Computação evolucionária e máquinas de comitê na identificaçãode sistemas não-lineares

Abstract

In the last decades, Machine Learning, the research area that aims to study computer algorithms that extract information from data automatically, has grown in importance due to the development of computer capacity and therefore due to the increase of available information. The main challenge of learning algorithms is to improve generalization ability of estimators. In this context, evolutionary algorithms and committee machines (combination of more than one model) may be seen as competitive alternatives to solve this challenge. Thus, the identification of nonlinear systems, increasingly required in advanced control problems, can benefit from these alternatives. From this premise, this work aims at applying such techniques in identification problems. Looking at the problem of identification in an optimization perspective, two entities are of utmost importance: the prediction error and the simulation error. With the use of evolutionary algorithms, multi-objective or not,the role of these entities in the parameters estimation of nonlinear models is discussed. Among the obtained results, it could be emphasized the one that recommends the use of prediction error based criteria in equation error problems and the use of simulation error based criteria in output error problems (or measurement error), the latter being generally more robust. Although it is known that the use of prediction error based criterion in output error problems, without the proper settings (noise model), finds biased estimates, the novelty is that the simulationerror also finds biased estimates when applied to equation error problems. A new bi-objective approach was proposed using simulation error and the model static function error in gray-box identification, showing its effectiveness against the black-box identification and against prediction error approaches on a real problem. PWA hybrid systems, examples of committee machines, were also estimated by these entities (through the application of genetic algorithms) finding that the definition of each submodel partition can be performed by prediction error based criteria regardless the noise model. However, the estimation of the submodels parameters should be undertaken by the proposed algorithmcalled MQEP (extended and weighted least squares estimator) in output error problems to avoid bias. Finally, co-evolutionary algorithms and artificial immune systems were implemented to build committees of neural networks being possible to obtain good results in some benchmark regression problems. It was shown that the use of adiversity measure in the learning process is not advisable and that it is possible to find small committees automatically