Applying Anytime Heuristic Search to Cost-Optimal HTN Planning

Abstract

This paper presents a framework for cost-optimal Hierarchical Task Network (HTN) planning. The framework includes an optimal algorithm combining a branch-and-bound with a heuristic search, which can also be used as a near-optimal algorithm given a time limit. It also includes different heuristics based on weighted cost estimations and different decomposition strategies. The different elements from this framework are empirically evaluated on three planning domains, one of which is modeling a First-Person Shooter game. The empirical results establish the superiority on some domains of a decomposition strategy that prioritizes the most abstract tasks. They also highlight that the best heuristic formulation for the three domains is computed from linear combinations of optimistic and pessimistic cost estimations.