A Hybrid Genetic Algorithm For System Identification

The problem considered is that of approximating an unknown system (plant) by learning the coefficients of a linear model from data collected on the plant performance, using the criterion of minimizing the sum of squared errors between predicted and actual performance. It is well known that the error surface is not unimodal, and that the estimated model may converge to a locally optimal solution. Furthermore, the solution may be unstable. In this paper a genetic algorithm (GA) is described which attempts to deal with each of these problems. The question of stability is handled by encoding the solution vector in terms of the radii and angles of the poles and zeros directly. The problem of premature convergence to a local optimum is tackled by hybridizing a dynamically varying GA with the classical Golden Section method of search. Results will be presented which show that this approach is capable of very high accuracy for an artificial test problem, and that it also outperforms ordinary l..