Combining reinforcement learning with symbolic planning

One of the major difficulties in applying Q-learning to real-world domains is the sharp increase in the number of learning steps required to converge towards an optimal policy as the size of the state space is increased. In this paper we propose a method, PLANQ-learning, that couples a Q-learner with a STRIPS planner. The planner shapes the reward function, and thus guides the Q-learner quickly to the optimal policy. We demonstrate empirically that this combination of high-level reasoning and low-level learning displays significant improvements in scaling-up behaviour as the state-space grows larger, compared to both standard Q-learning and hierarchical Q-learning methods.