Use of genetic programming in the identification of rational model structures

Abstract

This paper demonstrates how genetic programming can be used for solving problems in the field of non-linear system identification of rational models. By using a two-tree structure rather than introducing the division operator in the function set, this genetic programming approach is able to determine the “true” model structure of the system under investigation. However, unlike use of the polynomial, which is linear in the parameters, use of rational model is non-linear in the parameters and thus noise terms cannot be estimated properly. By means of a second optimisation process (real-coded GA) which has the aim of tunning the coefficients to the “true” values, these parameters are then correctly computed. This approach is based upon the well known NARMAX model representation, widely used in non-linear system identification.