Safe Distributional Reinforcement Learning

TL;DR

This paper introduces a formalized approach to safety in reinforcement learning using distributional RL, proposing a flexible model and an efficient safe policy optimization method validated on various domains.

Contribution

It presents a novel constrained distributional RL framework and extends safe policy optimization to improve safety during learning.

Findings

Efficient safe policy optimization algorithm developed.

Validated on artificial and real domains against state-of-the-art methods.

Flexible safety definitions including bounds, CVaR, variance, and bad state probability.

Abstract

Safety in reinforcement learning (RL) is a key property in both training and execution in many domains such as autonomous driving or finance. In this paper, we formalize it with a constrained RL formulation in the distributional RL setting. Our general model accepts various definitions of safety(e.g., bounds on expected performance, CVaR, variance, or probability of reaching bad states). To ensure safety during learning, we extend a safe policy optimization method to solve our problem. The distributional RL perspective leads to a more efficient algorithm while additionally catering for natural safe constraints. We empirically validate our propositions on artificial and real domains against appropriate state-of-the-art safe RL algorithms.