Regret Bounds for Risk-Sensitive Reinforcement Learning
Machine Learning
2022-10-12 v1
Abstract
In safety-critical applications of reinforcement learning such as healthcare and robotics, it is often desirable to optimize risk-sensitive objectives that account for tail outcomes rather than expected reward. We prove the first regret bounds for reinforcement learning under a general class of risk-sensitive objectives including the popular CVaR objective. Our theory is based on a novel characterization of the CVaR objective as well as a novel optimistic MDP construction.
Cite
@article{arxiv.2210.05650,
title = {Regret Bounds for Risk-Sensitive Reinforcement Learning},
author = {O. Bastani and Y. J. Ma and E. Shen and W. Xu},
journal= {arXiv preprint arXiv:2210.05650},
year = {2022}
}