Learning in DepthFirst Search A Unified Approach to Heuristic Search in Deterministic, Non.pdfVIP

Learning in Depth-First Search: A Unified Approach to Heuristic Search in Deterministic, Non-Deterministic, Probabilistic, and Game Tree Settings ´ Blai Bonet and Hector Geffner Abstract depth-first searches, with learning in the sense (Korf 1990) and (Barto, Bradtke, Singh 1995). In each iteration, if Dynamic Programming provides a convenient and unified there is a solution with cost not exceeding a lower bound, framework for studying many state models used in AI but no then the solution is found and reported, else the process algorithms for handling large spaces. Heuristic-search meth- ods, on the other hand, can handle large spaces but lack a restarts with the lower bound and the value function updated. common foundation. In this work, we combine the bene- LDFS reduces to IDA * with Transposition Tables (Reinefeld fits of a general dynamic programming formulation with the Marsland 1994) over deterministic models, but solves also power of heuristic-search techniques for developing an al- non-deterministic, probabilistic, and game tree models, over gorithmic framework, that we call Learning in Depth-First which a slight variation reduces to the state-of-the-art MTD Search, that aims to be both general and effective. The basic algorithm (Plaat et al. 1996). LDFS algorithm searches for solutions by combining itera- The LDFS framework makes explicit and generalizes two tive, bounded depth-first searches, with learning in the sense