Reinforcement learning and neural reinforcement learning

摘要

In this paper, we address an under-represented class of learning algorithms in the study of connectionism: reinforcement learning. We first introduce these classic methods in a new formalism which highlights the particularities of implementations such as Q-Learning, Q-Learning with Hamming distance, Q-Learning with statistical clustering and Dyna-Q. We then present in this formalism a neural implementation of reinforcement which clearly points out the advantages and the disadvantages of each qpproach.