Identificação de sistemas a partir de modelos NARX com seleção de regressores por programação genética

Abstract

For the past few centuries, humankind has searched for ways to improve its capacity to mathematically model unknown systems. Modern system identification techniques arise as a response to the increasing complexity of the systems that need to be modeled. Identification can be achieved through the use of NARX, Nonlinear AutoRegressive with eXogenous input, models. This model structure uses regressors and parameters in order to represent the behavior of systems with nonlinear characteristics. Methods of determining regressors and their combinations often depend on low efficiency computer processes. As an alternative, this paper employs an evolutionary computation technique named Genetic Programming, GP. In addition, the aplication of a Least Squares Method, LSM, operator was defined, in assistance to the evolutionary process. The algorithm was utilized to identify two systems: an industrial fluid flow plant and a magnetic levitator. In spite of an increase in the mean processing time, the use of the LSM operator produced a near 70% reduction of the Mean Square Error, MSE, of the best individuals in relation to the results without the operator. With the obtained results, it is concluded that system identification through GP exhibited precise models for the proposed systems.