A GRADIENT-BASED METHOD FOR THE OPTIMIZATION OF NEURAL NETWORK PARAMETERS

Authors
  1. DeCruyenaere, J-P.
Corporate Authors
Defence Research Establishment Suffield, Ralston ALTA (CAN)
Abstract
Un-manned air vehicles stand to benefit from the inclusion of computing structures that are based on artifical neural networks. The use of such a network in the role of an adaptive air vehicle stabilizer is felt to be a good choice for investigating the feasibility of this approach. Before such a system can be implemented, a suitable network training algorithm must be developed. This paper presents the development of a gradient-base minimization algorithm for unconstrained, continuous multivariate functions. This technique was based on a set of heuristics which suit the algorithm to functions similar to those encountered during the optimization of multilayered artifical neural networks. A series of trials demonstrate the good overall convergence of the gradientbased method in comparison with other appropriate minimization techniques, namely backpropagation and a modified version of Newton's method.
Report Number
DRES-M-1354 — Memorandum
Date of publication
15 Mar 1992
Number of Pages
77
DSTKIM No
92-02548
CANDIS No
105415
Format(s):
Hardcopy;Originator's fiche received by DSIS

Permanent link

Document 1 of 1

Date modified: