自适应RBF-LBF串联神经网络结构与参数优化方法  

Adaptive Structure and Parameter Optimizations of Cascade RBF-LBF Neural Networks

作　　者：高大启[1] 

机构地区：[1]华东理工大学计算机科学与工程系

出　　处：《计算机学报》2003年第5期575-586,共12页Chinese Journal of Computers

基　　金：国家自然科学基金 (60 2 75 0 17);上海市重点科技攻关项目(0 2 5 115 0 2 8)资助

摘　　要：研究了前向单层径基函数 (RBF)网络和前向单层线性基本函数 (LBF)网络的分类机理 ,提出了RBF的中心和宽度应通过学习自动确定 ,在学习过程中根据错分样本被错分入的类别自动生成新的核函数这一观点 .如果两个或两个以上核函数属于同一类 ,在输入空间相距较近且未被其它类别的样本分隔开来的情况下 ,则应考虑将之合并 ,或者使它们的作用区域部分重叠 .从理论上阐明了采用Sigmoid活化函数的单层感知器的分类阈值为0 .5 ,进而提出了由单层RBF网络和单层感知器组成的串联RBF LBF神经网络 .文中详细给出了确定该串联RBF LBF神经网络结构、核函数个数、位置与宽度的优化算法 .一般来说 ,该算法的计算复杂性比前向单隐层感知器采用的误差反传算法要小或至少相当 .对几个经典的模式分类难题的处理结果表明 ,与一般RBF网络和前向单隐层感知器网络相比 ,该串联RBF LBF网络及其自适应学习算法具有收敛速度快 ,分类精度高 ,易于得到最小结构 ,在学习过程中不易陷入局部极小点等优点 ,有利于实现实时分析 .实验结果同时也验证了单层LBF网络对提高RBF LBF网络分类正确率的重要性 .Based on researching into the classification mechanisms of feedforward single-layer radial basis function CRBF) and linear basis function (LBF) networks, the author presents the viewpoints that the RBF centers and widths should be determined through a self-learning procedure, that some new kernels naturally come into being according to which class the labeled patterns are misclassified to. Author holds that two or more kernels should be automatically merged or overlapped in part if they belong to the same original category and are neighboring each other and not divided by other classes. The reason is clarified why the classification threshold of feedforward single-layer perceptron networks with Sigmoid activation functions should be 0.5. And furthermore, a kind of cascade RBF-LBF networks consisting of a single-layer RBF network and perceptron are proposed. The learning algorithm for optimally determining the number, locations and widths of kernels in cascade RBF-LBF networks is gone into details. Generally speaking, the presented algorithm has much lower computational complexity than its backpropagation correspondent used in feedforward single-hidden-layer perceptron networks. The results for some classically tricky classification problems show that the proposed cascade RBF-LBF networks as well as their adaptively learning algorithm, have advantages over standard RBF networks and feedforward single-hidden-layer perceptrons. For examples, the cascade networks have fast convergence rate, high classification accuracy, large probability to get optimal structures, good capacity to reach global minimum points, etc. Therefore, the cascade networks are quite favorable for real-time analyses. The experiments still verify that single-layer perceptrons are very important for improving the classification correct rate of cascade RBF-LBF networks.

关 键 词：自适应RBF-LBF串联神经网络 网络结构 参数优化方法 径基函数 模式分类 感知器 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]