An embedded sigmoidal neural network for modeling of nonlinear systems

This paper discusses the problem of applying sigmoidal neural network to prediction and control of nonlinear dynamical systems. Instead of directly using neural networks as nonlinear models, we first develop a shield based on application specific knowledge, and then embed a sigmoidal neural network model in the shield. An embedded sigmoidal neural network model obtained in this way not only has a structure favorable for certain applications such as controller design, but also has useful interpretation on part of model parameters. Corresponding to the meaningful part and the meaningless part of model parameters, a hierarchical training algorithm consisting of two learning loops is introduced to train the model, which has good performance on solving local minimum problems. The usefulness of the proposed prediction model is demonstrated by applying it to prediction and control of a simulated nonlinear system