Solving planning problems with deep reinforcement learning and tree search

Abstract

Deep reinforcement learning methods are capable of learning complex heuristics starting with no prior knowledge, but struggle in environments where the learning signal is sparse. In contrast, planning methods can discover the optimal path to a goal in the absence of external rewards, but often require a hand-crafted heuristic function to be effective. In this thesis, we describe a model-based reinforcement learning method that bridges the middle ground between these two approaches. When evaluated on the complex domain of Sokoban, the model-based method was found to be more performant, stable and sample-efficient than a model-free baseline.