|
This thesis presents appropriate model structures for fuzzy
systems, and accompanies these model structures with
parameter-level learning and structure-level learning. The
emphasis of the present exercise is on basic principles of the
design, operating characteristics, and adaptation of fuzzy
systems. In order to incorporate adaptation into fuzzy systems,
a refined mathematical model format for fuzzy systems is
developed in such a way that the fuzzy logical rules in the
systems are parameterized. Two general learning paradigms are
considered: competitive learning and supervised learning.
By incorporating competitive learning into fuzzy system, we
demonstrate that fuzzy systems can be used effectively for
categorization and clustering of unlabeled input patterns.
Methods for dynamically adjusting the parameters and structures
based on fuzzy competitive learning are discussed. To facilitate
system design, we present several supervised learning algorithms
for adjusting parameters. Also, a novel structure-level
supervised learning algorithm that is able to self-organize the
number of fuzzy rules is proposed. The results of simulations
reveal that the proposed parameter-level as well as structure-
level competitive learning and supervised learning algorithms
are practically feasible. Potential applications include
function approximation, pattern classification, vector
quantization, clustering, and system identification.
|