On-line profit sharing works efficiently

Abstract

Reinforcement learning constructs knowledge containing state-to-action decision rules from agent's experiences. Most of reinforcement learning methods are action-value estimation methods which estimate the true values of state-action pairs and derive the optimal policy from the value estimates. However, these methods have a serious drawback that they stray when the values for the "opposite" actions, such as moving left and moving right, are equal. This paper describes the basic mechanism of on-line profit-sharing (OnPS) which is an action-preference learning method. The main contribution of this paper is to show the equivalence of off-line and on-line in profit sharing. We also show a typical benchmark example for comparison between OnPS and Q-learning.