对学习样本无误分类的对数型目标函数BP算法

提出了一种对学习样本无误分类的改进ＢＰ学习算法。该算法采用对数型目标函数，可以减少每次迭代的计算量。同时将输出节点分为正确分类节点和误分类节点两类。对于误分类节点，将其误差项加入到目标函数中，然后采用梯度下降算法进行学习。在学习过程中，对学习率μ（ｋ）采用动态优化确定方法，以加快算法收敛速度。为保证收敛后的网络对学习样本能够正确分类，算法终止条件要求对所有输入样本，无误分类输出节点。算例仿真表明了算法的有效性。

The Log Like Objective Function BP Algorithm with No Misclassification for the Learned Samples

This paper presents an improved BP algorithm with no misclassification for the learned samples.The approach uses the Log Like Objective function to reduce the amount of computation.In the algorithm the output nodes are classified into two categories of misclassified nodes and correct classified nodes.The errors of misclassified nodes are added to the objcctive function,and using the well known gradient descent algorithm to train the network.The learning rate μ(k) is determined by dynamic optimization method to accelerate the convergence rate.To guarantee the network can correctly classify the learning patterns, the condition of no misclassification nodes for all input samples is added to the algorithm terminated conditions.Computer simulation demonstrates the effectivity of the proposed algorithm.