Online Reinforcement Learning for Real-Time Exploration in Continuous State and Action Markov Decision Processes

Abstract: This paper presents a new method to learn online policies in continuous state, continuous action, model-free Markov decision processes, with two properties that are crucial for practical applications. First, the policies are implementable with a very low computational cost: once the policy is computed, the action corresponding to a given state is obtained in logarithmic time with respect to the number of samples used. Second, our method is versatile: it does not rely on any a priori knowledge of the structure of optimal policies. We build upon the Fitted Q-iteration algorithm which represents the $Q$-value as the average of several regression trees. Our algorithm, the Fitted Policy Forest algorithm (FPF), computes a regression forest representing the Q-value and transforms it into a single tree representing the policy, while keeping control on the size of the policy using resampling and leaf merging. We introduce an adaptation of Multi-Resolution Exploration (MRE) which is particularly suited to FPF. We assess the performance of FPF on three classical benchmarks for reinforcement learning: the "Inverted Pendulum", the "Double Integrator" and "Car on the Hill" and show that FPF equals or outperforms other algorithms, although these algorithms rely on the use of particular representations of the policies, especially chosen in order to fit each of the three problems. Finally, we exhibit that the combination of FPF and MRE allows to find nearly optimal solutions in problems where $\epsilon$-greedy approaches would fail.