Learning measures of progress for planning domains

摘要: We study an approach to learning heuristics for planning domains from example solutions. There has been little work on the types of used in deterministic and stochastic competitions. Perhaps one reason this is challenge providing a compact heuristic language that facilitates learning. Here we introduce new representation based lists set expressions described using taxonomic syntax. Next, review idea measure progress (parmar 2002), which any guaranteed be improvable at every state. take finding as our goal, describe simple algorithm purpose. evaluate across range planning-competition domains. The results show often greedily following learned highly effective. also can combined with rule-based policies, producing still stronger results.