Computational intelligence-based techniques in the construction and reduction of rule-based systems.

摘要

This dissertation focuses on applying Computational Intelligence, a consortium of the technologies of fuzzy sets, neurocomputing and evolutionary computing, to the design and analysis of fuzzy rule-based systems (FRBS). We discuss two methods to construct FRBS, where the crux of the method is to seamlessly be based on the fuzzy neural network (FNN) and fuzzy c-means (FCM) clustering respectively. The rule set for the FRBS derived from the above two methods commonly consists of quite a number of rules and including all the attributes from the problem inputs. It becomes necessary and intuitive to reduce the dimensionality (number of input attributes in the rule) of the rules in the rule set. Also, some rules in the rule set might be conflicting with others. To make the FRBS more concise, the less important rules could be removed from the rule set. So after finishing the construction of FRBS, the rule complexity reduction algorithms are applied.;The key results of this study include: · Construction of the FRBS with the aid of FNNs where the network is developed through genetic optimization. · Reduction of complexity in terms of dimensionality and quantity (viz. the number of rules) by configuring pruning thresholds for AND neurons and OR neurons. The optimal values of the thresholds are determined in a way one strikes a sound balance between the interpretability of the rules and the accuracy associated with the reduced (simplified) rules. To develop the model optimal against these two competing objectives, multi-objective optimization is considered. · Application of FRBS constructed with the use of FNN to well-known datasets and a real-world application such as deployment of wireless sensor networks. · Construction of FRBS involving mechanism of information granulation (fuzzy clustering) and local linear models and studies on their complexity management through reduction of condition space and a relational expansion of fuzzy clusters.