On the Pathological Search Behavior of Distributed Greedy Best-First Search

Book One

Volume

Proceedings of the International Conference on Automated Planning and Scheduling, 29

Main Track

Although A* search can be efficiently parallelized using methods such as Hash-Distributed A* (HDA*), distributed parallelization of Greedy Best First Search (GBFS), a suboptimal search which often finds solutions much faster than A*, has received little attention. We show that surprisingly, HDGBFS, an adaptation of HDA* to GBFS, often performs significantly worse than sequential GBFS. We analyze and explain this performance degradation, and propose a novel method for distributed parallelization of GBFS, which significantly outperforms HDGBFS.

10.1609/icaps.v29i1.3485