Author: Daniel Shawul
Date: 06:12:04 01/07/05
Go up one level in this thread
On January 06, 2005 at 18:07:47, Eric Oldre wrote:
>On January 06, 2005 at 14:33:33, Norm Pollock wrote:
>
>>On January 06, 2005 at 13:20:27, Eric Oldre wrote:
>>
>>>I have added the ability to add opening books to my latest version of Latista.
>>>I'll release it once I create a decent opening and add a couple other smaller
>>>features.
>>>
>>>So far I created a book my parsing the gm2600.pgn file and saving all positions
>>>which occur at least 5 times.
>>>
>>>This is certainly much better than no book at all but leaves much much room for
>>>improvment I'm sure.
>>>
>>>What general advice do you book makers have for me?
>>>
>>>Thanks,
>>>Eric
>>
>>First, may I suggest that you have two opening books, one for white's moves and
>>one for black's moves. This way you can "feed" pgns into white's book that have
>>white winning (and drawing). Likewise for black's book, you can "feed" pgns into
>>black's book that have black winning (and drawing).
>
>currently i store the number of white wins, black wins, and draws for each
>position in the book. I can see how it would probably be better from a
>space/memory perspective to create 2 separate.
>
>
>>
>>Second, I would also use more recent pgns than in gm2600.pgn because current
>>over-the-board players know what the latest changes are in openings, and they
>>have the benefit of computer analysis. (Btw, I have made a pgn collection of
>>recent games that is available at crafty-chess.com in the user download
>>section.) Also limiting the book to games with both players elo 2600+ will mean
>>a small book, unless you allow the number of position occurrences to be very
>>small (see next paragraph).
>
>Thanks Norm, I'll download that tonight!
>
>>
>>Third, no matter how hard you try, bad opening moves are going to be in the
>>opening book. The question is how to limit the number of them. I believe the
>>best and only reliable approach is to have a very high number for the number of
>>occurrences before including that position in the book. I recommend that you use
>>12 occurrences if you are starting from a pgn collection of 60K games.
>>
>>Another approach to quality control is "book learning". I'm not a fan of book
>>learning. All I think it does is adjust the statistics for each position (#
>>occurrences, win-lose-draw) based on the engine's performance. But if your
>>engine is playing an engine 100 elo stronger, why punish the opening because
>>your engine lost. The opening might have been excellent. Likewise, why reward a
>>poor opening because your engine defeated a weaker opponent.
>
>I had thought about trying to do book learning at some point but had thought of
>this also, I'll put it on the "maybe someday" list.
why postpone when you can do it easily. Add one variable learn_value,then
you can change this value when the engine loses/wins(learning) , or at book
loading stage if you want to prefer one move over the other.
daniel
>
>>
>>Another point. Some pgn collections have loads of old games, low elo games,
>>duplicate games, twin games, blitz, rapid, lightning, email, Internet,
>>simultaneous exhibition games and grandmaster short draws. My collection that I
>>mentioned before is one of those that has filtered those types of games out. So
>>be careful.
>>
>>Good luck,
>>
>>Norm
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.