Author: pavel
Date: 19:19:12 09/19/00
Hi,
I have been thinking about a learning idea for chess programs for some time,
and brought this idea once before in this forum, but probably didnt get enough
feedback, or explanation werent clear enough.
here is what I wanted to know....
lets make a chess program (program X) analyze a set of games which contains 3000
positions, program X will analyze for each position for,lets say, 10secs; other
then the moves that are already in the book.
It will store the evaluation value, along with the 'depth' and 'next best move'
it thought of, for each position in a *bat or *txt file,
Later, when it comes to that position again, it will start calculating from
where it left, eg, if it left in depth 6 in e2e4 it will start analyzing from
then on, since the depth and the pV will be stored.
It will make it search faster, or(and) avoid searching bad positions.
lets say i come to the same position in a game where i left e2e4 in depth 6 i
search from then on and stoped at depth 8, so I erase the previous evaluation in
that position, and save the new value (depth and PV), so when I encounter it
again, it will start from where it left (depth 8), and on and on.....thus making
it faster to search the next time it come to the same position....
I know I know, the possibility of a chess program to encounter the same move
twice is very rare.... probably there are zillions of moves in chess.
but here 3000 position was just an example, in real life it can be use to
analyze millions of position.
eg, the positions that comes right after the book is more important, and/or more
likely to encounter, then other moves. so more importance will be given (or
analyzed) those moves.
there can be a "universal" learning file for a program X which can be downloaded
and use by others, or one can just produce its own learning file.
now, I think, the real challenge is how to use the learning file (more likely
its huge) in a game of chess. If the file is 10mb size then its gonna be a lotta
stress to the program to put it in hash. So it is not a very good idea.
the other option is to use the learning file, like the one use by TableBase.
In this way it will be faster.
its easy to come up with a idea but hard to prove it or do it.
its probably really hard to code your program to use learning file like TBs.
so thats why my questions to chess programmers.
what are the pros and cons, of such learning feature?
what are the possibilities?
Is it possible at all?
thanks for your feed back...
pavel
ps, ofcourse it mainly goes to the programmers, but others feedbacks are
welcome.
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.