Effectiveness of a recurrent neural network in emergence of discrete decision making through reinforcement learning

Abstract

In a discrete decision making task, using a neural network suffers from the problem of discrete decision making. On the other hand, using a lookup table suffers from the problem in generalization and the curse of dimensionality. To overcome this problem, simple localized inputs in neural network are used. Furthermore, this paper focus on examining whether by utilizing the internal dynamics in RNN, quick decision making can be obtained through learning or not. In this paper, it is shown that a robot learned to make a discrete decision making even though no special technique other than a localized inputs in RNN through RL was utilized.