On Tue, 2007-01-09 at 16:31 +0100, Benjamin Teuber wrote:
> I just lost my first game against MoGo on KGS, 9x9, 0.5 komi, I was white.
> Impressing!
> But as a human, you don't like the useless endgame-moves MC-programs
> play against you when they know they win anyways.
> In order to make these programs more attractive for humans, I would like
> them to play the move winning by the biggest amount of points once
> several moves have the same high winning probability at the endgame.
> What do you think about this?
I've tried - it's hard to make it work. Almost anything you do in
this regard weakens the program.
The easiest and safest way to make the endgame seem more natural is
is to pre-process the moves once the program knows it is winning or
losing. Since MC considers all moves as equal once it is winning,
you have to "impose your will" on the algorithm.
One way to do this is just to determine which moves seem reasonable
to a person, then "hint" to the program that all others are
slightly losing (by give them a few pseudo losses at the root node
of the tree.) In this way the algorithm could even recover if
the routines which determine this get it wrong.
I would apply such an algorithm only after MC said the position
was at least 95 percent a win or loss. This may sound strict,
but MC usually knows this pretty early - way before the game is
played out to the bitter end.
It's easy to track or identify points that are ambiguious, just
play 1000 random games (in a fraction of a second) and keep
statistics on those points not clearly being occupied (or
surrounded) by one side or another. Then you can emphasize
moves to those points as I outlined above.
I have tried things like increasing or decreasing komi
appropriately so the programs continues to fight hard for
points once it has a won (or lost) game, but this is
quite risky. An MC program can give away the game if you
put in a position where it thinks it might be losing when
it isn't.
You can also switch to bean-counting once a game is won
or lost so that the program will try to maximize it's
territory gain. This is actually quite risky too - it
is far weaker in normal play and even when the game is
virtually "over" one must be very careful - since you
are indeed weakening the program to achieve the
artificial appearance of human-like play.
All of this reminds me of the Japanese/Chinese debate.
You can do things to make it more human-like or
convenient for us humans, but whatever you do it's
not scientifically relevant or interesting from the
perspective of better play. It's a cosmetic issue.
These type of things can be interesting and even
scientific in their own right (you can make a study
of anything) - I'm not saying that it isn't. It's
definitely interesting (and a challenge) to consider how
this might be done without weakening the program too
much.
It's like this with Japanese - it's an interesting
problem how to do it really well and correct - it's
just not relevant to program strength although the
things learned could contribute to program strength.
- Don
> _______________________________________________
> computer-go mailing list
> computer-go@computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
