The Research Of Several Robust New Intelligent Modeling Algorithms And Their Applications

In recent decades, modeling technology has obtained fast development, and has been widely used in a variety of domains of the social production and life. Traditional modeling methods mainly include the theoretical modeling method and the traditional identification modeling method, both of which show a very good effect in solving some simple linear problems. However, with the continuous development of the society, lots of realistic problems become more and more complex. Most of the actual systems are all complex, nonlinear and time-varying, and these properties all bring great difficulties to the traditional modeling methods. In recent years, the intelligent modeling method which is formed by combining the artificial intelligence technology and traditional identification has won great attention of people. Long-term research and practice have proved that the intelligent modeling method can effectively overcome the defects of the traditional modeling methods. The intelligent modeling method is an important approach of the complex system modeling. So far, the artificial intelligence technology and theory which have been used in the system identification mainly include the artificial neural network, the fuzzy logic, the genetic algorithm, the wavelet network, and so on, and there have been many related research achievements. The artificial neural network system identification and the fuzzy system identification are system modeling methods which are most widely researched and used. This paper first introduces some related knowledge of the artificial neural network system identification and the fuzzy system identification, and then respectively puts forward a new intelligent modeling method in the fields of the two technologies.Single hidden layer feed-forward neural network is one of the most widely used models for intelligent modeling. But the model faces a following critical challenge: For small sample sets, the traditional learning algorithm may train a model to fall into the over-fitting sate. In particular, when the dataset contains a large amount of noise, the trained model has weak robustness and is very sensitive to the noise. In order to overcome this shortcoming, a robust learning algorithm of SLFN is derived for small and noisy datasets. Due to introduction of theε-insensitive learning measure and the structural risk term, the proposed algorithm can effectively overcome the shortcoming of the traditional learning algorithms. The experimental results on simulated and real-world datasets also confirm the above advantages.It has been mathematically proved that a gaussian mixture model can be translated to a generalized fuzzy model. Based on the above theory, this paper presents a two-phase method of constructing an additive generalized fuzzy system. The first phase is to obtain the gaussian mixture model from the sample, and translate the model to the form of the additive generalized fuzzy model; In the second phase, a new objective function is proposed to further train this model by introducing theε- insensitive learning measure and the structural risk term, and the solution to the proposed new objective function is transformed into a classic quadratic programming problem. Due to introduction of theε- insensitive learning measure and the structural risk term, the learning algorithm in the second phase can effectively overcome some defects of the traditional learning algorithms, including easy to fall into a fitting, lack of robustness, slow training speed, etc. The experimental results on simulated and real-world datasets also confirm the above advantages.