Dear DetlefÂ, Please don't forget the rollout, which is still important. Zen's rollout is much better than others' so far. This could be the main reason of your question.
The value network introduced by the Nature paper is not perfect, rather represents very rough (coarse) approximation of the value function of 19x19 Go. The number of possible positions is 10^170 and is greater than 10^30 in practice, I guess. The order of the number of the freedom (and the training samples) of the value network is 10^7 which is much much smaller than the number of freedom being necessary to represent the value function. So, the value network only approximate rough, smooth, low-order part of the function. Thus, the detal, high-order part of the function must be represented by the rollout. Best, Hideki Detlef Schmicker: <d8a14055-2f4e-7e99-d9ba-2c79a34af...@physik.de>: >Hi, > >I'd like to start a discussion on what zen might do being so strong on >CGOS with only one core and no graphic card :) > >The version actual playing (and therefore best comparable to the >programs actual playing) is > >Rank Name Elo + - Games >12 Zen-13.5-1c0g 3639 48 48 569 >16 Aya792p2v2cn50_12t 3558 46 46 476 >24 Rn.3.3-4c 3459 34 34 918 >26 No335-4.5-gpu-4c 3438 115 115 112 >29 CGI1407_1_475_7c 3399 73 73 142 >53 CrazyStone-0002 3239 114 114 141 >54 NG-05 3235 67 67 228 > > > >No335 is probably Hirabot >NG is oakfoam >from crazystone I dont know, if this version uses gpu >Rn is probably ray > >All but Zen use gpu and so have probably about 300 policy network calls >per second. > >Zen does this on cpu and I got Hideki to let me know, that this can do >only about 20 calls per second :) > >It might be that zen uses cnn only in the upper nodes and relies on the >faster node generation zen used before cnn for deeper nodes? > >I can not believe, that the policy network is so much stronger than the >ones of all the others?! > > >Any ideas > > >Detlef >_______________________________________________ >Computer-go mailing list >Computer-go@computer-go.org >http://computer-go.org/mailman/listinfo/computer-go -- Hideki Kato <mailto:hideki_ka...@ybb.ne.jp> _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
