Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / learning from database
- - By fritz10thinker (*) [ae] Date 2009-05-04 22:37
I decided I'll use fritz10 by checking the "learning from database" using the games  I've played with rybka engines and others, is this a good way to strenghten the engine and ELO?
Parent - By Mark Mason (***) [gb] Date 2009-05-05 06:32

First a quick clarification of what this function is. Learn from Database is not about strengthening an engine - it is a feature you can use to enhance your openings book.  It looks at the games of the chosen database and pretends it has played them itself (which of course it hasn't). On the basis of the results it changes the weightings of individual variations - so your book weightings are updated as though you have played all of the games. (the 'weighting' of an existing move makes it more/less likely to be chosen. This is shown in the Prob / % column. The value is from -125 (very unlikely to be picked) to +125 (very likely to be picked). You can change these manually to any value but this is what the automatic book learning does based on whether games are won or lost).

Because learn from database adds the statistics from a selected database without actually importing the moves into the tree - when applying the learning you can either use the database you have made your book from, or a subset of it, or even a completely different one  (If  a different one then the program "learns" from the experience of other players). One idea is to use a smaller subset of games as your Learn Base eg. it could be selection of high quality games from high elo players. Kevin Frayer explains this well in his opening book development & tuning blogs - 

When using this feature you can set a number of learning parameters:
Wins: Only won games are used, the program increases the weights of white moves if White won
and black moves if Black won.
Losses: Only losses are considered. If White lost, the white moves are reduced, and if Black lost, the black moves are reduced.
White: The program evaluates the games from the point of view of White only. The weights for white moves are increased if White won, and decreased if White lost.
Black: The weights for black moves are increased if Black won, and decreased if Black lost.
Player: Only considers the games of one specific player in the database.
These five options can be combined in any way you like. You can also restrict the import to a certain number of games in the database by selecting "Games" at the top of the dialog box.

My conclusion would be that yes this a very powerful and useful function to enrich you book, but you have to be careful to select the right games to learn from.

Hope this helps.

Up Topic The Rybka Lounge / Computer Chess / learning from database

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill