Algorithms for various cost criteria in Reinforcement Learning

Abstract

In this thesis we will look at various Reinforcement Learning algorithms. We will look at algorithms for various cost criteria or reward objectives namely Finite Horizon, Discounted Cost, Risk-Sensitive Cost. For Finite Horizon and Risk-Sensitive Cost we derive the policy gradient, and for Discounted Cost we propose a new algorithm called Critic-Actor. We analyze and prove the convergence for all the proposed algorithms. We also analyze the empirical performance of our algorithms through numerical experiments.