Planning in Incomplete Domains

Engineering complete planning domain descriptions is often very costly because of human-error or lack of domain knowledge. While many have studied knowledge acquisition, relatively few have studied the synthesis of plans when the domain model is incomplete (i.e., actions have incomplete preconditions or effects). Prior work has evaluated the correctness of plans synthesized by disregarding such incomplete features, but not how to synthesize plans by reasoning about the incompleteness. In this work, we describe several techniques for reasoning with the action incompleteness to make plans robust, as measured by the number of incomplete domain interpretations under which a plan succeeds. Among the techniques, we show that representing explanations of plan failure with prime implicants provides a natural approach to comparing plans by counting prime implicants (diagnoses) instead of domain interpretations (i.e., model counting) – leading to better scalability and higher quality plans. We present and empirically evaluate a forward heuristic search planner, called DeFAULT, that synthesizes plans by propagating information about faults due to incompleteness both within the state space and the relaxed planning space by using intuitions from model based diagnosis and assumption-based truth maintenance systems. We also provide a translation from incomplete planning domains to conformant probabilistic planning, where action incompleteness is represented by state incompleteness. We compare DeFAULTwith a control planner that uses the FF heuristic (measuring plan length and ignoring incompleteness), and with the conformant probabilistic planner POND. The results show that DeFAULT i) scales better than POND, ii) finds better solutions than a planner using the FF heuristic, iii) scales best and finds its best quality solutions when counting prime implicants rather than models.