Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: why probability evaluation is better than pawn evaluation

Author: Dave Gomboc

Date: 17:09:53 01/18/02

Go up one level in this thread


Computer programs still report their assessments in centipawns, but internally
they have a notion of these things.  At least, as far as I can tell, the best
ones do.

Dave

On January 17, 2002 at 17:21:59, Jay Scott wrote:

>Too many of the posters to this thread do not get it. Pawns add; a pawn
>evaluator typically considers that winning two pawns is twice as good as winning
>one. Probabilities do not add; being ahead two pawns does not produce twice the
>probability of winning as being ahead one.
>
>The fundamental difference is that a probability evaluator, if it takes full
>advantage of probability theory, must account for feature interactions. In a
>pawn evaluator, entirely hand-made, feature interactions are handled ad hoc by
>the programmer saying, "Gosh, these two things are related [king safety and the
>endgame, say], I'd better put in a term to account for it. Hmm, let's try this
>value and see how it works." In a probability evaluator, whose values are filled
>in by collecting large amounts of statistical data, feature interactions are
>detected by statistical analysis and accounted for mathematically. It's
>principled instead of ad hoc, and as evaluators become more complicated,
>organizing principles for them become more important.
>
>Because of feature interaction, and because it has to be benchmarked to its own
>measured results, a probability evaluator is more complex to create than a pawn
>evaluator. The payback is that you have a mathematical theory that tells you how
>to extract the most information from it. If it includes all the features it
>needs, and you have collected enough data to accurately model the feature
>interactions, then it is a theorem that the evaluator will be accurate. Having
>math on your side can only be good.
>
>A concrete example:
>
>In an otherwise even position, it's smart to win a pawn at the cost of a
>positional disadvantage, unless the disadvantage is too big.
>
>Suppose you've won the pawn and accepted the disadvantage, and now you have a
>chance to win a second pawn at the cost of further positional damage. Everything
>depends on the specific position, of course, but if the one pawn is already
>enough to win it makes sense to consolidate your position instead of winning the
>second pawn.
>
>A pawn evaluator (unless it has extra smarts) compares its value of a pawn to
>its value of the positional disadvantage, and chooses the bigger one--whether it
>has already won one pawn or not. That will sometimes be a mistake. A human does
>not do that. A probability evaluator does not do that either: if it is correctly
>constructed, according to probability theory, it will understand that winning
>extra material produces ever-smaller increases in winning probability, and that
>accepting a greater positional disadvantage produces ever-greater chances of
>allowing a perpetual or even underestimating an attack and losing. If the
>probability evaluator includes good chess features and has good data on them,
>then it will correctly weigh the odds and choose the move that gives it the
>statistically best chance of winning--genuinely the best given its knowledge,
>according to an unimpeachable mathematical theory. You can't ask for more than
>that.



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.