A Systematic Methodology to Obtain a Fuzzy Model Using an Adaptive Neuro Fuzzy Inference System. Application for Generating a Model for Gas-Furnace Problem

摘要

In this paper we present a complete design methodology to obtain a fuzzy model with an Adaptive Neuro Fuzzy Inference System (ANFIS). This methodology consists of three phases: In phase I, the automatic selection of input variables and other parameters such as number and type of membership functions is made with a Genetic Algorithm with a special fitness function, obtaining a basic structure of the fuzzy model. The second phase is a massive training of the fuzzy model previously obtained. Finally, the third phase is a post-adjusting of the weights of the rules with a local search algorithm, based on an adjusted fitness function from the first phase. An application of the proposed design method for the gas-furnace time series, a well-known benchmark dataset used by many researchers in the area of neural networks and fuzzy systems is presented, and finally, we present a comparative with other Box-Jenkins models.