Reinforcement learning using a recurrent neural network

Reinforcement learning methods that do not take into account previous states cannot deal with domains which have perceptually indistinguishable states that require different actions. This paper presents a neural network approach to the problem that uses William and Zipser´s RTRL (1989) network to incorporate temporal information. In addition, an off-line technique to speed up learning is discussed. The techniques have been successfully applied to a simple navigation task