An average-reward reinforcement learning algorithm based on Schweitzer's Transformation

In this paper, we propose a relative value iteration reinforcement learning(RVI-RL) algorithm based on Schweitzer's Transformation for Markov decision processes (MDP) with average reward. An equivalent average reward optimality equation and a new form of action-value function are presented via Schweitzer's Transformation. Then, combined with the theory of relative value iteration, this RVI-RL algorithm doesn't only omit the estimation of the average reward in the learning, but also improves the convergence rate. Finally, a simulation experiment for the navigation of autonomous mobile robot is considered, which illustrates the effectiveness and applicability of the algorithm.