Predicting Chaotic Time Series by Reinforcement Learning

Although a large number of researches have been carried out into the analysis of nonlinear phenomena, little is reported about using reinforcement learning, which is widely used in artificial intelligent, intelligent control, and other fields. Here, we consider the problem of chaotic time series using a self-organized fuzzy neural network and reinforcement learning, in particular, a learning algorithm called Stochastic Gradient Ascent(SGA). The proposed fuzzy neural network is similar to a radial basis function network(RBFN), but has self-organization ability dealing with its dynamical inputs,and provides stochastic outputs. The outputs are values of predicted time series, which called actions in reinforcement learning. After feeding some training data of chaotic time series to the initial frame of system, the structure and synaptic weights will be organized, and the predictor begins to provide correct dynamics of time series. Applying our proposed method to the Lorenz system, we obtained a high accuracy estimation of short-term prediction, and a reasonable result of long-term prediction.