Explaining the Space of Plans through Plan-Property Dependencies

Anonymous

Abstract: A key problem in explainable AI planning is to elucidate decision rationales. User questions in this context are often contrastive, taking the form “Why do A rather than B?”. Answering such a question requires a statement about the space of possible plans. We propose to do so through plan-property dependencies, where plan properties are Boolean properties of plans the user is interested in, and dependencies are entailment relations in plan space. The answer to the above question then consists of those properties C entailed by B. We introduce a formal framework for such dependency analysis. We instantiate and operationalize that framework for the case of dependencies between goals in oversubscription planning. More powerful plan properties can be compiled into that special case. We show experimentally that, in a variety of benchmarks, the suggested analyses can be feasible and produce compact answers for human inspection.