Action Schema Networks for Numeric Planning

Afifa Tariq, Richard Valenzano, Mikhail Soutchanski

Keywords: Numeric Planning, Action Schema Network, Neural Network, ENHSP, Generalized Planning

TL;DR: Solving numeric planning problems using Action Schema Networks.

Abstract: Planning is the fundamental ability of an intelligent agent to reason about what decisions it should make in a given environment to achieve a particular set of goals. Generalized planning is the task of finding a generalized policy that applies to a set of planning instances that share a standard model. Action Schema Networks (ASNets) is an approach to find generalized policies for classical planning problems. In this paper, we extend ASNet to work with numeric planning problems. We use a technique to propositionalize numeric variables, which converts them from infinite ranges to a finite domain, and update the training procedure to use exploration in order to increase the diversity of states encountered. We also use a non-generalized numeric planner, Expressive Numeric Heuristic Search Planner (ENHSP), to teach ASNet to solve numeric planning problems by learning to mimic the actions chosen by ENHSP for problem instances. ASNet finds a generalized policy and weights after training, allowing it to share these to solve unseen problem instances of the same domain. We analyze our approach through an extensive experimental study aimed at evaluating the performance of ASNet on several numeric planning domains. The results show that our numeric ASNet can effectively solve problems in many numeric planning domains.