Improving probabilistic planner with generalized neural policy

Abstract

In this paper, we propose a method of combining generalized neural network-based policy and the existing search-based planner when solving probabilistic planning problems with large states and action spaces. The policy based on the graph neural network structure is learned by mimicking the existing search-based planner in small-sized planning problems, and the learned policy guides the search direction of the planner in large planning problems that the original planner cannot solve. Comparing the proposed framework with the original planner and policy learning based on reinforcement learning, the proposed methodology has been shown to help improve the performance of the planner. Also, our work can be used as a baseline in the field of automatic planning based on deep learning.