Pattern-Database Heuristics for Partially Observable Nondeterministic Planning

Heuristic search is the dominant approach to classical planning. However, many realistic problems violate classical assumptions such as determinism of action outcomes or full observability. In this paper, we investigate how – and how successfully – a particular classical technique, namely informed search using an abstraction heuristic, can be transferred to nondeterministic planning under partial observability. Specifically, we explore pattern-database heuristics with automatically generated patterns in the context of informed progression search for strong cyclic planning under partial observability. To that end, we discuss projections and how belief states can be heuristically assessed either directly or by going back to the contained world states, and empirically evaluate the resulting heuristics internally and compared to a delete-relaxation and a blind approach. From our experiments we can conclude that in terms of coverage and guidance, it is preferable to represent both nondeterminism and partial observability in the abstraction (instead of relaxing them), and that the resulting abstraction heuristics significantly outperform both blind search and a delete-relaxation approach where nondeterminism and partial observability are also relaxed.