Reward Biased Maximum Likelihood Estimation for Learning in Constrained MDPs

TL;DR

This paper introduces the RBMLE algorithm for learning optimal policies in constrained Markov Decision Processes and analyzes its learning regret to evaluate performance.

Contribution

It proposes the Reward Biased Maximum Likelihood Estimation (RBMLE) method specifically for CMDPs and provides theoretical analysis of its regret bounds.

Findings

RBMLE effectively learns optimal policies in CMDPs.

Theoretical regret bounds for RBMLE are established.

The approach advances learning in constrained decision-making environments.

Abstract

We use the Reward Biased Maximum Likelihood Estimation (RBMLE) algorithm to learn optimal policies for constrained Markov Decision Processes (CMDPs). We analyze the learning regrets of RBMLE.