Author: Jay Scott
Date: 14:57:12 02/25/98
Go up one level in this thread
On February 25, 1998 at 15:19:50, Don Dailey wrote: >But now a good question is: What does the probability measure? Is it >the probability that the computer will win? I think the correct >interpretation should be "the probability that the position is a won >position." A dead draw should be considered as 50% of a win, or 50% >probability of winning. If we say it's the probability that the >computer will win, then it's completely ambiguous, because we do not >know what assumptions to make about the strength of the opponent! If >Cilkchess is down half a pawn against most humans then it's chances of >winning are still greater than 50%. Another possible interpretation >is the probability of winning against an equal opponent (whatever that >is!) A crucial question, in my view! What your evaluation probability "really" measures depends on how you constructed the evaluation. Usually the game-theoretic values of positions aren't available, so you have no way to construct an evaluator which measures the true probability that a position is a win given what the program knows about it. That still makes sense as an ideal to strive for, though. If you construct the evaluator from self-play, then the probability will be the program's probability of winning against itself from that position. If you construct it from games against a variety of opponents, like KnightCap, then it's the probability of winning against some notion of the average opponent. I'm not happy with either of those possibilities. I would like the probability to be more meaningful. I would like some kind of opponent model, fancy or crude, so that I could say that it was the probability of winning this game, against this opponent. A crude model might contain no more information than the opponent's rating, which may or may not be useful information to a real program. A fancy model could take into account whether the opponent is a human or a program (crafty does this) and any other helpful info you can think up. If you don't know anything about the opponent then you have to fall back on an empty opponent model anyway, which is equivalent to having no special opponent model, so this isn't too big a deal. Another point, to repeat something I've said before, is that you don't really want to measure the probability of winning from a position; you want to measure the utility of reaching a position, which depends on the probability of winning and the probability of drawing. Counting a draw as half a win is not always right; it doesn't distinguish between positions which are dead drawn and positions which are dynamically equal, and sometimes you care. Jay
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.