Abordagem de computação paralela para geração de bases de regras fuzzy em problemas de grande volume e de alta dimensionalidade dos dados usando algoritmos genéticos multiobjetivo de ordenação por não dominância

Abstract

The development of Rule-Based Fuzzy Systems (FRBS) brings with it the need to address issues related to the desirable balance between the conflicting objectives of accuracy (measure related to the system's ability to describe a given problem) and complexity (measure related to interpretability of the FRBS according to the number of rules and rule antecedents). This situation also occurs when considering problems where the input data have large volume and high dimensionality. However, in these cases, a question arises: the number of rules is so large (thousands or even millions of rules) that the extraction of accurate and interpretable Fuzzy Rule Bases (FRB) through classical sequential algorithms becomes computationally very costly and often unfeasible. In this context, Parallel Computing appears as a means to, from a large set of rules, enable the extraction of FRB that are weighted between accuracy and complexity. Therefore, this work proposes a parallel computing approach to generate FRB from high-volume and high-dimensional rule sets. The idea is that, using parallel computing and Multiobjective Evolutionary Fuzzy Systems (MOEFS), it is possible to extract accurate and compact (less complex | more interpretable) FRB from a large set of rules. For the generation of MOEFS, the Multiobjective Non-Dominated Sorting Genetic Algorithms NSGA-DO, MNSGA-DO and the classic NSGA-II were used. Each of these algorithms was tested to generate FRB from sets of Fuzzy rules obtained through the FCA-Based method. With the tests it was noticed that, with the use of AGMO NSGA-DO and MNSGA-DO, it was possible to achieve, in addition to the two main objectives, a better distribution of the solutions along the Pareto Frontier compared to the solutions obtained by the NSGA -II. The tests showed that, through the proposed approach, it was possible to extract accurate and compact BRFs from large sets of rules. Furthermore, tests were performed following a method based on the state-of-the-art FARC-HD approach. It was possible to perceive that the BRF obtained through the approach proposed here had, in general, greater accuracy in relation to the BRF extracted through the method based on the FARC-HD approach.