NeoRL: Efficient Exploration for Nonepisodic RL

TL;DR

NeoRL introduces an optimistic RL approach for nonepisodic nonlinear systems, learning from a single trajectory without resets, and achieves near-optimal regret bounds with empirical success.

Contribution

The paper presents NeoRL, a novel method for nonepisodic RL that provides the first regret bounds for continuous, single-trajectory learning in nonlinear systems.

Findings

NeoRL achieves the optimal average cost in experiments.

NeoRL incurs the least regret compared to baselines.

Regret bound of O(Γ_T √T) for nonlinear systems with Gaussian process dynamics.

Abstract

We study the problem of nonepisodic reinforcement learning (RL) for nonlinear dynamical systems, where the system dynamics are unknown and the RL agent has to learn from a single trajectory, i.e., without resets. We propose Nonepisodic Optimistic RL (NeoRL), an approach based on the principle of optimism in the face of uncertainty. NeoRL uses well-calibrated probabilistic models and plans optimistically w.r.t. the epistemic uncertainty about the unknown dynamics. Under continuity and bounded energy assumptions on the system, we provide a first-of-its-kind regret bound of $O(\Gamma_T \sqrt{T})$ for general nonlinear systems with Gaussian process dynamics. We compare NeoRL to other baselines on several deep RL environments and empirically demonstrate that NeoRL achieves the optimal average cost while incurring the least regret.