神经网络的具有自适应动量和步长的伪牛顿算法
QUASI NEWTON ALGORITHM OF ADAPTIVE DECOUPLED STEP AND MOMENTUM FOR NEURAL NETWORK
-
摘要: 以单隐层的3层前向神经网络为基础，由自适应BP算法和牛顿优化算法导出了自适应步长和动量解耦的伪牛顿算法（QNADSM）．该算法计算量小，收敛速度快．文中还给出了该算法的收敛性证明、算法的仿真结果及其它算法的比较结果，并对网络的训练及该算法的特点作了进一步的讨论．仿真结果表明QNADSM算法是一种有效的工程实用算法．Abstract: In this paper, a quasi Newton algorithm with adaptive decoupled step and momentum(QNADSM) used in three layered feedforward neural network is derived from adaptive BP algorithm and Newton algorithm, which is of less computational amount and fast convergence. In this note we also present the proof of the convergence and simulation results. The simulation results demonstrate that QNADSM algorithm is an effectively practical algorithm.
下载:
