> From: Don Dailey <[EMAIL PROTECTED]> > > On Sun, 2008-10-12 at 08:08 -0700, terry mcintyre wrote: > > Yes, there are analogies. The databases of games in Chess include many > high-quality grandmaster-level games, do they not? I hope that Go databases > also > sample professional Dan-level games, otherwise we're just diving into a pool > of > ignorance and drawing up a sample. > > The downside is of course that little exploration is done. In chess it > is still common to find "theoretical novelties" or previously unknown > moves that turn out to be be interesting or even good. So mini-maxing > the choices of Grandmasters is not going to be innovative.
I'd certainly encourage blending existing knowledge with exploration. > I've always thought Joseki for computers is pretty dicey, but I'm not > really qualified to judge this well as I don't play go. But perhaps > joseki is nothing more that a pattern database to suggest moves and you > still just combine it with MCTS? If joseki databases are "nothing more than a pattern database to suggest moves", they will probably be brittle, not knowing how to continue in the face of novel moves. Assuming a dan-level opponent, the difference between knowing the right continuation and not is the difference between a game at the knife edge of 50% win/loss probability, and a certain loss - but any program which doesn't know the continuation will believe it is still in the game. It's like not seeing a looming "scholar's mate" - but in Go, the consequence of a mistake is not so obvious just 4 moves into the future; it can be a capturing race 15, 20, 30 moves deep, but as inexorable and certain as a looming train wreck. The pro, seeing the result, evaluates the game as a certain win. The program, not able to peer that far into the future, would think it's still in the game. You mentioned that chess programs now use very narrow branching trees. Go players likewise use narrow trees; they use a lot of rules of thumb and stored experience with joseki and fuseki to eliminate a large percentage of legal moves as uninteresting. They also use short and focused trees, to study particular objectives: capture this group, defend that group, threaten the other, escape to the center, Life-and-death issues sometimes require a tree 20 or 30 moves in depth; many joseki involve intricate life-and-death problems at their heart; some cutting plays are rejected because they can be captured. Ladders are a common feature of joseki and fusekil; a play may be good only if a certain ladder works, otherwise the same play becomes a liability. A joseki "pattern database" needs to know the details, in depth, accurately; otherwise it can actually be a liability. It needs to know how to correctly handle out-of-book plays as well as the main lines. Given that caveat, yes, such a database should be combined with MCTS. Pros do not follow joseki and fuseki blindly; they choose a particular joseki based on how it works with the other stones on the board. Better to build a wall which faces one's stones than facing one's opponent, for instance. Looking at http://eidogo.com, one often sees several choices, MCTS could help select one depending upon the likely outcomes. Best not to neglect handicap openings. The Nihon Ki-in has an excellent book on handicap play; one chooses joseki which work well with the existing handicap stones, dividing and confining white's forces. ( ISBN 1889554-28-6 ) Look at http://senseis.xmp.net/?Hamete and http://senseis.xmp.net/?TrickPlay for examples of how stronger opponents use overplays to trick weaker players ( and programs ) into doing foolish things. Knowing how to respond to such plays would enable programs to defeat stronger opponents; not knowing will often lead to certain defeat. _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
