Author: Robert Hyatt
Date: 09:05:47 04/28/03
Go up one level in this thread
On April 27, 2003 at 12:35:31, Drexel,Michael wrote: >On April 27, 2003 at 10:56:45, Robert Hyatt wrote: > >>On April 26, 2003 at 19:15:04, Drexel,Michael wrote: >> >>>Today I started an interesting experiment. >>>A match Chessmaster 9000 against Shredder 7.04 in Chessbase GUI with >>>over-optimistically settings for The King 3.23. >>>With this settings The King engine evaluates his positions almost always as >>>better for himself, except it is completely lost. >> >>You are overlooking a key point. The alpha/beta algorithm is relative to >>the initial score. If a program starts off at +3.0, and another program >>starts off at 0.0 in the same position, the two programs can play >>identically. All alpha/beta tries to do is maximize the raw score at the >>root, whether it is +3 or -3 is irrelevant... > >That point is too simple (even for me), so I didnt overlook it. >If I evaluate a position regarding material, and give my own bishop,knight,etc. >0.1 points more than the opponent bishop,knight,etc. for no apparent reason. >Isnt that over-optimistically? Perhaps or perhaps not. A computer is generally better in complex positions than in simple ones. Such a score bonus would tend to drive it toward more complicated positions where it would do better. Of course this will skew the score, but if you ignore the absolute value of the score all is ok. For example not every program uses 1.0 for pawns. Deep Blue, for example, used 1.28 (a power of 2). That means that being a piece ahead would make the score almost +4.0. So long as you know +4 is good, and don't try to interpret it on some uniform scale, you are "in the ballpark." >Of course this can lead to wrong decisions. >I try to avoid trading of pieces and this might have long-term positional >drawbacks. That's a point. But that is true for _all_ evaluation terms. Protect king safety but get positional problems or avoid positional problems and get an exposed king. Or factor in pawn majorities vs an outpost knight, and so forth. It's not an _exact_ science by any stretch. >Against humans for example this might nevertheless lead to better results (thus >stronger). The probability that a human blunders with many pieces and pawns on >the board is much higher. >It also can lead to better results against something which plays very strong >in the endgame. Correct... > >I cant give many examples at the moment (some are in my database) but Shredder >very often gives high scores for positional advantages which are only >temporarily and will disappear later on. I've seen that. But I have seen it in many other programs as well, including my own... If positional scores were 100% accurate, we'd toss out the tree search anyway. >IIRC Shredder gives advantage for white in the Morra-Gambit of the Sicilian >Defence. If you play on the 'advantage' for white will soon disapear. >Thats what I call over-optimistically. >It doesnt seem to hurt (at least not against other programs). > >Michael > > > > > > > > > >> >> >>> >>>Surprisingly a 30 game match ended: >>>Chessmaster 9000 - Shredder 7.04: 15.5-14.5 (+12 =7 -11) >>>5 min, AMD 2200+, ponder off, Remis.ctg, alternate colours >>> >>>I told The King that the own qeen is better than the opponents qeen, the own >>>rooks are better than the opponent rooks, the own bishops are better than the >>>opponent bishops and so on... >>> >>>Its over-optimistically evaluations dont hurt at all. >>> >>>The evaluations were way off but it nevertheless won the match and played a lot >>>of exciting games although it lacks resistance in worse positions. >>>Chessmaster played very strong in positions it had an advantage. >>> >>>Therefore I think it should be a good idea to have completely different >>>evaluations. >>>For clearly better positions an optimistically evaluation (Shredder obviously >>>has very high scores in such positions) and for worse positions a more realistic >>>evaluation. >>> >>>Michael
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.