Identification of Nonlinear Dynamic Systems via the Neuro-Fuzzy Computing and Genetic Algorithms

In a modeling process of a real world problem, there usually are a huge number of potential inputs involved. A large number of inputs may increase the complexity in computation and cause other problems related to running time, memory spaces, etc. In the case of modeling process with large input, the number of inputs should be reduced and the priority inputs should be determined by an optimal selection method. In this paper, an effective method for selecting significant input variables in building ANFIS (Adaptive Neuro-Fuzzy Inference System) [1] model of a nonlinear system is proposed. The important input variables which independently and significantly affect the system output can be extracted by Genetic Algorithms (GAs) [3, 13]. GAs are a powerful search algorithms that use operations found in natural genetics to guide the trek by space searching. Moreover, GAs are theoretically and empirically proven to provide robust search capabilities in complex spaces, offering a valid method to problems requiring efficient and effective searching. Using GAs, dominant inputs for a nonlinear system identification are extracted through evaluating error function. Regarding to the ANFIS, it combines the power of fuzzy systems with that of neural networks so that the structure of ANFIS is a hybrid system and mainly based on Sugeno-fuzzy model [4]. The way of modeling proposed in this paper is as following. Firstly, we determine the initial conditions for the ANFIS structure. Then, a GA is used as a dominant input selector to enable the system to have reduced input dataset by eliminating irrelevant inputs in problems with a large number of inputs. And then, the selected dominant inputs are applied to the ANFIS to identify the nonlinear system. In order to verify the availability of the performance, Box and Jenkins gas furnace data [2] is applied to our proposed method. In the simulation results given in section 3, the proposed method combining GA with ANFIS shows a relatively high performance and it indicates that our model is an adequate model for a nonlinear identification. This paper is composed of four sections. In section 2, the concepts of the Neuro-Fuzzy system and the Genetic Algorithms are introduced. In section 3, the experiment results with a nonlinear dynamic system are shown. This paper is concluded in section 4.