Learning Optimal Policies in Markov Decision Processes with Value Function Discovery

摘要

htmlabstractIn this paper we describe recent progress in our work on\udValue Function Discovery (VFD), a novel method for discovery\udof value functions for Markov Decision Processes (MDPs).\udIn a previous paper we described how VFD discovers algebraic\uddescriptions of value functions (and the corresponding\udpolicies) using ideas from the Evolutionary Algorithm field.\udA special feature of VFD is that the descriptions include the\udmodel parameters of the MDP. We extend that work and\udshow how additional information about the structure of the\udMDP can be included in VFD. This alternative use of VFD\udstill yields near-optimal policies, and is much faster. Besides\udincreased performance and improved run times, this\udapproach illustrates that VFD is not restricted to learning\udvalue functions and can be applied more generally.