[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [gnugo-devel] joseki database
Re: [gnugo-devel] joseki database
Tue, 3 Dec 2002 17:10:27 +0200
> On Mon, Dec 02, 2002 at 10:57:06PM +0200, Paul Pogonyshev wrote:
> > i'll take it. besides, tree-based matching will be a step towards a
> > joseki reader (a one that chooses a joseki move by evaluating all
> > possible joseki outcomes in terms of moyo/influence).
> If you haven't fixed your sigths to some solution already, I have a few
> points I'd like to mention for possible discussion.
> * Although choosing from "all possible" joseki outcomes sounds like the
> proper thing, it may be too expensive (thousand variations of the taisha).
> But often a move has a meaning, a direction it is pushing the game. It
> might be useful to express this as a likely outcome position, and choose
> according to that. Maybe (later?) a selection of a few possible outcomes.
> The whole decision tree (who decides what and where) can be overly complex.
well, i'm going to start with a mere pattern matcher, so there will
be plenty of time to ponder on joseki reader. however, you are
probably right that considering all possible outcomes may be too
> * Not sure what sort of "tree-based" matching you are going to do, but if
> it based on following the game, watch out for situations where a player
> deviates from the joseki, and later the game reaches a well-known position.
> If white starts at 6-3, and black approaches 4-4, and white tenukis, we
> have a classical 4-4 joseki position... This may not happen all too often
> in the games of the Great Masters, but when humans play against a "mere
> computer", this is the kind of tricks they hope for the program to fall
> into. And we hope not to...
the thing i'm going to implement will handle such situation simply because
it will never know how a given position have been achieved, either by
playing a standard joseki or by some untraditional sequence. it will
use a tree internally, but match patterns on positions, not on move
> * Choosing the best-looking joseki is no use, if we can not handle the
> resulting position. No point in aiming for a slightly better fighting
> position if we know our fighting ability is weak - or staying away from a
> slightly bad fighting joseki if fighting is our best side... Tuning this by
> hand will be tedious, and needs to be redone when even the balance of the
> engine changes. Doing it automatically (learning) has a great number of
> problems as well.. In short - choosing the right joseki is not easy :-)
agreed. but a faster joseki pattern matcher with a smaller pattern
database will help nevertheless.
> * A joseki is not only the well-known sequence of plays, but also the
> reason behind every one of them. If the opponent blunders in the joseki, we
> ought to know what to do about that. As Kageyama said in "Lessons in the
> Fundamentals of Go": Joseki are not to be momrized, but to be discovered.
this is the toughest challenge, i think. i believe that there are too
ways: 1) make tactical/semeai/owl reading (and probably the rest of the
engine) _very_ strong so it can handle such situations and punish the
opponent (very difficult); 2) create a huge bundle of corresponding
patterns (not so difficult but very boring and uninteresting).
thank you for the comments.