Presentation 2003/3/10
Behaviors of Globally Suboptimized H_∞-Learning
Kiyoshi NISIYAMA, Koushi OCHISAI,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) Back propagation (BP) method and the extended Kalman filter (EKF) learning algorithms are known as a conventional learning algorithm for layered neural networks. However, their learning speed (or the number of learning iterations ) is strongly affected by the initial values of weight coefficients and thresholds. This paper focuses on a globally suboptimized H_∞-learning algorithm (g-EHF), investigating details of its behaviors on the error surface as well as in the weight space using computer simulations. Our aim in the present paper is to clarify the mechanism for robustness of the H_∞-learning to variations in the initial weight vector.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) H_∞-learning / H_2-learning / BP / weight space / error surface / learning algorithm
Paper # NC2002-143
Date of Issue

Conference Information
Committee NC
Conference Date 2003/3/10(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) Behaviors of Globally Suboptimized H_∞-Learning
Sub Title (in English)
Keyword(1) H_∞-learning
Keyword(2) H_2-learning
Keyword(3) BP
Keyword(4) weight space
Keyword(5) error surface
Keyword(6) learning algorithm
1st Author's Name Kiyoshi NISIYAMA
1st Author's Affiliation Department of Computer and Information Science, Iwate University()
2nd Author's Name Koushi OCHISAI
2nd Author's Affiliation Department of Computer and Information Science, Iwate University
Date 2003/3/10
Paper # NC2002-143
Volume (vol) vol.102
Number (no) 729
Page pp.pp.-
#Pages 8
Date of Issue