Author: Don Beal
Date: 10:53:31 10/05/98
Go up one level in this thread
On October 02, 1998 at 18:05:41, jonathan Baxter wrote: >>BTW, we applied the same method to Shogi, and learnt piece values there. >>That's directly useful, because Shogi doesn't have a standardised set of >>values that programmers can pick up and use. > >I saw one of your papers on this (and liked it :). Hey, we're going to get along real well :-) :-) >One question I have is: the piece values drift >an awful lot, with the drift seeming to reduce only with the decrease in step >size. I was wondering, if you keep the step size constant (but small) and run >for 100's of 1000's of games, do the pieces swap their relative ordering a lot >or do they finally settle to a constant relative ordering? With fixed learning rates (aka step size) we found piece values settle to consistent relative ordering in around 500 self-play games. The ordering remains in place despite considerable erratic movements. But piece-square values can take a lot longer - more like 5000. The learning rate is critical - it has to be as large as one dares for fast learning, but low for stable values. We've been experimenting with methods for automatically adjusting the learning rate. (Higher rates if the adjustments go in the same direction, lower if they keep changing direction.) The other problem is learning weights for terms which only occur rarely. Then the learning process doesn't see enough examples to settle on good weights in a reasonable time. I suspect this is the main limitation of the method, but it may be possible to devise ways to generate extra games which exercise the rare conditions.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.