Hi Tristan, nice to read from you. All the best for your "new" Project!
Ingo. > Gesendet: Mittwoch, 19. Juli 2017 um 21:43 Uhr > Von: cazen...@ai.univ-paris8.fr > An: "Patrick Bardou" <patrick.bar...@laposte.net>, computer-go@computer-go.org > Betreff: Re: [Computer-go] What architecture behind Golois2 KGS 7dan ? > > > Hi Patrick, > > KGS admin told me to reuse my Golois2 account for the new version of > Golois using a value network, that's why it jumped from 1 dan to 7 dan, > the old Golois2 was a policy network, the new one uses the policy network > as a prior for MCTS and a value network to evaluate states. It runs on 3 > GTX 1080Ti GPUs, making approximately 2 000 evaluations in 12 seconds for > each move. It uses resnet as explained in my recent paper in TCIAIG and > ACG. There should be all versions of Golois on KGS from 1 to 7, I recently > only used Golois2 so maybe some other accounts disappeared. I could reach > 4 dan with a resnet based policy network. > > Thank you for your interest, > > Tristan. > > > Hi, > > > > Can anyone (starting with their author maybe) comment on the go > > programs playing on KGS as the various "Golois" versions ? > > > > In particular, I noticed that Golois2 KGS rank recently jumped from a > > solid 1dan to a solid 6dan (actually now in the low 7dan in fast games, > > 15"/move time control; user's info :"Golois version 3.8"). > > > > Golois and Golois2 have been described by their author as greedy > > players, based on a Res-Net policy network. So I'm wondering if the > > currently Golois2 7dan is still based on greedy play (what sounds > > amazing, even for 15s time control) or if it does incorporate some level > > of tactical search. > > > > In his papers "Residual Networks for Computer Go" > > (http://www.lamsade.dauphine.fr/~cazenave/papers/resnet.pdf), "Improved > > architectures for computer go" (https://openreview.net/pdf?id=Bk67W4Yxl) > > and "Combining tactical search and deep learning in the game of Go" > > (http://www.lamsade.dauphine.fr/~cazenave/papers/godeep.pdf) , Tristan > > Cazenave's describes different versions of Golois, but it is not clear > > for me while reading these papers and checking KGS archive users Golois% > > what is behind each version. For example: > > > > "We made the 20 layers residual network with bagging and a 58.5450% > > accuracy play games on the KGS internet Go server. The program name is > > Golois4 and it is quite popular, playing 24 hours a day against various > > opponents. It is ranked 3 dan." > > (http://www.lamsade.dauphine.fr/~cazenave/papers/resnet.pdf). However, I > > don't find a Golois4 in KGS archive, only Golois, Golois2, Golois5, > > Golois6 and Golois7 ... > > > > Thanks and regards, > > Patrick > > > > --- > > L'absence de virus dans ce courrier électronique a été vérifiée par > > le logiciel antivirus Avast. > > https://www.avast.com/antivirus > > > > _______________________________________________ > > Computer-go mailing list > > Computer-go@computer-go.org > > http://computer-go.org/mailman/listinfo/computer-go > > > > > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
