On Sat, 2008-10-11 at 22:33 +0100, Raymond Wold wrote: > Ingo Althöfer wrote: > > During the last few days I have been meditating a lot about > > the questiion whether taking into account the margin of > > win into MCTS (UCT) may help or hurt.
You are not alone! I think most of us have looked into that. > > I do not have a go program by my own, so for the moment > > I have to believe what programmers are saying, namely > > that "MCTS with margin of win as a criterion" gives worse > > performance in play against other computer programs. It seems that way so far but you never know. > > I think the computer go field has a lot of magical thinking like this. > Where a couple of different teams try out an idea, no one can get it to > work, and so everyone believes the idea is no good. I couldn't say it better myself. In fact I have dropped my own ideas, only to make them work later. One can never be completely sure there is not a bug, or some implementation detail that seemed unimportant caused a problem. > > The truth is we don't know. Most of computer go is very empirical and > ad-hoc. We don't know why playing out a game randomly many times gives a > good approximation of winning chance (I'm sure I could design games > where this is not true). Using the margin and not just win/loss would > intuitively seem to mean a higher margin of error in this estimate, so > that a lot more play-outs are needed to take advantage of any added > information. But that would really just be another guess. Are you speculating that if enough play-outs are done, the idea might prove to be superior? I never actually considered that. So perhaps with 5000 playouts using the win/loss score wins, but at 50,000 using the margin might be better? This is easy to test with simple MC go programs. If I get a chance I will test it for you (since you do not have a program of your own) and report the results. I'm not holding my breath however ... > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/
signature.asc
Description: This is a digitally signed message part
_______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
