Evolving Fuzzy Rule Based Classifiers with GA-P: A Grammatical Approach

Genetic Programming can be used to evolve Fuzzy Rule-based classifiers [7]. Fuzzy GP depends on a grammar defining valid expressions of fuzzy classifiers, and guarantees that all individuals in the population are valid instances of it all along the evolution process. This is accomplished by restricting crossover and mutation so that they only take place at points of the derivation tree representing the same non-terminal, thus generating valid subtrees [13]. In Fuzzy GP, terminal symbols are fuzzy constants and variables that are chosen beforehand. In this work we propose a method for evolving both fuzzy membership functions of the variables and the Rule Base. Our method extends the GA-P hybrid method [6] by introducing a new grammar with two functional parts, one for the Fuzzy Rule Base (GP Part), and the other for the constants that define the shapes of the fuzzy sets involved in the Fuzzy Rule Base (GA Part). We have applied this method to some classical benchmarks taken from the collection of test data at the UCI Repository of Machine Learning Databases [9].