A New Global Optimization Method and Supervised Learning of Multilayer Neural Networks

Jinhui CHAO, Wijak RATANASWAN, Shigeo TSUJII

Summary :

This note presents a new global optimization method and derives a learning schema based on the method for multilayer artificial neural networks. The schema consists of (1) pasting" the admissible region in Rⁿ to a n-D torus Tⁿ and smoothly connecting the potential function at the boundary; (2) global searching along the flow of a nonvanishing vector field on the compact smooth manifold Tⁿ. This flow is featured by the ability of automatically passing through distinct local minima one after another along the negative/positive gradient field. It has also a unit norm everywhere on the Tⁿ, so the searching speed will not slow down in the neighborhood of critical points.

• Publication: IEICE TRANSACTIONS on transactions Vol.E73-E No.11 pp.1796-1799
• Publication Date: 1990/11/25
• Type of Manuscript: Special Section LETTER (Special Issue on 1990 Autumn Natl. Conv. IEICE)
• Category: Neural Networks

Authors

Jinhui CHAO
Wijak RATANASWAN
Shigeo TSUJII

Keyword