Year

2006

Degree Name

Doctor of Philosophy (PhD)

Department

School of Information Technology and Computer Science - Faculty of Informatics

Abstract

This thesis presents a novel extension to the Adaptive Neuro-Fuzzy Inference System (ANFIS) which we call extended ANFIS (EANFIS). The extension includes the introduction of an output class based membership function architecture, in which each output class in a discrete output situation has its own membership function and in the case of a continuous output, only one class; the possibility of determining the structure of the rule base from the underlying structure of the input variables; the determination of a possibly non-symmetric membership function the parameters of which can be determined automatically from the given input variables; the possibility of incorporating global information on the input variables through a Linear Discriminant Analysis in combination with the local input variable structure as represented by the membership functions. The possibility of determining the structure of the rule section before the training process commences means that the proposed EANFIS architecture can be applied to possibility large scale practical problems, as it does not require the formation of all possible combination of rules before the training process commences. In other words, the EANFIS architecture together with its structure determining procedures overcomes the current limitation facing ANFIS architecture when applied to systems with large number of inputs. The possibility of determining a membership function from the input variables means the user no longer needs to select a membership function from a set of candidate membership functions. The possibility of incorporating global information on the input variables in addition to the local information on input variables means that the EANFIS architecture can take advantage when such global information might be useful in improving the performance of the Neuro-Fuzzy System. The new EANFIS architecture is evaluated on a number of standard benchmark problems, and have been found to have superior performance. In addition, as this is an EANFIS, rules can be extracted from the trained system, thus providing information on the way in which the underlying system is operating. The proposed EANFIS recommends itself readily for applications in practical systems.

02Whole.pdf (1741 kB)

Share

COinS