Beyond Average Return in Markov Decision Processes
This advances the theory of MDPs for risk-aware decision-making, though it appears to be a theoretical extension rather than a practical breakthrough.
The paper investigates which reward functionals can be exactly computed and optimized in Markov Decision Processes, proving that only generalized means are optimizable exactly even in Distributional Reinforcement Learning, while providing error bounds for approximate evaluation of other functionals.
What are the functionals of the reward that can be computed and optimized exactly in Markov Decision Processes?In the finite-horizon, undiscounted setting, Dynamic Programming (DP) can only handle these operations efficiently for certain classes of statistics. We summarize the characterization of these classes for policy evaluation, and give a new answer for the planning problem. Interestingly, we prove that only generalized means can be optimized exactly, even in the more general framework of Distributional Reinforcement Learning (DistRL).DistRL permits, however, to evaluate other functionals approximately. We provide error bounds on the resulting estimators, and discuss the potential of this approach as well as its limitations.These results contribute to advancing the theory of Markov Decision Processes by examining overall characteristics of the return, and particularly risk-conscious strategies.