Learning Time Series by Simple Recurrent Network with Short-term Memory

摘要

Recently, various neural networks are applied to learning time series. Elman's network is a major method in learning time series. However, it has the following problems; (1) learning time is comparatively long, (2) there has not been much discussion on generalization ability, and (3) convergence success rate is insufficient. In this paper, Elman's network with short-term memory is proposed in order to simplify time series learning, which adds past histories to input layer directly. Moreover, cross entropy method is employed which has been applied to static and discrete valued problems. Our method is applied to a real time series; the number of chickenpox cases observed every month in New York City. It is shown in a computer simulation that learning speed attains 30 times faster, generalization error is reduced to 70%, and convergence success rate rises up to 100% all together.