Hi - yes, that is me, and greenpeep is my program. About 10 years ago I worked on coevolution applied to Go, but greenpeep is an entirely new program based on UCT. I think the greenpeep is mostly similar to what some other people are doing with UCT, and I'm using it to test ideas. greenpeep uses the usual UCT, plus all-moves-as-first based mostly on the MoGo paper from ICML 2007.
greenpeep also uses patterns derived from 20000 UCT self-play games. These are simple local patterns with scores that (roughly) indicate the probability that the move at the center of the pattern was selected by UCT during these games. These patterns are then used both to bias moves at UCT nodes which have few visits, and also to bias the playouts. What I've seen is: - Biasing playouts by patterns is much better than unbiased playouts - Playouts using self-play patterns together with MoGo-style move preferences (favor defensive moves and captures, as well as local moves biased by the self-play patterns, before resorting to a global move biased by patterns) yield much better results than just using the patterns by themselves globally. I tried to do some comparisons to MoGo's hand-coded local patterns as described in the original MoGo report, and the self-play patterns seemed to give overall results that are at least comparable. But I think that there is a lot of room for improvement here. The fact that it is possible to improve on the patterns by forcing additional simple preferences like captures, means the patterns are certainly not as good as they could be. Also, it was necessary to "flatten" the pattern probabilities quite a bit; the quality of the patterns doesn't seem to be good enough to bias moves too strongly. greenpeep uses some other tweaks to improve results, but nothing in the current version that by itself had any large effect. greenpeep on CGOS and KGS has run on an 8-core machine, which certainly helps a lot. I don't think the playouts are especially fast though; the lookups into a large pattern table are one bottleneck. The version on CGOS uses about 500k playouts/move in the opening, then quickly goes down to about 250k playouts/move. I'm curious as to how many playouts other UCT/Monte Carlo programs on CGOS are typically using. -Chris Rosin On 10/11/07, terry mcintyre <[EMAIL PROTECTED]> wrote: > > This may be the same Chris Rosin: > > http://www.cs.wisc.edu/areas/ai/aisem/abstracts/1995.2.summer/rosin.html > http://www-cse.ucsd.edu/users/crosin/ > > Other than the senseis.xmp reference, I have been able to google nothing > about greenpeep. > > Terry McIntyre <[EMAIL PROTECTED]> > > ----- Original Message ---- > From: Olivier Teytaud <[EMAIL PROTECTED]> > > Following this idea of the "public" nature of experiments in cgos, > I am very interested in greenpeep ("playouts guided by > patterns extracted from offline self-play", according to > http://senseis.xmp.net/?ComputerGoServer#toc33), I would be > very > grateful if someone could provide links/infos about it, it is seemingly > quite innovative as it introduces an original way of learning across > games (an efficient coevolution in Monte-Carlo planning would be > great!). > > > > ________________________________ > Yahoo! oneSearch: Finally, mobile search that gives answers, not web links. > _______________________________________________ > 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/
