A proposal of neural network architecture for nonlinear system modeling

Abstract

SUMMARYThis paper proposes new neural network architecture for nonlinear system modeling. The traditional modeling methods with neural network have the following problems: (1) difficulty in analyzing the internal representation, namely, the obtained values of the coupling weights, (2) no reproducibility due to the random scheme for weight initialization, (3) insufficient generalization ability for the input space in which no training sample exists. In order to overcome these deficiencies, the proposed method presents the following approaches. The first is the design of a sigmoid function with localized derivative. The second is a deterministic scheme for weight initialization. The third is an updating rule for weight parameters. Simulations were conducted based on several nonlinear systems with two inputs and one output. These results indicated small initial error, small modeling error, smooth convergence, and improvement of the difficulty in analyzing the internal representation.