Re: [computer-go] pure 3x3 pattern playouts weaker than light playouts?
Orego is currently using the MoGo policy (escape, local patterns, capture, random). Including these as priors helps a little, but just including them in the playouts helps a lot, even with time (rather than # of playouts) fixed. Peter Drake http://www.lclark.edu/~drake/ On Sep 12, 2009, at 5:19 PM, David Fotland wrote: I think most engines that follow the mogo 3x3 pattern design only apply them locally. That’s what I do. I never tried using them globally, so I can’t say what might be wrong. Did you learn weight for every possible pattern using the crazystone algorithm, or are you just using the basic mogo patterns? Are you using global patterns as UCT priors, or to choose moves during playouts? During playout it’s important to favor local responses. David From: computer-go-boun...@computer-go.org [mailto:computer-go-boun...@computer-go.org ] On Behalf Of Isaac Deutsch Sent: Saturday, September 12, 2009 4:27 PM To: computer-go Subject: Re: [computer-go] pure 3x3 pattern playouts weaker than light playouts? 50k playouts. they apply globally. Am 13.09.2009 um 01:22 schrieb Jason House: Same number of playouts? What are your pattern weights? Do they apply around the last move played or for all board areas? ___ 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/
Re: [computer-go] pure 3x3 pattern playouts weaker than light playouts?
Did you learn weight for every possible pattern using the crazystone algorithm, or are you just using the basic mogo patterns? No, all 1000 and something weights are learned using the algorithm. Are you using global patterns as UCT priors, or to choose moves during playouts? During playout it’s important to favor local responses. David No, there are no priors yet, only RAVE. I only use it during playouts. Currently, the playouts only follow the patterns, no other weights. Thanks, ibd___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] pure 3x3 pattern playouts weaker than light playouts?
I think most engines that follow the mogo 3x3 pattern design only apply them locally. That's what I do. I never tried using them globally, so I can't say what might be wrong. Did you learn weight for every possible pattern using the crazystone algorithm, or are you just using the basic mogo patterns? Are you using global patterns as UCT priors, or to choose moves during playouts? During playout it's important to favor local responses. David From: computer-go-boun...@computer-go.org [mailto:computer-go-boun...@computer-go.org] On Behalf Of Isaac Deutsch Sent: Saturday, September 12, 2009 4:27 PM To: computer-go Subject: Re: [computer-go] pure 3x3 pattern playouts weaker than light playouts? 50k playouts. they apply globally. Am 13.09.2009 um 01:22 schrieb Jason House: Same number of playouts? What are your pattern weights? Do they apply around the last move played or for all board areas? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] pure 3x3 pattern playouts weaker than light playouts?
50k playouts. they apply globally. Am 13.09.2009 um 01:22 schrieb Jason House: Same number of playouts? What are your pattern weights? Do they apply around the last move played or for all board areas? ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] pure 3x3 pattern playouts weaker than light playouts?
Same number of playouts? What are your pattern weights? Do they apply around the last move played or for all board areas? Sent from my iPhone On Sep 12, 2009, at 7:18 PM, Isaac Deutsch wrote: I now have playouts based on 3x3 pattern weights. When I tested it on CGOS it seemed to be weaker than my old engine which used light playouts only. I used a comparable setup, and the elo difference is about 100 after more than 150 games. I can also seem some weird RAVE value patterns (not 1-1 points are the worst, but the 2-1/2-2 points), but they are symmetrical so I doubt it's a bug. Is this a known phenomenon and are 3x3 patterns only strong in combination with more knowledge, or should I start debugging? Or is it possibly an effect caused by the rating of the cgos bots not being the same as when I first tested? regards, ibd ___ 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] pure 3x3 pattern playouts weaker than light playouts?
I now have playouts based on 3x3 pattern weights. When I tested it on CGOS it seemed to be weaker than my old engine which used light playouts only. I used a comparable setup, and the elo difference is about 100 after more than 150 games. I can also seem some weird RAVE value patterns (not 1-1 points are the worst, but the 2-1/2-2 points), but they are symmetrical so I doubt it's a bug. Is this a known phenomenon and are 3x3 patterns only strong in combination with more knowledge, or should I start debugging? Or is it possibly an effect caused by the rating of the cgos bots not being the same as when I first tested? regards, ibd ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/