The Automatic Creation of Memory - basedSearch Heuristics

Our goal is to automatically generate heuristics to guide state space search. Our heuristics are deened by distances in an abstract space that is automatically derived from the deenition of the original space. Such abstractions provide monotone heuristics for search algorithms such as A* and IDA*. To facilitate the automatic generation of heuristics we introduce a notation for describing states and operators, PSVN, that is simple but suuciently expressive to succinctly describe a wide range of state spaces. We then describe domain abstraction , a way of abstracting PSVN state spaces. With some restrictions, domain abstraction of PSVN descriptions gives the user ne control over the size of the abstract space. The heuristic corresponding to the abstract space is eeciently precomputed and stored in a lookup table, so that its use during search is very eecient. Three applications of automatically generated heuristics are presented. First, thousands of automatically generated heuristics are used to experimentally investigate the conjecture by Korf 14] that m t is a constant, where m is the size of a heuristic's lookup table and t is the number of nodes expanded when the heuristic is used to guide search. Second, a similar large-scale experiment is used to verify that the complexity analysis of search heuristics by Korf and Reid 15] can be used to rapidly and reliably choose the best among a given set of heuristics. Finally, we show that by using this analysis method to select the best among a large set of randomly generated heuristics, good heuristics for large, novel search spaces can be discovered quickly and automatically.