PROCESS OF LEARNING DISCRETE DYNAMICAL-SYSTEMS BY RECURRENT NEURAL NETWORKS

This paper considers the learning of discrete dynamical systems using recurrent neural networks. The discussion is based on the theory of th e probabilistic descent method, and the learning algorithms are compar ed by numerical experiment. In the discussion based on the theory of t he probabilistic descent method, it is shown that, from the viewpoint of the learning speed in the early stage of the learning, is equivalen t to the backpropagation method with a large learning constant. For th e case where the variable is not constrained to the value of the teach er signal and a chaotic time series with a large Lyapunov exponent is to be learned, it is found that the effect of the recurrent connection s is not manifest at the early stage of the learning but the learning is accelerated with the progress of the learning by the fluctuation ca used by the chaos.