Classification And Time Series Prediction Based On Fuzzy Neural Network And Ensemble Learning

In some cases,a single neural network that used to handle classification and predictionproblems is difficult to achieve the desired results,so ensemble ideas is commonly used to deal with these problems.In this paper,classification and time series prediction based on fuzzy neural network and ensemble learning are discussed.The fuzzy neural network is used as the basic classifier of ensemble learning to deal with the classification problem.And incremental learning is used to update network parameters when input training data increases.In addition,ensemble ideas is applies to time series prediction.The main work of this article is as follows:Firstly,ensemble learning and fuzzy neural network are combined to deal with classification problems.Wavelet fuzzy neural network,fuzzy TS neural network and fuzzy TS neural network based on second-order Taylor formula are used as the ensemble learning classifier.The Adaboost correlation ensemble algorithm is used to integrate the training results of these networks.In the fuzzy neural network,restricted Boltzmann machine is uesd to initialize the fuzzy parameters with the benefit of reducing the impact of the parameter randomness.Besides,the learning rate with the scaling factor is used to learn the network parameters.Through the learning rate adaptive adjustments,the classification accuracy of the model increases and the number of iterations are reduced.The simulations in the kidney disease datasets and some datasets in the UCI database show that the proposed model can achieve better classification accuracy than other classification algorithms.Secondly,for the situation that the amount of input data increases with time,ridge regression and incremental learning are introduced into the training of fuzzy TS neural network.Ridge regression is used to update the matrix pseudo-inverse so as to update the network parameters.When the network training input data and fuzzy rules increase,corresponding incremental algorithms are given to update the network parameters to reflect this change,thus avoiding the parameter retraining process.Simulations in the MNIST handwritten font dataset and the NORB dataset are uesd to verify the validity of the proposed model.Finally,the integrated idea is applied to the prediction of time series.Linear model and quadratic model are used as a group of base predictors.Wavelet fuzzy neural network,fuzzy TS neural network,and fuzzy TS neural network based on second-order Taylor formula are used as another set of base predictors.The results of the base predictor are integrated and compared using different weighted summation methods or regression models.Simulations in the Mackey-Glass time series,TSDN database,and hemodialysis datasets are uesd to verifythe validity of the proposed model.