[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: computer-go: A problem with understanding lookahead
>In go it's much harder to make an openings book as we
>all very well realize.
>
>A chessprogram without an openingsbook would never beat
>a professional chessplayer as they are directly killed
>strategically.
And this is absolutely correct.
The fact that the opening book is so important in Chess, should make
it painfully obvious that Chess programs still have very far to go.
Obviously, if each ply of search improves a program so much (and this
is a proven fact) and they STILL cannot figure out what opening moves
are best, then we need to readjust our viewpoint of how simple and
easy it is to evaluate Chess positions.
From: Darren Cook <darren@xxxxxxxxxxxxxxxxx>
>In go it's much harder to make an openings book as we
>all very well realize.
>
>A chessprogram without an openingsbook would never beat
>a professional chessplayer as they are directly killed
>strategically.
There was a paper presented by Ernst Heinz at CG2000 on self-play in chess
[1], which showed the value of each additional ply of search (and also
showed diminishing returns).
He used an opening book. It would be very interesting to see the results of
self-play with different sizes of opening books rather than different
search depths. And to repeat Heinz's experiment without the opening book.
Any students out there looking for a research project? :-).
It would give us some hard data to use in the arguments about relative
importance of search and knowledge.
Darren
[1] E.A. Heinz.
New self-play results in computer chess.
In 2nd International Conference on Computers and Games, I. Frank and
T.A. Marsland (eds.), to be published (LNCS series by
Springer), December 2000.
Ernst Heinz experiment involved more games that had ever been done
before. It was an excellent paper, and he always writes high quality
papers.
Ernst used the opening books to provide variety. Without this, you
would just get a lot of the same games being played over and over
again! It could still be done but you would have to provide a
mechanism to vary the games, and this almost always negatively impacts
the quality of the games.
Don