抄録
Due to the orbital instability, the non-periodicity and the long-term unpredictability of chaotic time series, mathematical modeling of different chaotic phenomena is computationally heavy and difficult to find values of their numbers of parameters. So, many of researches of soft-computing methods have shown us their satisfactory effects, especially neural networks.Meanwhile, there are some hypothesize assert chaotic time series that we deal with are not some of determining chaos but artifact produced by stochastic process. And there are some researches have shown us statistic estimations(i.e., surrogate data) make important roles in nonlinear dynamic analysis. Here, we consider a neural network whose outputs have stochastic feature to predict chaotic time series. The neural networks has fuzzy inference ability, which classifying inputs of time series, and self-organization ability to control the structure of predictor itself. And its neurons in output-layer are some variables of stochastic function. Especially, we employ a kind of reinforcement learning named Stochastic Gradient Ascent(SGA) to train prediction system's synaptic weights. To investigate our proposed method, an application to the Lorenz system were executed. The result of short-term predictions showed high estimation accuracy.
