With the development of neural network theory, neural network technique has become an efficient ways for Pattern Classification. Neural network learning algorithm consists of supervised and unsupervised ways. In unsupervised learning algorithm, classification results are often not satisfactory. However in supervised learning algorithm, training samples needed to be labelled by manual, which is sometimes very difficult, and moreover whose classification results mainly depend on the selected training samples. According to these limitations, This paper improves unsupervised Self-Organizing Map (SOM) Neural Network and supervised General Regression Neural Network(GRNN), and lastly proposes a hybrid neural network of combining SOM Neural Network with GRNN.The main contributions of this thesis are given as follows:(1) By analyzing several classical neural network model, we understand the future trend and direction of neural network, and mainly study the models of both SOM Neural Network and GRNN.(2) After deeply studying kernel function principle, we use kernel function and hybrid kernel function to improve traditional SOM algorithm, and the advantages of new method are shown by several experiments.(3) We use Particle Swarm Optimization algorithm (PSO) to automatically optimize the parameter of GRNN, which avoids the influence of different artificial parameter by manual on classification results. Several experimental comparative results suggest their efficiency.(4) We propose a hybrid neural network model of combining unsupervised SOM with supervised GRNN. Experimental results for both Iris data , Wine data and remote sensing data verify the validity of our hybrid model. |