Action Schema Networks for Numerical Planning

Planning is the fundamental ability of an intelligent agent to reason about what decisions it should make in a given environment to achieve a certain set of goals. Action Schema Networks (ASNet) find generalized policies for classical planning problems. In this thesis, we extend ASNet to work with numerical planning problems. We present the technique to propositionalize numerical variables to convert them from continuous infinite ranges to a finite domain. We use a non-generalized numerical planner, ENHSP, to teach ASNet to solve numerical planning problems by learning to mimic the actions chosen by this teacher planner for problem instances. We have optimized the training algorithm with action generation techniques, objective functions, and evaluation strategies. ASNet finds a generalized policy and weights after training which allows it to share these policy and weights 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 numerical planning domains. The results show that our numerical ASNet can effectively handle many numerical planning domains and significantly outperforms the baseline planner in terms of execution time. This work is a first step to applying neural networks to numerical generalized planning.