Abstract
When humans are given a policy to execute, there can be pol-icy execution errors and deviations in execution if there is un-certainty in identifying a state. So an algorithm that computesa policy for a human to execute ought to consider these effectsin its computations. An optimal MDP policy that is poorly ex-ecuted (because of a human agent) maybe much worse thananother policy that is executed with fewer errors. In this pa-per, we consider the problems of erroneous execution and ex-ecution delay when computing policies for a human agent thatwould act in a setting modeled by a Markov Decision Process(MDP). We present a framework to model the likelihood ofpolicy execution errors and likelihood of non-policy actionslike inaction (delays) due to state uncertainty. This is followedby a hill climbing algorithm to search for good policies thataccount for these errors. We then use the best policy found byhill climbing with a branch and bound algorithm to find theoptimal policy. We show experimental results in a Gridworlddomain and analyze the performance of the two algorithms.We also present human studies that verify if our assumptionson policy execution by humans under state-aliasing are rea-sonable.
Abstract (translated)
URL
https://arxiv.org/abs/2109.07436