This preview shows pages 1–7. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: Lecture 25: Efficiency of Proximal Bundle Method Variable Metric Subgradient Methods April 23, 2007 Lecture 25 Outline ComputingSubgradients Simplified Proximal Bundle Method Convergence Rate of the Method Subdifferential of a Composite Function Some issues in optimization Convex Optimization 1 Lecture 25 Comment onSubgradient Let f be a convex function with dom f and bounded subgradients over a set X . Let x X and > Let y X be such that k y x k 2 L with L = max {k s k  s f ( z ) , z X } Any vector s f ( x ) is ansubgradient of f at y , i.e., f ( x ) f ( y ) Proof : We have for any z R n , f ( y ) + s T ( z y ) = f ( y ) f ( x ) + f ( x ) + s T ( z x + x y ) f ( x ) + s T ( z x ) + [ f ( y ) f ( x )] + k s kk x y k f ( z ) + [ f ( y ) f ( x )] + L k x y k Convex Optimization 2 Lecture 25 where the last inequality follows from s f ( x ) . By convexity of f , we have f ( y ) f ( x ) g T ( y x ) for some g f ( y ) By subgradient boundedness, it follows that f ( y ) f ( x ) L k x y k . Hence, for all z R n f ( y ) + s T ( z y ) f ( z ) + 2 L k x y k f ( z ) + showing that s f ( y ) Convex Optimization 3 Lecture 25Subgradient for Dual Function q ( ) = inf x X { f ( x ) + T g ( x ) } Suppose x X attains the dual value q ( ) with an error > , i.e., x X is such that q ( ) + f ( x ) + T g ( x ) (1) Then g ( x ) q ( ) Proof : We need to show that q ( ) + g ( x ) T (  ) q ( ) for any dom q . Using relation (1), we obtain q ( ) + g ( x ) T (  ) f ( x ) + g ( x ) T  + g ( x ) T (  ) = f ( x ) + T g ( x ) inf x X { f ( x ) + T g ( x ) }  = q ( ) Convex Optimization 4 Lecture 25 A Proximal Bundle Method Using Subgradients This method is also given by Kiwiel [2000]. It is an adaptation of the proximal method of Kiwiel proposed in 1995 What has been improved: Subgradients are used instead ofsubgradients Stopping rule is based on actual error in the linearized model Scheme for bundle update is simpler [does not use error updates] Scheme for weight update is simpler [no need for a factor 10] As a result: There are fewer parameters to choose The overall procedure is simpler [less computationally demanding] The algorithm is easier for analysis Convex Optimization 5 Lecture 25...
View
Full
Document
This note was uploaded on 08/22/2008 for the course GE 498 AN taught by Professor Angelianedich during the Spring '07 term at University of Illinois at Urbana–Champaign.
 Spring '07
 AngeliaNedich

Click to edit the document details