Abstract
This thesis proposes a co-evolutionary-based approach
to solve the problem of automatic fuzzy system design.
The co-evolution supports hierarchical and
collaborative relations among individuals representing
different parameters of fuzzy models. The proposed
approach takes species which encode partial solutions
to fuzzy modeling problems, organized into four
hierarchical levels. Each hierarchical level encodes
membership functions, individual rules, rule-bases and
fuzzy systems, respectively. A special fitness
evaluation scheme is proposed to measure the
performance of each individual of different species.
Constraints and local objectives must be observed at
all hierarchical levels to guarantee the occurrence of
individuals characterized by the simplicity of fuzzy
rules, rule compactness, rule base consistency and
visibility in the universe partition. The approach
allows the evolution of Mamdani or Takagi-Sugeno fuzzy
models. In addition to performance improvement in terms
of accuracy and interpretability, the co-evolutionary
approach increases autonomy by minimizing user
intervention, since it allows automatic tuning of a
number of critical parameters, like type and total of
fuzzy rules, relevant variables (for each rule and for
the whole application), shape and location of
membership functions, antecedent aggregation operator,
and, for Mamdani models, aggregation operator, rule
semantic, and the defuzzification method. The
performance of the approach is evaluated via function
approximation and pattern classification problems.
Users
Please
log in to take part in the discussion (add own reviews or comments).