Presentation 1998/3/20
SUPERLINEAR AND AUTOMATICALLY ADAPTABLE CONJUGATE GRADIENT TRAINING ALGORITHM
Peter GECZY, Shiro USUI,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Conjugate gradient optimization methods (e. g. BP with momentum) are among the most used MLP network training techniques. Difficulty with the use of the methods is due to the necessity of finding the appropriate values of learning parameters (e. g. learning rate and momentum). This paper introduces a novel algorithm with automatically and dynamically adaptable learning parameters. Learning rate and momentum term are optimally determined at each iteration in a single-step calculation. The newly proposed algorithm has the same computational complexity as BP with momentum, however, it is convergent with superlinear convergence rates, i. e. the fastest convergence rates for first order techniques. Apart from the theoretical justification the simulation results indicate superior performance of the proposed algorithm over the standard BP with momentum term.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) first order optimization / conjugate gradient / line search subproblem / adjustable parameters
Paper #
Date of Issue

Conference Information
Committee NC
Conference Date 1998/3/20(1days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair
Vice Chair
Secretary
Assistant

Paper Information
Registration To Neurocomputing (NC)
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) SUPERLINEAR AND AUTOMATICALLY ADAPTABLE CONJUGATE GRADIENT TRAINING ALGORITHM
Sub Title (in English)
Keyword(1) first order optimization
Keyword(2) conjugate gradient
Keyword(3) line search subproblem
Keyword(4) adjustable parameters
1st Author's Name Peter GECZY
1st Author's Affiliation Department of Information and Computer Sciences, Toyohashi University of Technology()
2nd Author's Name Shiro USUI
2nd Author's Affiliation Department of Information and Computer Sciences, Toyohashi University of Technology
Date 1998/3/20
Paper #
Volume (vol) vol.97
Number (no) 624
Page pp.pp.-
#Pages 8
Date of Issue