An Improved Hyper-ellipsoidal Basis Function Network Classifier

Radial basis function neural network is a classical neural network, and it is successfully applied in pattern recognition problems. With the development of radial basis function networks, many neural network models with different basis functions are introduced. These networks generally take two learning methods:online learning and batch learning. The online learning neural network has strong adaptive ability. However, using online learning method, the radial basis function neural network will have redundant hidden neurons. Therefore, in order to improve the generalization ability of neural networks, we need a pruning strategy for hidden layer neurons.A self-adaptive resource allocation network classifier (MSRAN) which is based on hyper-ellipsoidal base function with pruning criterion is presented in this paper. In order to solve the classification problem with large-scale and high dimensional data, MSRAN, which uses S. Suresh’s learning method and divides the space with hyper-ellipsoid, forms an enclosed decision region. Thus it is more clearly and efficiently to divide the space. Also, it uses a center distance and angle criterion to prune the hidden neurons, thus the network generalized ability is improved. Finally, through the comparison of the MSRAN, SRAN and ordinary MSRAN, we show that according to different characteristics of classified data, the average classification ability of MSRAN and SRAN is almost the same, and the average classification ability of MSRAN after pruning is3.245%better than that of MSRAN without pruning.The structure of this thesis is organized as follows. Chapter1gives a brief introduction on artificial neural network. Chapter2includes the basic knowledge of radial-basis-function neural network and extended kalman filter. Chapter3is concerned with SRAN which is proposed by S. Suresh. Finally, MSRAN based on hyper-ellipsoidal base function with pruning criterion is presented, we also give a comparison between it and ordinary MSRAN by numerical experiments.