A hybrid intelligent system: genetic algorithm and rough set incorporated neural fuzzy inference network
Date of Issue2014
School of Computer Engineering
Centre for Computational Intelligence
Neural Fuzzy Inference System (NFIS) has been intensively investigated due to its aptitudes in accurate data processing as well as extractable and human interpretable inference rule base. However, most NFIS architectures focus primarily on modeling accuracy. Lesser attention has been devoted to improve the interpretability of the fuzzy inference rule base. In this thesis, a high-level interpretability of the rule base is manifested through the use of a small number of features and a small number of simple rules without redundancies and inconsistencies. Furthermore, a small number of fuzzy membership functions without much overlap between the adjacent ones promote greater legibility. With a high-level accuracy and better interpretability, intelligent systems become more convincing to developers, experts, and users. However, many established methods that perform knowledge reduction are applied to the derived rule base. Lesser attention has been devoted to derive a compact fuzzy inference rule base from the initial design and during an iterative optimization process. This thesis aims to construct an NFIS that achieves a high-level interpretability with competitive accuracy in an integrated framework focusing on leveraging the trade-off between accuracy and interpretability from the design phase and throughout the iterative optimization process. To realize this objective, rough set theory is incorporated to perform knowledge reduction and genetic algorithm is incorporated to search for pseudo optimal solutions. A hybrid architecture, named Genetic Algorithm and Rough Set Incorporated Neural Fuzzy Inference System (GARSINFIS), is proposed in this thesis. GARSINFIS employs Genetic Algorithm based Rough Set Clustering (GARSC) technique, which systematically divides data into clusters and represents them using minimal amount of necessary knowledge. GARSC ensures a relatively small network size of GARSINFIS. The trade-off between interpretability and accuracy can be tuned by adjusting the coefficients of different components defined in the fitness evaluation function. Moreover, the Recursive Least Squares (RLS) algorithm, which boosts accuracy but degrades interpretability, can be applied to increase the order of the employed fuzzy rules and further fine-tunes them after the network structure has been automatically determined according to the clustering result. A committee of multiple GARSINFIS networks with the weighted voting scheme can be employed to increase the overall accuracy. GARSINFIS is applied to eight publicly available and widely used data sets for performance comparisons against other classical and similar models. Furthermore, it is applied to two real-world applications for performance evaluations, wherein the automatically selected input features and systematically derived inference rules are compared to the ones used by the experts to demonstrate how well GARSINFIS performs in terms of knowledge reduction without prior knowledge and human intervention.
DRNTU::Engineering::Computer science and engineering::Computer systems organization::Performance of systems