Neural modules: networks with constrained architectures for nonlinear function identification

Abstract

The aim of this work is the design of a class of neural networks for nonlinear function identification: the so-called neural modules. A neural module is a neural network with an internal structure specially designed to be able to learn and mimic the behaviour of a certain class of dynamic systems. Neural networks are abstract models well suited for approximating nonlinear functions. The training cost and the structural complexity of neural networks can be drastically reduced if a-priori knowledge of the function to be learned is internally incorporated in the form of structural constraints. The resulting neural network has less parameters than a conventional one, much faster learning convergence and it can provide meaningful information about the learned nonlinear function. This paper describes the design of a useful set of neural modules for system identification and gives general guidelines for the design of neural modules. The resulting networks are evaluated and their use on general systems identification is pointed out.