论文部分内容阅读
针对神经元网络ＢＰ算法的局部极小问题，给出了一种避免陷入局部极小点的具有全局最优的神经网络ＢＰ算法。其特点是利用自适应线性单元和ＢＰ网络结合构成新的混合网络，网络权值仍用ＢＰ算法进行修正，其中自适应线性单元是以样本的零次幂到样本数减１次幂输入。通过数学证明此方法得到的网络的权值是全局最优。实例验证此方法的正确性。此方法简单、易行，并且不是以学习时间长为代价而得到全局最优的，所以对于ＢＰ算法在实际应用具有重要意义。
Aimed at the local minima of BP neural network, a BP neural network algorithm with global optimization that avoids falling into local minima is proposed. Its characteristic is to use the adaptive linear unit and the BP network to combine to form a new hybrid network. The network weights are still modified by the BP algorithm. The adaptive linear unit is input from the zero power of the sample to the number of samples minus 1 power. It is proved that the weights of networks obtained by this method are globally optimal. The example verifies the correctness of this method. This method is simple and easy to implement, and it is not globally optimal at the expense of long learning time. Therefore, it is of great significance to BP algorithm in practice.