摘要
RBF径向基函数神经网络具有训练简洁、学习效率快、不易陷入局部极小等优点,广泛应用于信号处理与模式识别。虽然常用的RBF网络比较容易构建,但因其结构通常固定或者复杂度较高,从而导致学习时间过长或网络资源的浪费。针对上述原因,提出利用扩展卡尔曼滤波器作为RBF的学习算法,并在隐层中使用双径向函数。通过对逼近基准的结果分析,清楚地表明该算法比其他分类网络模型具有更强的泛化性。
RBF neural network, with simple training, high efficiency of study, and be not easy to lost in local minimum, etc., is widely used in signal processing and pattern recognition. Although the commonly used RBF network is easier to built, their fixed structural or high complexity, usually causes too long learning time or a waste of network resources. For the above reasons, the extended Kalman filter is proposed as RBF learning algorithm, and the biradial function is used in hidden layer. By approaching the results of the benchmark analysis, we make it clear that this alorithm is more generalized nature than any other network model.
