Evolution–fuzzy Rule Based System with Parameterized Consequences

While using automated learning methods, the lack of accuracy and poor knowledge generalization are both typical problems for a rule-based system obtained on a given data set. This paper introduces a new method capable of generating an accurate rule-based fuzzy inference system with parameterized consequences using an automated, off-line learning process based on multi-phase evolutionary computing and a training data covering algorithm. The presented method consists of the following steps: obtaining an initial set of rules with parameterized consequences using the Michigan approach combined with an evolutionary strategy and a covering algorithm for the training data set; reducing the obtained rule base using a simple genetic algorithm; multi-phase tuning of the fuzzy inference system with parameterized consequences using the Pittsburgh approach and an evolutionary strategy. The paper presents experimental results using popular benchmark data sets regarding system identification and time series prediction, providing a reliable comparison to other learning methods, particularly those based on neuro-fuzzy, clustering and ε-insensitive methods. An examplary fuzzy inference system with parameterized consequences using the Reichenbach implication and the minimum t-norm was implemented to obtain numerical results.