On Tue, 2008-10-28 at 16:37 -0400, Michael Williams wrote: > It makes me wonder if there are other classes of moves that can be segregated > as AMAF-worthy and AMAF-unworthy. > > For instance, maybe capture moves played in the playout are only worth > treating > as first when the enemy did not add any stones to the chain that got captured. > Of course, you can't do anything too complex that would significantly slow > down the playouts.
It's pretty tricky doing anything very clever without slowing things down a lot. But I am highly interested because I want to build a stronger Ogo type of program for handheld devices. I could probably improve Ogo already based on this idea of decaying weights. Ogo is the same as anchorMan but I don't like the tricks I use to get to 1500 ELO - they are pretty ugly hacks in my opinion. I could manage 5000 playouts without it taking too long on 9x9 but that is really an upper limit. One thing I once tried was to take statistics on each 3x3 pattern on the board instead of each move. So a move to c3 when not surrounded by anything is different that C3 when there is something around it. I didn't get any improvement from this - it was just slow and weak. The patterns really kill how many samples you can get from each playout and of course it's a major slowdown to an optimized program. So this is not just academic for me. I would like to find some more simple things (emphasis on simple) to improve basic bots. Of course tree search is not something I want to get into for a handheld device although it's possible. You clearly don't want a memory hog on a handheld device and you basically need fast response time to be practical. Of course there is a limit to how far you can go without tree search and fancy heuristics, but I think it might be possible to push this pretty far with some cleverness. I wonder if 1700 is out of the question? So I think it's possible to push the AMAF-worthy idea a little farther. If you look at actual games, you will see that one of the most common problems is self-atari to create an atari of it's own. The Ogo solution was artificial, I just detected that and gave it a little penalty. But I would like to find something more in the spirit of what we are doing now - not so domain specific. I'm also wondering what we could do by taking statistics on the opponents moves. I can't think of anything immediately obvious. - Don > > Don Dailey wrote: > > On Tue, 2008-10-28 at 15:07 -0400, Michael Williams wrote: > >> It's worth noting that the standard ref bot also has the first move test. > > > > Yes, so presumably even the weakest version is doing it the "best" way, > > i.e. doing the check. > > > > - Don > > > > > > > >> > >> Don Dailey wrote: > >>> It appears from testing 3 of us are doing, that the Mark Williams > >>> enhancement is good, and that testing to see if the move was played > >>> before is VERY GOOD when combined with the enhancement. > >>> > >>> The enhancement without the checking is perhaps 40 ELO stronger but with > >>> the testing it appears to be about 110 ELO stronger. Of course these > >>> numbers are rough estimates. I can give more precise numbers in a few > >>> hours. I'm round robin testing 3 versions against each other: > >>> > >>> 1. standard ref bot > >>> > >>> 2. mw enhanced - no first move test. > >>> > >>> 3. mw enhanced - with first move test. > >>> > >>> > >>> - 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/ > >> > >> ------------------------------------------------------------------------ > >> > >> _______________________________________________ > >> 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/
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/
