一种自适应RBF分类神经网络模型的构造方法

An Adaptive Constructing Algorithm of RBF Network and Its Applica tions to Classification

该文在自适应算法——DDA算法的基础上,提出了学习算法过程中加入裁减规则,从而进一步提高测试数据识别率观点。另外,对于这种算法进行了实验性的研究,通过实例分析,对于阈值选取的合理性、初始宽度的选取以及迭代停止的条件等方面进行了一定的探讨。最后的实验表明了采用这种改进算法来构造的RBF网络相对DDA方法而言,能够进一步提高网络的识别率,但同时保持了DDA这种自适应的RBF网络构造方法快速、简单且有效等诸多优越性能。

Many RBF network training algorithm confronts difficulties especially facing determination of network topology.None-Supervised/Supervised hybrid approaches were introduced.Inspired by the P-RCE method,the aim of this paper is to propose a novel approach to establish the structure of RBF dynamically by surmounting the fault of irreversibility of decays.The nature of this algorithm allows training to reach stability much faster than is the case for gradient -descent based methods.Also,we further discuss the rationality of the selection of positive and negative threshold,determination of the initial width of prototype and the stop criteria of iterative.The results of the given classical experiments illustrate the superior performance in classification applications with some consideration of improvement try.