristic functions for the search. Our theoretical work examines new results on the worst-case number of node expansions for inconsistent heuristic functions with bounded estimation errors. Additionally, we consider several alternative termination criteria in order to more quickly terminate the bidirectional search, and we provide worst-case approximation bounds for our suggested criteria. We prove that our approximation bounds are purely additive in nature (a general improvement over previous multiplicative approximations). Experimental evidence on large-scale road networks suggests that the errors introduced are truly quite negligible in practice, while the performance gains are significant. Copyright © 2012, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.