A TECHNIQUE FOR OPTIMIZING A NEURAL NETWORK

Authors
  1. Barton, S.A.
Corporate Authors
Defence Research Establishment Suffield, Ralston ALTA (CAN)
Abstract
The report describes a matrix-based method for optimizing the parameters in a neural network. With this technique, the optimum set of weights and biases is found for the output side of a network with a single hidden layer of neurons, given any set of weights and biases for the input side of the hidden layer. All the input patterns are included in a single calculation cycle. A simple numerical minimization procedure is used to iteratively optimize the weights and biases on the input side of the hidden layer. Several test problems, using both continuous and binary inputs, have been solved, confirming the validity of this matrix optimization technique. In all of the applications tested, the matrix technique optimized networks much more rapidly than the backpropagation method. Based on this work, it should be possible to use the matrix optimization method to define neural networks that can generate the control signals of an autopilot.
Report Number
DRES-M-1351 — Memorandum
Date of publication
15 Dec 1990
Number of Pages
35
DSTKIM No
91-01469
CANDIS No
68676
Format(s):
Hardcopy;Originator's fiche received by DSIS

Permanent link

Document 1 of 1

Date modified: