Quoting Jacques Basaldúa <[EMAIL PROTECTED]>:
Hi, Magnus
Magnus Persson wrote:
Weak tactics is a problem of the playouts in my opinion. UCT as a
general search method has thus little to do with ladders and other
game specific details. If there are no tactical mistakes in the
playouts the problems disappear. Also tactics has a large impact on
where territory appears in the playout, since much of the noise in
the playouts come from tactical mistakes near the end of the game.
I was meaning the specific problems already mentioned in this list
where you should not start something (e.g. a ladder)
unless you a sure to win. The more you play (if you don't win)
the more you lose. The best move is "hidden" by the increasingly
negative evaluation of continuing the ladder another step and
losing it. In a 9x9 board, ladders may no be very long, but in
19x19 they can. Of course, in the case of ladders that has simple
solutions as forcing the playout to follow the atari-lines, but in
more complex situations there is no known (to me) solution.
Of course, UCT would find the optimal solution with infinite time,
but that is not the question. In fact, it is harder
to find the solution with UCT than with non stochastic methods.
But what I tried to say is that this is easy to fix, if one one adds the
following rules to the playouts.
If a move can escape from atari (ladder broken): play it.
If a move with 2 liberties can be captured in a ladder: play the ladder.
If a stone is in a ladder that is broken: capture it before it escapes
Essentially the playouts will not play out broken ladders.
Now the nice thing about UCT is that when these rules are added the
program will
play much better in almost all tactcal situations, even those that are not
captured explicitely by these rules. The reason is that a lot of noise is
removed from the playouts.
So my point is that the ladder problem is not much of a problem (although
writing fast and bugfree ladder code may be tricky ) for UCT itself. The code
that fixes the problem can also be completely independent of the tree search.
(But I do use it in the search part too, to improve move ordering for
progressive widening).
-Magnus
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
