Perception modelling using type-2 fuzzy sets
Type-1 fuzzy logic has, for over thirty years, provided an approach for modelling uncertainty
and imprecision. This methodology has been highly successful with a history
of successful applications in a number of areas - particularly control. However, type-1
fuzzy systems are essentially `crisp' in nature. This is not only paradoxical but also raises
concerns for knowledge representation and inferencing. In particular type-1 fuzzy logic
is flawed when representing perceptions such as colour, beauty, comfort etc. since these
perceptions do not have a measurable domain.
This fundamental paradox is tackled in this research by employing a type-2 fuzzy
paradigm. The investigation of the type-2 approach concludes that the uncertainty or
imprecision that exists in most real problems can be more effectively modelled by a type-2
approach. The research reported in this thesis explores the properties of type-2 fuzzy sets
as well as showing how useful they can be for knowledge representation and inferencing.
It is shown that type-2 fuzzy sets have an important role to play in modelling perceptions.
Results are given of using type-2 fuzzy sets to represent perceptions of a medical expert
for shin image analysis indicating that the type-2 fuzzy paradigm is particularly helpful
for perception representation.
A methodology has been developed that allows linguistic inputs to an adaptive system
that implements a type-2 fuzzy system(the Adaptive Fuzzy Perception Learner (AFPL)).
In this thesis, the rationale and full mathematical detail of the AFPL is presented. The
approach has been applied successfully to the, so called, linguistic AND (analogous to
the Boolean AND) as an aid to illustrating the methodology. Results are presented of
applying the method to a real problem of classifying the acceptability of a car based
on perceptions that describe certain features of the car. The AFPL is applied to this
large, complex, set of data where the inputs to the network are linguistic. A detailed
evaluation of the AFPL is given with recommendations for effective use of the AFPL.
The results indicate that we now, truly, have an approach for learning the perceptions
and relations in a type-2 fuzzy system