神经网络已广泛应用于各类问题，然而ＢＰ算法要求有连续且可微的激活函数，文中提出一种用于训练非可微激活函数的神经网络学习算法。同时，利用相对熵误差测度，算法被完整地导出。实验结果表明，在解决异或问题、编码／解码问题及其补问题时，算法收敛速度非常快，收敛结果也令人满意。 Neural networks have been widely used in various kinds of problems. However, the BP algorithm requires continuous and differentiable activation functions. In this paper, a neural network learning algorithm for training non-differentiable activation functions is proposed. At the same time, the algorithm is derived completely using the relative entropy error measure. The experimental results show that the convergence speed of the algorithm is very fast and the convergence result is satisfactory in solving XOR problems, encoding / decoding problems and their complement problems.