Author: Robert Hyatt
Date: 15:43:24 02/24/99
Go up one level in this thread
On February 24, 1999 at 11:43:12, Vincent Diepeveen wrote: >On February 24, 1999 at 10:19:56, Robert Hyatt wrote: > >>On February 24, 1999 at 08:15:09, Vincent Diepeveen wrote: >> >>>On February 24, 1999 at 06:57:57, blass uri wrote: >>> >>>> >>>>On February 23, 1999 at 18:53:28, Dann Corbit wrote: >>>> >>>>>On February 23, 1999 at 18:47:36, Vincent Diepeveen wrote: >>>>>>On February 23, 1999 at 14:53:11, Dann Corbit wrote: >>>>>[snip] >>>>>>>I don't think this is evidence of "cooking" -- more likely learning. Quite >>>>>>>likely, they have already analyzed this position set in their database and >>>>>>>therefore know the best responses for each position. >>>>>>> >>>>>>>Rather than indicating something underhanded, it probably just means that they >>>>>>>have run extensive testing in this area. Crafty would do the same thing. If I >>>>>>>ran crafty at very long time controls, it would update the learning data file >>>>>>>about what to do in these situations. Is crafty cheating? Certainly not. Not >>>>>>>'cooking' anything either. >>>>>>> >>>>>>>I don't see any evil conspiracy here. That does not mean there is nothing >>>>>>>afoot. But I doubt it. >>>>>> >>>>>>f5.32 gets shipped with Learning files already compiled for those >>>>>>positions and another few thousands for auto232 at the auto232 version, >>>>>>is that what u mean? >>>>>I only suspected something of that nature. But it certainly makes sense. If I >>>>>were a chess program manufacturer, I would have best move entries for thousands >>>>>of known difficult positions. Some of the Crafty and Covax postions turn up a >>>>>lot in games, for instance. >>>> >>>>I do not think that they used learning files already compiled for these >>>>positions >>>>because if they did it then they did not need to calculate moves when they >>>>played. >>>> >>>>They probably changed the evaluation function to have better results in the nunn >>>>match. >>>> >>>>Uri >>> >>>Well first look 'what kind of learning' do chessprograms have. >>>Chessprograms have only learning for bookmoves, and those are played at >>>once. All experiments with position learning at non-book moves failed so far, >>>commercially seen. >> >> >>I don't agree. I have spent more hours debugging stuff caused by forgetting >>about crafty's "position learning" than anything else. Because after a deep >>search it stores the result as a permanent hash entry that goes into a file. >>And when you play again, these 'learned positions' are loaded before doing >>anything. In my case, it wouldn't be hard to 'cook up something' for (say) >>the Nolot test to get 'em all right. And I would still do normal searches, >>but the 'scores' would be doctored by the permanent hash entries. > >You are right here, but in the case of fritz it is something else that >explains it, it is called hardcoding things in preprocessor, >which is way simpler than position learning, and has also a much bigger >impact. > >>> >>>So any kind of 'learning file' in the Nunn positions cannot be true unless >>>they play the move within a flash of a second (book move). >>> >>>The only thing that gets done is HARDCODING in the piece square tables >>>the position. Fritz3 was clearly provable cooking BK test set, >>>so much even that the b2-b4 position it played always b2-b4 if the >>>pawn structure in the centre matched. It even did it >>>when putting whites pawn from g2-g4 and blacks pawn from h7-h6. Then b4 is a big >>>blunder as black gets 2 passed pawns after Qc8, but still it played it. >>>all other chessprograms saw this within seconds, so would never play b4, >>>but fritz did. It just didn't like b4 initially, so i guess Frans gave it >>>so much bonus that it then always wanted to play b4, otherwise it got >>>death penalty... >>> >>>To quote someone: "the people just test positions at it to see how strong >>>programs are, so they ask for it". I'll not reveal his name, that would >>>not be nice. >>> >>>I think he is completely right however. >>> >>>Most programs, including crafty solve win at chess position 2, >>>at low depths just seeing that pawns get to 3d rank, where such >>>kind of evaluation is very dangerous. >> >> >>I don't think this is dangerous. My 'evaluation' is not "just a pair >>of connected passers on the 6th". It is more complex. And it was _not_ >>done for WAC2. It was done because in bullet games it would be well >>ahead and then lose when it allowed two connected passers to reach the >>point where a rook can't stop them. > >It is dangerous in this case, as king is in quadrant. > I don't think it is as dangerous as it seems... because I have specific code to turn this off if the king gets too close. With a reasonable search depth, the king can't get there in time.. >>This has saved a _lot_ more games, and _won_ a lot more games, than it has >>lost. >> >> >>> >>>But abou the nunn positions: the only thing that can be done >>>in order to play these positions ok with >>>Fritz is to have new code implemented in the program that fills the piece >>>square tables. >>> >>>This is a clear case of cooking. >>> >>>Of course Nunn selected in the past only more or less closed positions, >>>as he got paid by chessbase. Later they even improved play here by doing >>>this. >>> >>>Greetings, >>>Vincent
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.