Mutual Learning: Part II --Reinforcement Learning
Kumpati S. Narendra, Snehasis Mukhopadhyay · 2020
The concept of "Mutual Learning" was introduced by the authors in Part I of this paper which was presented at the 2019 ACC. This is the second of the series of papers the authors propose to write on this subject. The principal question addressed in all of them concerns the process by which two agents "learn" from each other. More specifically, the question is how two (or more) agents should share their information to improve their performance.In Part I, the concept of mutual learning was introduced and discussed briefly in the context of two deterministic learning automata learning from each other in a static random environment. In this paper, we first provide some reasons why the concept can become complex even in such simple situations, and propose some (weak) necessary conditions for the problem to be well defined. Following this, we consider similar questions which arise when two agents use reinforcement learning in both static and dynamic environments. In particular, in the latter category, mutual learning in Markov Decision Processes (MDP) with finite states is discussed. Simulation results are presented wherever appropriate to complement the theoretical discussions.