节点文献

解无约束优化的非单调信赖域法和Perry-Shanno无记忆拟牛顿法

【作者】 杭丹

【导师】 颜世建;

【作者基本信息】 南京师范大学 , 计算数学, 2007, 硕士

【摘要】 本文对求解无约束优化问题(?)f(x)给出三个算法:(1)不重解子问题的非单调自适应信赖域算法。(2)非单调Perry-Shanno无记忆拟牛顿方法,(3)非单调带参数的Perry-Shanno无记忆拟牛顿法。本文主要工作如下:(1)文[2]给出了一种自适应信赖域算法,其调整信赖域半径的公式是△k+1=Rc2(rk)‖dk‖。其中Rη(t)称为R-函数。我们给出一个比文[2]简单的新的R-函数Rη(t)并采用公式△k+1=Rc2(rk)△k调整信赖域半径。在数值试验中我们发现当试探步dk被接受时,有时dk可能是f(x)的一个极好的下降方向。取xk+1=xk+dk可能并没有充分利用这个好的下降方向dk,对这种情形,我们采用一种不精确线搜索来确定xk+1。另外当试探步dk不被接受时,我们没有重解子问题或向后线搜索,而是采用了一个固定的公式给出新的迭代点xk+1。对采用上述技巧的信赖域算法,在适当条件下,我们证明了它的全局收敛性。数值试验表明该算法是有效的。(2)对非单调线搜索的Perry-Shanno无记忆拟牛顿法,我们不仅证明了f(x)是凸函数时的全局收敛性,同时在f(x)是非凸函数时的收敛性也作了深入的探讨,并给出了几个收敛的充分条件。初步的数值试验表明了算法的有效性。(3)在第二个工作的基础上给出了非单调带参数的Perry-Shanno无记忆拟牛顿算法,我们不仅证明了f(x)是凸函数时的全局收敛性,同时在f(x)是非凸函数时的收敛性也作了深入的探讨,并给出了几个收敛的充分条件。并且可以通过参数的选取来控制解的误差,最后给出了几个演示性的算例。

【Abstract】 In this paper, we propose three methods for unconstrained optimiza-tion.(1).a new nonmonotonic self-adaptive trust region algorithm without resolv-ing the subproblem.(2). nonmonotonic Perry-Shanno Memoryless Quasi-Newtonmethod.(3).nonmonotonic Perry-Shanno Memoryless Quasi-Newton method with pa-rameter, the dissertation can be summarized as follows:1. Paper [2] proposes a self-adaptive trust region algorithm, and the formula oftrust region radius is△k+i= Rc2(rk)‖dk‖.Where Rη(t) names R-function.wepropose a new R-function which is simpler than Paper [2],and define a formula△k+1= Rc2(rk)△k as trust region radius, when a trial step is accepted,sometimesdk is a good descend direction,we take a inexact line search to obtain xk+1.When atrial step is not accepted,the method does not reslove the subproblem but generatesa iterative point whose steplength is defined by a formula.Under mild conditions,weprove that the algorithm is global convergence.Numerical results are presented whichshow that the method is effective. Numerical results are also presented.2. This chapter will try to discuss nonmonotonic Memoryless Quasi-Newtonmethod.The convergence of this method is proved for convex function and non-convex function. Primary numerical results demonstrate that the algorithm is efficient.3. The chapter takes nonmonotonic Perry-Shanno Memoryless Quasi-Newtonmethod with parameter. The convergence of this method is proved for convex functionand non-convex function.Preliminary numerical experiments show that the proposedalgorithm is effective.

  • 【分类号】O224
  • 【下载频次】99
节点文献中: 

本文链接的文献网络图示:

本文的引文网络