scotbri@rosevax.Rosemount.COM (Scott Brigham) (04/18/91)
I have been attempting to duplicate the truck backer-upper experiment of Nguyen and Widrow, following the information in their papers and in the Application Domains section of "Neural Networks for Control" (ed. Miller, Sutton and Werbos). I am doing this chiefly to gain confidence in the techniques and my understanding of them. There is enough info available on what they did so as to provide me with a lot of 'hand-holding' as I go through it. The problem I am having is in the first part -- training a neural network to emulate the truck dynamics. I run it through 30,000 ten-step runs with random starting points, etc, etc as is reported in the literature. But the errors are quite large. The truck moves in increments of 0.2 meters and my network's estimate of the next position is off by about 2 meters in x and y on the average. The actual numbers coming out of the network look 'good' in that they are 'close' to the target value (for example, the target value is .3077 and my network output is .3201) but this difference (.0124) represents .62 meters if my output range of +/- 1 is scaled for the +/- 50 meter grid that the truck moves around on. This is my first attempt at doing something where my target values aren't fullscale (pattern classification) and where the training set is not a fixed set that is shown repeatedly. If anyone has any ideas, I'd really appreciate hearing from you. I haven't explained the whole thing real well, but I hope you get the general idea. ============================================================================== scotbri@rosemount.com Scott Brigham When speaking Truth to Power Rosemount, Inc. don't forget your duck! 12001 Technology Drive Eden Prairie, MN 55344 (612) 828-3234 (voice) USA (612) 828-3259 (FAX) ==============================================================================