CLASSIFICATION OF SIMULATES SEQUENCES USING A FIR NEURAL NETWORK: EXPERIMENTAL RESULTS

PDF

Authors
  1. Arsenault, N.
  2. Stevenson, M.
Corporate Authors
Defence Research Establishment Atlantic, Dartmouth NS (CAN);New Brunswick Univ, Fredericton NB (CAN) Center for Research in Engineering and Applied Science
Abstract
This report examines the experimental results achieved by the FIR Neural Network on data sets which simulate underwater acoustic transient signals. A software package to simulate FIR neural networks has been developed in the C programming language. The effects of several user controlled options will be given along with suggestions on how to choose these and the network parameters (number of layers, number of nodes, etc.) which define the structure of the FIR neural network. Experiments have been conducted using three separate data sets (each data set contains a training and test set composed of a collection of spectrograms obtained from the Short Time Fourier Transform): a Three Tone Data Set (TTDS), the Simulated DARPA Data Set (SDDS), and a Multipath Data Set (MPDS). The effects of four user controlled options (weight initialization, weight update frequency, choice of desired response, and error scaling sequence) on the classification rate and learning speed is studied in order to determine the best configuration in which to set up the FIR software package. A suggested procedure to determine the transient time of occurrence is presented along with some results obtained on the SDDS. The performance of the FIR neural network is shown to be superior to that of static neural networks on both the TTDS and the SDDS. Finally, conclusions and suggestions for future work are given.
Keywords
Finite impulse response;Dynamic neural networks;Transient signals;Sonar Information Management;Finite-duration Impulse Response (FIR)
Report Number
DREA-CR-97-459 — Contractor Report
Date of publication
01 Apr 1995
Number of Pages
73
DSTKIM No
98-01942
CANDIS No
508789
Format(s):
Hardcopy;Document Image stored on Optical Disk

Permanent link

Document 1 of 1

Date modified: