On Tue, Oct 19, 2010 at 5:33 AM, Petr Baudis <pa...@ucw.cz> wrote: > On Mon, Oct 18, 2010 at 07:18:51PM -0700, James Pettit wrote: > > I have a basic UCT bot with AMAF/RAVE implemented as part of my master's > > research. The bot can muster a small win-rate versus Gnugo level 6 (~20% > as > > black), but usually loses games by not forming 2 eyes and being > completely > > captured. I have uniform random playouts, with simple ko and self-eye > moves > > disallowed. Is this normal for a basic UCT bot? From looking at territory > > visualizations (the viz. post has been most interesting), the bot seems > to > > want to capture white groups that can usually be kept alive, and ignores > > forming eyes to keep itself alive. I'm wondering if I messed up somewhere > in > > the tree portion searching possible opponent moves. Any thoughts? > > You are likely to have some bugs in there. Accidentally reversing > black/white in the minimax, starting the playout with the opposite color > than I should, counting the result wrongly, typos in tree descent > equations, updating wrong tree nodes, bugs in eye detection, suicide > mishandling, etc. etc., been there, done all that. :-) > > Trace through all your code - both tree part and playout part and > carefully verify that it all makes sense; that the tree is growing along > sensible sequences, all moves are made properly, colors are alternating > properly, simulations go through correctly and end in sensible > positions, the tree values are updated properly. Debugging UCT programs > is lenghty and frustrating. > > Note that from my experience, AMAF/RAVE is only very marginally useful > (or not at all) with uniform random playouts, and only starts to be > effective when you add 3x3 patterns and basic capture heuristics. > Perhaps try using raw UCB1 for starters and getting that work properly. > > Then, try out your program on CGOS and check the rating against > comparable implementations: > > http://senseis.xmp.net/?CGOSBasicUCTBots > > If it's roughly in line (you need to wait for few hundreds of games for > the rating to converge reasonably), you can be sure that you have > something that's reasonably bug-free and work from there. >
You can see the bayeselo ratings from the cgos web page which gives you error margins. I would like to point out that the error margins can be a bit misleading because the calculation is based on the assumption that you have decided in advance exactly how many games you will play but it will give you a rough idea of how many games you need to play to have reasonable confidence in the rating. > > -- > Petr "Pasky" Baudis > The true meaning of life is to plant a tree under whose shade > you will never sit. > _______________________________________________ > Computer-go mailing list > Computer-go@dvandva.org > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@dvandva.org http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
