Eyal Shlomo Shimony

Recent research on bidirectional search describes anomalies, or cases in which improved heuristics lead to more node expansions. Aiming to avoid such anomalies, this paper characterizes desirable properties for bidirectional search algorithms, and studies conditions for obtaining these properties. The characterization is based on a recently developed theory for bidirectional search, which has formulated conditions on pairs of nodes such that at least one node from every pair meeting these conditions must be expanded. Moreover, based on this must-expand-pairs theory, we introduce a method for enhancing heuristics by propagating lower bounds (lb-propagation) between frontiers. This lb-propagation can bestow the desirable properties on some existing algorithms (e.g., the MM family) while avoiding the above anomaly altogether. Empirical results show that lb-propagation reduces the number of node expansions in many cases.