Generalization Error Research And Apply Of Neural Network Ensemble

Now, neural network ensemble technology has been widely applied in regression and classification fields. Improving classification and forecasting accuracy as one of it is an important yet difficult task facing decision maker in many areas. Combing multiple models can be an effective way to improve forecasting performance. Neural network ensemble (NNE) as one of multiple models, easy to use and good performance, is becoming a hot research field in machine learning and neural computing.In this paper, we have done some research on theory of NNE and applied it to the field of time series forecasting and classification, through increasing NNE generalization error to improve time series forecasting accuracy. Mainly works about improving NNE generalization error are shown as following.Firstly, from the point of original training data, in order to fully using original training data and improving generalization error, by adding noises into the input data and thus augment the original training data set to form models based on different but related training samples. Individual neural networks train on different training samples and get high difference degree. Form this angle, it improve generalization error.Secondly, in order to augment difference degree of individual networks, proposing a method of evolutionary ensemble of neural network based on niche technique. Using niche technique's good performance in increasing population diversity and improving local search of evolutionary, adjusting individual network's fitness by the similarity degree's sharing function among individuals. Then select networks according to the new adjusted fitness to get individual networks with diversity.Lastly,apply proposed methods to the power short-term load forecasting of a city of south and two criteria knowledge data sets in the UCI machine learning. Experiment results show that these methods can get good performance.