System Identification Using Neural Networks with Parametric Sigmoid Functions

Abstract

Nonlinear systems can be modeled by neural networks. However, choice of suitable network architecture is the most important problem. And “how to find the best activation function” is a persistent aspect of the architecture design. Here we have proposed a sigmoid function with one parameter which provides us not only the reduction of error bound but also the opportunity of obtaining better insight into the systems. The proposed function has the ability of recognizing linear and/or nonlinear parts of the system under study. After automatic training of this parameter along the weights, more information about the system will be available. Using this additional knowledge about structure of the system, one will be well equipped to attack control problems such as controller design using neural network model.