Author: Tim Mirabile
Date: 00:29:30 06/20/98
Go up one level in this thread
On June 19, 1998 at 15:55:13, Steven J. Edwards wrote: >I dislike the uniform distribution method mostly because it seems that a lot of >information is being thrown away. I wans't really arguing againt non-uniform distribution, just against trusting final results too much. But uniform distribution has some advantages - it makes it more difficult to predict what your program is going to play. >But I also dislike the ECO/Informator >transcription (I hear these can be had on CD-ROM nowadays) because it is >particularly vunerable to attack by typo hunters who peruse the source input. Ok, lets say I've discovered a typo in ECO B33, second edition, page 214, note 205 in the variation starting with 14.c3. At the end of the line it shows +- when it should be -+. If the program has the whole ECO in its book, and chooses from among the reasonable moves with a fairly flat distribution, what are the odds that I can ever hope to see this line? Only 10-20% of the time would it even play 1.e4. Then after 1...c5 only a 10-20% chance of 2.Nf3, etc. all the way to move 14 after which there are no more branches. Even if I could hope to discover and remember a dozen such lines, I still wouldn't get the chance to use them 9999 times out of 10000. >Additionally, the Informator evaluations are really not scalars and some sort of >multivariate weighting function would have to be used. And so we're back to the >same type of problem as with weighting win/draw/loss results. Right. The typical weighting is to make a draw equal to half a win. But this could be adjusted depending upon the tournament situation. If a draw is needed to clinch first place, the value of a draw can be raised up very close to the value of a win. The opposite can be done if a draw is as good as a loss. Contempt factors can also be used to decide the relative weight of a draw. Can this be done on the fly, or would it require rebuilding the whole book? In the case of the informant rates, a similar function which took into account the tournament situation could be used to weight the "unclear" and "compensation" evals relative to the others. Or the program could be set to tune just these lines, looking to see which side it would do better with.
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.