Abstract
The problem which motivated this research was that of stationary target identification (STI) with millimeter wave seekers in a heavy clutter environment. While investigating the use of neural networks to perform target discrimination phase for ST1 problem, we began to search for a method to reduce the computational overhead associated with training a neural network to recognize low probability events. Our search yielded the development of a likelihood ratio weighting function (LRWF), which is very similar to the weighting function used in importance sampling techniques employed in the simulation of digital communication systems. By incorporating the LRWF into the backpropagation algorithm, we were able to significantly reduce the computational burden associated with training a neural network to recognize events which occur with low probability. This reduction in computational overhead is realized due to the reduction in the size of the data sets required for training.
Date of this Version
May 1993
Comments
Page 59 is missing from original document as well as copy held in the Siegesmund Engineering Library of Purdue University.