Novel clustering methods for neurofuzzy systems design

In this thesis, novel clustering methods with optimized parameters in order to have NeuroFuzzy inference systems design are developed. A modified version of Simulated Annealing (SA) optimization and Subtractive Clustering (SC) techniques are adapted to Fuzzy System to form a fuzzy classifier (SASCFC) in order to obtain optimum fuzzy rule base, parameters and to find out most important inputs. Four distinct classifiers namely SASCFC-Type1, Type2, Type3 and Type4 are derived in order to form different optimization scenarios. Although there are some similarities in each type of SASCFC, Type4 has best classification performance because a hybrid feature selection algorithm is also developed in Type4. Two new NeuroFuzzy Classifiers (NFC) are proposed as NFC1 and NFC2. Initial structures of both classifiers are set up via Rival Penalized Competitive Learning (RPCL) based clustering. A new RPCL type gradient descent training algorithm is also proposed for the NFC2. Rule adaptation mechanism is embedded into training of the NFC2 that both parameters and structural optimization performed twice which enables to change the structure of classifiers by adding new rules and deleting unnecessary rules in training phase dynamically. It is found that the proposed classifiers, which are tested on some benchmarks problems, have good performance in comparing to their counterparts in recent literature