Heuristic Learning in Domain-Independent Planning: Theoretical Analysis and Experimental Evaluation

Automated planning deals with the problem of finding a sequence of actions leading from a given state to a desired state. The state-of-the-art automated planning techniques exploit informed forward search guided by a heuristic which is used to estimate a distance from a state to a goal state. In this paper, we present a technique to automatically construct an efficient heuristic for a given domain. The proposed approach is based on training a deep neural network using a set of solved planning problems as training data. We use a novel way of extracting features for states developed specifically for planning applications. Our experiments show that the technique is competitive with state-of-the-art domain-independent heuristic. We also introduce a theoretical framework to formally analyze behaviour of learned heuristics. We state and prove several theorems that establish bounds on the worst-case performance of learned heuristics.