非凸函数的异步并行拟牛顿算法的收敛性分析

Asynchronous-parallel Pseudo-Newtonian Algorithm for Solving Non-convex Function Minimals

若假设可供使用的处理机具有p+q台,将其分成两组,两组处理机之间进行异步并行计算。提出了一种求解非凸函数极小的异步并行拟牛顿算法。若假设目标函数是二阶连续可微的,二阶导数矩阵在极小点x*处正定,步长由Wolfe原则确定,证明了所提出异步并行算法的全局收敛性。

p+q sets of processors are assumed, which are available for use.They are divided into 2 groups, for which asynchronous and parallel calculations are conducted.A method is proposed for solving non-convex function minimals.If the target function is 2 order continuously differential and the Hessian matrix is positive definite at x* and its steplength is ensured by Wolfe method,global convergence is demonstrated for the parallel pseudo-Newtonian algorithm.