Local Search Topology: Implications for Planner Performance

Hoffmann’s topological analysis of the h and h F state spaces explained why Fast Forward dominated the performance on early planning domains (Hoffmann 2004). His taxonomy segmented domains according to the presence/size of local minima and dead-end class. Surprisingly, the taxonomy has not been used to explain the performance of other planners that use similar heuristics. In this paper, we extend this analysis in several ways: 1) We apply the taxonomy to 10 heuristic search (including FF-2.3) and 17 non-heuristic search planners to determine the extent to which it explains planner performance. 2) We test model scalability by examining a subset of challenging problems to determine the extent to which the topology explains performance. We conclude that topological analysis is a valuable tool in explaining the conditions under which h is favored. Similar analyses of newer domains and planners could benefit the community and may lead to more informed application, extensions, or simplifications of either the heuristics or the planners that use them.