Improved constructive learning algorithms for fuzzy inference system identification

This paper presents an improved constructive learning algorithm for fuzzy inference system identification. An incremental training procedure that starts with a single pattern and a single-fuzzy rule has been used: if after several attempts, the fuzzy model cannot reduce the error within the specified tolerance; it grows by adding a new fuzzy rule. In order to overcome the over-training problem and to ameliorate the performance of the previous algorithm, two techniques of reduction have been introduced. In the first one, the growing of the fuzzy rules is conditioned by the generalisation error. In the second approach, a technique based on the similarity measures has been applied. The presented approaches have been applied for two examples to show the identification performance.