When do discounted-optimal policies also optimize the gain?
arXiv:2304.08048
Abstract
In this technical note, we establish an upper-bound on the threshold on the discount factor starting from which all discounted-optimal deterministic policies are gain-optimal, that we prove to be tight on an example. To address computability issues of that theoretical threshold, we provide a weaker bound which is tractable on ergodic MDPs in polynomial time.