Hi, I did not mention here the prior initialization that is done in each node. When you create a node, you can look at all possible move and if a pattern matches (the exact same as in the playout) you initialize rw and rc to 14. If the move saves a capture (same as in the playout), same initialization, rw and rc to 14. If it is a self atari, you initialize rw to 0 and rc to 14. Else you initialize rw to 7 and rc to 14. Of course you can do put much more clever prior if you are a player and know the subtleties of the game.
You can put an exploration term, but the cases where it is needed are rare. I did a lot of experiments on that, and even at long thinking time, no exploration term was always better (statistically). Sylvain 2009/1/21 Mark Boon <tesujisoftw...@gmail.com> > > On Jan 21, 2009, at 10:23 AM, Magnus Persson wrote: > > Quoting Thomas Lavergne <thomas.laver...@reveurs.org>: >> >> - the best play is a good only if played immediatly and very bad if >>> played later in the game : >>> - the first playout for this play resulted in a lost. >>> score and RAVE score will be very low and this play will never be >>> considered again until a very long time. >>> >> >> >> You raise an interesting concern. >> >> The simple solution to your question is to add an exploration term using >> UCT for example. Then it becomes an empirical question what parameter for >> exploration gives the strongest play. My experience is that the best >> parameter is so small it can be set to zero. >> > > Well, empirically, when I set the exploration component to zero it starts > to play a lot worse. Like I wrote: the winning percentage drops to 24% vs. > the same program with the exploration component, which is a huge difference. > > So if you have a different experience, you must have something else that > overcomes this hurdle that's not part of a simple MCTS-RAVE implementation. > I'd be very interested to learn what that is. Sylvain didn't take the bait > ;-) > > Mark > > > _______________________________________________ > 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/