Recurrent neural network system using probability graph model optimization

摘要

We propose a novel recurrent neural network model based on a combination of the echo state network (ESN) and the dynamic Bayesian network (DBN). Our contribution includes the following: (1) A new graph-based echo state network (GESN) model is presented for nonlinear system modeling. The GESN consists of four layers: an input layer, reservoir layer, graph model optimization layer and output layer. Unlike the original ESN model, the GESN model uses probability graph inference to optimize reservoir states, and the optimized GESN is more robust. (2) A novel DBN graph is proposed for describing random probability signals in the GESN. Simultaneously, an inference algorithm of the DBN is deduced according to Bayesian rules and probability graph theory. Finally, GESN performance is tested using the Mackey–Glass time series and laser time series data forecasting The simulations and comparison results show that the proposed model is promising.