Transcription of 6.252 NONLINEAR PROGRAMMING LECTURE 4 …
{{id}} {{{paragraph}}}
NONLINEAR PROGRAMMINGLECTURE 4 CONVERGENCE ANALYSIS OF GRADIENT METHODSLECTURE OUTLINE Gradient Methods - Choice of Stepsize Gradient Methods - Convergence IssuesCHOICES OF STEPSIZE I Minimization Rule: kis such thatf(xk+ kdk) = min 0f(xk+ dk). Limited Minimization Rule: Min over [0,s] Armijo rule: f(xk)'dk f(xk)'dk0 Set of AcceptableStepsizes s sUnsuccessful StepsizeTrials 2sStepsize k = f(xk + dk) - f(xk) Start withsand continue with s, 2s, .., until msfallswithin the set of withf(xk) f(xk+ dk) f(xk) OF STEPSIZE II Constant stepsize: kis such that k=s:a constant Diminishing stepsize: k 0but satisfies the infinite travel condition k=0 k= GRADIENT METHODS WITH ERRORSxk+1=xk k( f(xk)+ek)whereekis an uncontrollable error vector Several special cases: eksmall relative to the gradient; , for allk, ek < f(xk) f(xk)ekgkIllustration of the descentproperty of t
6.252 NONLINEAR PROGRAMMING LECTURE 4 CONVERGENCE ANALYSIS OF GRADIENT METHODS LECTURE OUTLINE • Gradient Methods - Choice of Stepsize • Gradient Methods - Convergence Issues
Domain:
Source:
Link to this page:
Please notify us if you found a problem with this document:
{{id}} {{{paragraph}}}