The Analysis of the Feasibility of Optimizing Elman Neural Network

Abstract

Elman Neural Network (ENN) is a widely used typical feedback-type neural network model. Considered to represent an optimization of Back Propagation Neural Network (BPNN), while ENN does perform better than BPNN in some certain properties, is also inherits some defects of BPNN. For example, both BPNN and ENN are highly nonlinear neural network models with poor generation capability. Such deficiency determine that these two models perform worse when dealing with some prediction problems involving linear influencing factors. To address this, the rest of this paper presents an optimized model based on ENN (ENN-DIOC) by introducing direct input-to-output structure into ENN, and analyzes the feasibility of this optimization. Combined with the experiments completed by predecessors, it can be proved that it’s feasible for ENN-DIOC to improve the prediction accuracy and add linear factors to basic ENN.