RSOM算法及其应用研究

神经网络以其优越的非线性拟合能力和强大的自组织模式分类能力已被用于许多模式识别问题，并取得了很好结果．但是对于大样本集分类和复杂模式识别问题，大多数常规神经网络在决定网络结构与规模、设计自学习算法和应付庞大的计算量等方面存在诸多困难．为了克服这些困难，在Kohonen自组织映射模型(SOM)的基础上，提出了两种基于类别可分性判据RSOM分类树：非结构自适应的RSOM-Ⅰ分类树与基于奇异值分解方法的结构自适应RSOM-Ⅱ分类树，这两种分类树的每个节点由拓扑有序的自组织映射网络组成．RSOM分类树的优点在于处理大样本集和复杂模式的识别问题时能够自适应地确定网络的结构和规模，最后的数据试验就是很好的佐证．

Research on RSOM Algorithm and Its Application

Neural network have been widely used in pattern recognition for their great capability to non-linearity fitting and self-organizing pattern classification, and acquired great achievement recently. However, a number of general neural networks have several difficulties such as deciding their structures and scales, designing their self-learning procedure and coping with a bulk of computation for the case of large data set classification and complex patterns recognition. In order to solve these problems, it proposes two RSOM tree classifiers based on discrimination criterion approach. RSOM-I tree classifier is not structure-adaptive, while RSOM-II tree classifier is structure-adaptive by employing the SVD (Singular Value Decomposition) approach. Both RSOM tree classifiers are composed of topology-preserved SOM nets, and their scales are determined by discrimination criterion and SVD partially. The main advantage of these new neural networks is that they adjust their structure and scale automatically with the large training data set, so they map the training set very well. This makes them achieve high right recognition rate, and the experiments in the end are very good proofs of these new networks.