Planning Graph Heuristics for Incomplete and Non-Deterministic Domains

This doctoral work centers on developing domain-independent heuristics for planning problems characterized by state incompleteness and action non-determinism. The key means by which the heuristics are computed is through planning graph analysis. The approach has been used to construct conformant and contingent plans in two different search techniques. The initial focus of the research was investigating useful single and multiple planning graph heuristic measures to estimate the conformant distance between belief states. The work showed that multiple planning graphs can accurately estimate belief state distance, but are costly. A subsequent contribution reduces the cost of multiple planning graph heuristics but maintains the accuracy through using a Labelled Uncertainty Graph (LUG). The LUG uses a single planning graph where graph elements are labelled with propositional formulas to indicate the states of the projected belief state for which the element is reachable. This innovation has proven to place the conformant planner CAltAlt in competition with current techniques in conformant planning. More recently, the LUG is showing its worth in deriving heuristics for a contingent planner PBSP . This abstract will describe the steps that I have taken in completing the aforementioned work and outline future prospects for my thesis.