解无约束优化的渐弱过滤集方法.pdfVIP

Abstract be in ofnonconvexnonfinear c3n ter-ui8 multivariate importantproblemsexpressed Many classesof forthe unconstrained broad unconstrained optimizationproblems．Two Mgorithms line-search and filter optimizationproblenm口e algorittnnstrust—re{；ioilalgorithms．Since was Fletcherand in1997mid introdu(『cd tocJmiquc by Leyffcr subsequentlypublishedas【131． filter hasbeenstudiedand in it well． technique appliedmanyareas，andperforms We anew filter for unconstrained proposeddwindlingtechniquelarge-scale optimization the multidimensionalfilter problems．Generally：forpurposeofgiobalconvergeltce，the f21，25， 271iscoveredafixed afixed isnotsuitablefor line- by envelope．Sinceenvelope backtracking search furtheritinthisthesisand forwarda of process，westudy put conceptiondwindling filter mainideaisthatthe filter isnot becomes technique．The dwindlingenvelopefixed，but thinnerandthinnerasthe ofthe fine-searchbecomessmallerand step-lengthbacktracking smaller． ofthe filter andthesecond—orderline-search Naturally，combination techliique dwindling methodisa forunconstrained new is very significantalgorithm optimization．Thealgorithm showntobe t0onesecond-order atleast．Preliminary globallyconvergent stationarypoint