Author: Robert Hyatt
Date: 10:57:05 08/28/00
Go up one level in this thread
On August 28, 2000 at 13:35:15, Albert Silver wrote:
>On August 27, 2000 at 22:05:14, Robert Hyatt wrote:
>
>>On August 27, 2000 at 12:08:19, Ed Schröder wrote:
>>
>>>On August 27, 2000 at 10:05:17, Robert Hyatt wrote:
>>>
>>>>On August 27, 2000 at 00:28:08, Ed Schröder wrote:
>>>>
>>>>>On August 26, 2000 at 23:43:34, Robert Hyatt wrote:
>>>>>
>>>>>>On August 26, 2000 at 23:06:10, Ed Schröder wrote:
>>>>>>
>>>>>>>On August 26, 2000 at 21:08:37, Robert Hyatt wrote:
>>>>>>>
>>>>>>>>On August 26, 2000 at 11:05:24, walter irvin wrote:
>>>>>>>>
>>>>>>>>>On August 26, 2000 at 09:21:15, Robert Hyatt wrote:
>>>>>>>>>
>>>>>>>>>>On August 26, 2000 at 09:16:50, Uri Blass wrote:
>>>>>>>>>>
>>>>>>>>>>>I tried this position from the game century3-crafty17.11
>>>>>>>>>>>
>>>>>>>>>>>[D]rnbqk2r/pp1p1ppp/4pn2/4P3/1b1N4/2N5/PPP2PPP/R1BQKB1R b KQkq - 0 1
>>>>>>>>>>>
>>>>>>>>>>>Crafty17.13 played 6...Qc7
>>>>>>>>>>>
>>>>>>>>>>>I gave my Crafty17.11 as engine for Hiarcs with 128 Mbyes hash time control of 4
>>>>>>>>>>>hours/60 moves
>>>>>>>>>>>(I guess that this time control is similiar to 2 hours/60 moves on the alpha)
>>>>>>>>>>>
>>>>>>>>>>>My Crafty17.11 changed its mind in the last second at depth 12 from 6..Qc7(0.15
>>>>>>>>>>>pawn advantage) to the better move 6...Nd5(0.16 pawns advantage).
>>>>>>>>>>>
>>>>>>>>>>>I am interested to know crafty17.13's opinion.
>>>>>>>>>>>Can crafty17.13 avoid 6...Qc7 at depth 12?
>>>>>>>>>>>
>>>>>>>>>>>Uri
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>It avoids it by depth=10. The problem was that the mode we were using forced it
>>>>>>>>>>to play Qc7 as it was a book move. Somehow we followed a very rarely played
>>>>>>>>>>line, but until I get the log files from Graham I can't tell exactly why. I
>>>>>>>>>>did notice a pretty significant CAP score that might have pulled it down that
>>>>>>>>>>line erroneously...
>>>>>>>>>
>>>>>>>>>what was the speed relitive x86 for the alpha you used .would it have been
>>>>>>>>>faster if you could have gotten similar hardware that deep junior had at
>>>>>>>>>dortmund ??what nps was you seeing on the alpha .
>>>>>>>>
>>>>>>>>
>>>>>>>>We were doing about 400K nps. My quad xeon does about 1M, for reference.
>>>>>>>>
>>>>>>>>Speed wasn't the issue here however, it was simply following a bad book line
>>>>>>>>until it was too late...
>>>>>>>
>>>>>>>But 6..Qc7? came not from the book but from the Crafty engine. In this
>>>>>>>game the Crafty book did not work.
>>>>>>>
>>>>>>>Ed
>>>>>>
>>>>>>
>>>>>>No... it came from the book. We were using "book random 0" which says to
>>>>>>search (for a short time, normally) all the book moves, and play the one with
>>>>>>the best score. Qc7 was in the book, but the score fails low pretty quickly.
>>>>>>Then, after resolving the fail low, it decides "hey, this is bad, I am not
>>>>>>going to play a book move at all." It then does a normal search, but it had
>>>>>>used too much time and the search wasn't very deep....
>>>>>>
>>>>>>The book was actually working, but "book random 0" changes the way it looks,
>>>>>>and if you don't know, you could assume it is doing a normal search, unless
>>>>>>you see the line "searching only the following moves: {a b c d etc}
>>>>>>
>>>>>>I think Qc7 was the third most popular move in that very narrow (and bad) book
>>>>>>line...
>>>>>
>>>>>This is what I remember from our game: 1.e4 by Rebel and Crafty immediately
>>>>>replied with 1..c5 from the book. Then 2.Nf3 and then Crafty started to
>>>>>think shortly (5-10 seconds) and the second Crafty move was played. This
>>>>>behaviour (5-10 secons for the move) continued till move 5 or 6 and then
>>>>>Crafty started to use the normal 2-4 minutes for its moves. 6..Qc7 came
>>>>>at the last moment after a 2-4 minutes search. After 6..Qc7 Rebel was also
>>>>>out of book.
>>>>>
>>>>>I don't know that much about Crafty but everybody was saying Crafty was
>>>>>out of book after move 1. Also Graham said the above Crafty behavior was
>>>>>the same in game 1 (implied Crafty playing without book in game-1).
>>>>>
>>>>>Hope this gives you sufficient information to make up your mind what
>>>>>really went wrong with the book.
>>>>>
>>>>>Ed
>>>>
>>>>
>>>>That sounds like what I would expect here... If the score for Qc7 dropped
>>>>enough, it will refuse to play it as a book move and do a search. I had not
>>>>factored in the problem with a fail low at the end of the book search, which
>>>>could use over 1/2 of the total time allocated to this move. But I will cover
>>>>this hole quickly, of course. :)
>>>>
>>>>Qc7 was played several thousand times, however, which makes it surprising it
>>>>isn't in your book.
>>>
>>>Qc7 is a losing move just trash it from your book. Rebel had to find 7.exf6
>>>and sacrifice the rook on a1. I was happy Rebel did. After that the game was
>>>over. I think Jeroen will add 6..Qc7? 7.exf6 to the book so Rebel can find
>>>7.exf6 also on short time controls. After 7.exf6 things are easy to find.
>>>
>>>Ed
>>
>>
>>The sad thing is that Crafty has played this vs GM players 10 times now. And
>>it _won_ all ten games with black. That is why book learning had not already
>>culled this silly variation. :)
>
>What were the time controls of these games? Also, does it base its learning on
>the game result or the eval of the position after leaving its book? The latter I
>presume, but you mentioned it won 10 games hence my doubt.
>
> Albert Silver
Crafty uses the scores from the first 10 moves out of book. It looks at
these to get an idea of whether things were good, bad, bad but getting better
or good but getting worse. It then folds in the time it spent on those
searches, plus its rating and the opponent's rating, and comes up with a
'score' it thinks reflects the book position at the end of the book line.
It can be adjusted so that if the score is good, it will play the same opening
again, and if the score is bad, it will never play the opening again. I don't
run it like this on ICC, although if it loses a game, it definitely won't play
the last few book moves in the line again, but it will give other variations
in that line a try...
The time controls vary, obviously. But to compensate the search depth is used
to 'weight' the scores, with deeper depths being weighted higher than shallower
depths. IE one standard game is trusted more than 2-3 blitz games, easily...
>
>>
>>Learning doesn't cure all book woes unless the opponents punish it so that
>>it learns it was punished. :)
>>
>>Creating a good book (automatically) is a challenge. If you only use GM
>>games, the book is too narrow. If you use a bigger game collection, you get
>>garbage moves...
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.