解无约束优化的渐弱过滤集方法-计算数学专业论文.docxVIP

AbstractMany Abstract Many important problems be expressed in ter-ui8 of nonfinear multivariate unconstrained optimization problems．Two broad classes of Mgorithms for the unconstrained optimization problenm口e line-search algorittnns and trust—re{；ioilalgorithms．Since filter tocJmiquc was introdu(『cd by Fletcher and Leyffcr in 1997 mid subsequently published as【131． filter technique has been studied and applied in many areas，and it performs well． We proposed new dwindling filter technique for large-scale unconstrained optimization problems．Generally：for the purpose ofgiobal convergeltce，the multidimensional filter f21，25， 271 is covered by fixed envelope．Since fixed envelope is not suitable for backtracking line- search process，we further study it in this thesis and put forward conception of dwindling filter technique．The main idea is that the dwindling filter envelope is not fixed，but becomes thinner and thinner as the step-length of the backtracking fine-search becomes smaller and smaller． Naturally，combination of the dwindling filter techliique and the second—order line-search method is very significant algorithm for unconstrained optimization．The new algorithm is shown to be globally convergent t0 second-order stationary point at least．Preliminary numerical experiments set of CUTEr test problems indicate that the new algorithm is very competitive with classical line-search algorithms．We show that the linked-list construction in the FORTRAN language works well aald that the storage of the dwindling fiIter is medium． And then，we notice that the trtLst-region method is ahvays more effective than line- search method．This inspke to apply the dwindling filter technique in trust-re西On line- search framework to give another new algorithm．We prove that{under reasonable MSUlllp— tions．the new algorithm converges to second-order stationary point at least．Some preliminary numerical results set of CUTEr test problems reported．which show that this new algorithm has significant