摘要
根据神经网络模型的结构特点,将能量函数的二阶导数与最速下降方向相结合,构造出一种新型的BP算法,该算法比梯度法收敛快,较牛顿法计算量小.它适合于计算结构复杂的BP神经网络模型,理论分析表明该算法行之有效,计算机仿真达到了理想的效果.
Based on the architectural features of the neural network, a new BP algorithm is developed by combining the gradient direction with the second derivative of the energy function. It is shown that the rate of convergence of the algorithm proposed is faster the that of gradient method, and its amount of work is less than that of Newton method. It is fairly suitable to the calculation of a BP neural network model of complicated architecture.
出处
《华中理工大学学报》
CSCD
北大核心
1998年第3期105-107,共3页
Journal of Huazhong University of Science and Technology
基金
国家自然科学基金
关键词
BP算法
神经网络
能量函数
梯度法
二阶导数
descent direction
BP algorithm
neural network
energy function
gradient method
Newton method