Re: [SPAM] RE: [SPAM] [computer-go] Conflicting RAVE formulae
Hi David, Thanks for these information. Your patterns are not automatically extracted; I don't know to which extent we would benefit from patterns like yours in MoGo, or to which extent you would benefit from automatically extracted patterns as ours, and to which extent it is nearly equivalent or redundant. > I don’t think of Many Faces as having a big database of patterns since > there is so much code, and only 1900 general patterns, but I think in > comparison with other strong programs you might consider it to have a large > pattern database. > Sure! it's different in the way you generate it, but it's also a large pattern database. That's nearly what I meant by writing that you have a huge part of go expertise (in my mind I separated automatically extracted patterns and handcrafted rules for biasing the tree search - sorry for my unclear email). For us, we had a big improvement in mogo when adding patterns _after_ RAVE, and I don't know clearly if we should try to remove rave (and then tuning the formula - this is a tedious work and that's why I've not tested it yet). If you have added rave _after_ patterns, and also had a great improvement, this might indicate that both are necessary for optimal performance. Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [SPAM] RE: [SPAM] [computer-go] Conflicting RAVE formulae
We added (MoGo's original) patterns and RAVE at about the same time. Both helped a great deal, and using both was best of all. Peter Drake http://www.lclark.edu/~drake/ On Sep 15, 2009, at 5:28 AM, Olivier Teytaud wrote: Hi David, Thanks for these information. Your patterns are not automatically extracted; I don't know to which extent we would benefit from patterns like yours in MoGo, or to which extent you would benefit from automatically extracted patterns as ours, and to which extent it is nearly equivalent or redundant. I don’t think of Many Faces as having a big database of patterns since there is so much code, and only 1900 general patterns, but I think in comparison with other strong programs you might consider it to have a large pattern database. Sure! it's different in the way you generate it, but it's also a large pattern database. That's nearly what I meant by writing that you have a huge part of go expertise (in my mind I separated automatically extracted patterns and handcrafted rules for biasing the tree search - sorry for my unclear email). For us, we had a big improvement in mogo when adding patterns _after_ RAVE, and I don't know clearly if we should try to remove rave (and then tuning the formula - this is a tedious work and that's why I've not tested it yet). If you have added rave _after_ patterns, and also had a great improvement, this might indicate that both are necessary for optimal performance. Best regards, Olivier ___ 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] rave and patterns
I implemented RAVE first. Simple playouts with no eye fills and mogo 3x3 patterns and basic uct beat Gnugo 40% (at version 120) Adding RAVE boosted the win rate to 57% (about 30 more versions of tuning). I was trying to duplicate the mogo results before adding my own stuff, to make sure the basic code was debugged. Adding many faces prior to just the root node boosted win rate to 60% (at version 150) Adding one liberty tactics and nakade shapes boosted win rate to 70% (at version 250) Adding mfgo prior throughout the uct tree boosted win rate to 80% (at version 270) UCT rewrite and many bug fixes boosted win rate to 90% (at version 330) All win rates are on 9x9 vs gnugo 3.7.20 level 10 with 5000 playouts. After this I switched to testing 19x19, and stopped tuning for 9x9. It would be easy to turn off rave and run some tests to do the win rate. Would take about a day to get significant results. I think RAVE still helps a lot. I think many faces' patterns help much more on 19x19 than on 9x9. David From: computer-go-boun...@computer-go.org [mailto:computer-go-boun...@computer-go.org] On Behalf Of Olivier Teytaud Sent: Tuesday, September 15, 2009 5:29 AM To: computer-go Subject: Re: [SPAM] RE: [SPAM] [computer-go] Conflicting RAVE formulae Hi David, Thanks for these information. Your patterns are not automatically extracted; I don't know to which extent we would benefit from patterns like yours in MoGo, or to which extent you would benefit from automatically extracted patterns as ours, and to which extent it is nearly equivalent or redundant. I don't think of Many Faces as having a big database of patterns since there is so much code, and only 1900 general patterns, but I think in comparison with other strong programs you might consider it to have a large pattern database. Sure! it's different in the way you generate it, but it's also a large pattern database. That's nearly what I meant by writing that you have a huge part of go expertise (in my mind I separated automatically extracted patterns and handcrafted rules for biasing the tree search - sorry for my unclear email). For us, we had a big improvement in mogo when adding patterns _after_ RAVE, and I don't know clearly if we should try to remove rave (and then tuning the formula - this is a tedious work and that's why I've not tested it yet). If you have added rave _after_ patterns, and also had a great improvement, this might indicate that both are necessary for optimal performance. Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [SPAM] [computer-go] rave and patterns
Thanks for sharing all this information, David. > It would be easy to turn off rave and run some tests to do > the win rate. Would take about a day to get significant > results. I think RAVE still helps a lot. I agree that it's easy to turn off rave, but I think that for a fair comparison you would have to tune the formula for using patterns. For sure, just removing RAVE will make the code much weaker, but removing Rave + re-tuning the formula without rave is something else. > All win rates are on 9x9 vs gnugo 3.7.20 level 10 with 5000 playouts. > After this I switched to testing 19x19, and stopped tuning for 9x9. > Just a point around that: tuning in 9x9 and 19x19 is very different (at least for us), but also tuning for short time settings and tuning for long time settings is very different. We always check that the improvement remains significant for "real" numbers of simulations. Some of the best improvements in MoGo (in particular the "fill board") were of no use for small numbers of simulations (by the way, I hope you'll have improvements with this as well as us, it would be nice for this rule if it was more general than only efficient in MoGo :-) ). > I think many faces’ patterns help much more on 19x19 than on 9x9. > For us, in MoGo, the database of patterns has a very little effect in 9x9. But we made no effort for designing patterns specifically for 9x9. Best regards, Olivier ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [SPAM] Re: [SPAM] RE: [SPAM] [computer-go] Conflicting RAVE formulae
> We added (MoGo's original) patterns and RAVE at about the same time. Both > helped a great deal, and using both was best of all. > You mean mogo's 3x3 patterns I guess; the discussion here is about pattern databases for biasing the research in the tree (patterns with size until 19x19) and how they cumulate with Rave in 19x19. Essentially, MoGo, Mango, Zen, CrazyStone use a automatically built pattern database, whereas ManyFaces use a mix between an automatically built pattern database and a handcrafted pattern database. Mogo, ManyFaces, CrazyStone use Rave values; I don't know for Mango and Zen. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
