Optimistic Agents are Asymptotically Optimal

TL;DR

This paper introduces a class of reinforcement learning agents that use optimism to achieve asymptotic optimality across various environment classes, providing finite error bounds in deterministic cases.

Contribution

It presents a generic optimistic approach for asymptotically optimal reinforcement learning applicable to broad environment classes, with finite bounds in deterministic scenarios.

Findings

Achieves asymptotic optimality in diverse environments.

Provides finite error bounds for deterministic environments.

Demonstrates the effectiveness of optimism in reinforcement learning.

Abstract

We use optimism to introduce generic asymptotically optimal reinforcement learning agents. They achieve, with an arbitrary finite or compact class of environments, asymptotically optimal behavior. Furthermore, in the finite deterministic case we provide finite error bounds.