Research On Reservoir Computing Model And Time Series Prediction And Classification

In recent year, the Reservoir Computing model, represented by echo state networks(ESN), has become a new research focus. The basic feature of ESN is that it has large-scaleneurons (called as the "reservoir") which are randomly and sparsely connected. And only theweights of the connections from the reservoir to the readout (output) neurons need to beadapted with the least-square method. Such computing model revolutionizes the predictionaccuracy in chaotic time series. But there are still many problems in Reservoir Computingmodel. First of all, in the aera of chaotic time series forcasting, ESN interation forecast modelhas a better accuracy, but less robustness. And ESN direct forecast model has a betterrobustness, but less accuracy. How can we improve the forcasting model for both highaccuracy and robust? It has practical applicable meaning. Another problem is thatReservoir Computing model can’t apply in time series classification directly, which is a hotresearch subject in time series analysis. Can the ESN be extended from time series predictionto time series classification? This is a valuable research topic. For these two problems, thispaper proposes modular echo state network (MESN) and functional echo state network(FESN) separately, specified as follows:1) We propose modular echo state network model based on the idea of combinationbetween ESN direct forecast model and ensemble learning theory.2) According to ensemble learning theory and ESN learning algorithm, we deduce aglobal approximate optimal modular partition function. And then, we deduce a trainingalgorithm. Experimental results on Mackey-Glass and Lorenz chaotic systems show thatMESN maintains the same robustness as ESN direct forecast model, and has greatlyenhanced the accuracy.3) An functional echo state network model combining ESN and functional neurons ispresented in this paper, which can directly apply to time series classification.4) An efficient training algorithm for FESN is proposed, with method of pseudo-inversealgorithm and fourier expansion. At last, we have compared the performances of FESN withother popular approaches on the public data sets available through the UCR Time Series Data Mining Archive for both time series classification and clustering problems. The experimentalresults indicate that the FESN has a very high accuracy.