search

Tencent / PhoenixGo

Go AI program which implements the AlphaGo Zero paper
Other
2.88k stars 577 forks source link

[Feature Request 3] for main move path : add second move path, third move path, etc.. (PV) #85

Closed wonderingabout closed 5 years ago

wonderingabout commented 5 years ago

Hi :+1:

PhoenixGo native engine support is being added to goreviewpartner software here #96

while we are working on this, we noticed that the main move path behaviour can be improved, by adding second move path, third move path because goreviewpartner is an automated review tool that will send main move path, second move path, third move path, etc... into a .rsgf file to view it later after analysis

the new idea is to add more data to verbose levels, something like this : 

I0121 17:55:30.734530 30129 mcts_debugger.cc:43] ========== debug info for 54th move(w) begin ==========
I0121 17:55:30.734563 30129 mcts_debugger.cc:44] main move path: kd(7163,-0.77,0.14,-0.85),be(1775,0.78,0.13,0.80),bd(1171,-0.76,0.41,-0.84),de(476,0.75,0.48,0.76),cd(472,-0.75,0.91,-0.81),if(90,0.76,0.16,0.80),kf(50,-0.76,0.50,-0.84),eq(14,0.76,0.22,0.75),jc(4,-0.72,0.19,-0.73),gf(1,0.70,0.25,0.70),op(0,-nan,0.00,nan)
I0121 17:55:30.734563 30129 mcts_debugger.cc:44] second move path: ch(1374,-0.77,0.14,-0.85),kd(1775,0.78,0.13,0.80),bd(1171,-0.76,0.41,-0.84),de(476,0.75,0.48,0.76),cd(472,-0.75,0.91,-0.81),if(90,0.76,0.16,0.80),kf(50,-0.76,0.50,-0.84),eq(14,0.76,0.22,0.75),jc(4,-0.72,0.19,-0.73),gf(1,0.70,0.25,0.70),op(0,-nan,0.00,nan)
I0121 17:55:30.734563 30129 mcts_debugger.cc:44] third move path: di(394,-0.77,0.14,-0.85),de(1775,0.78,0.13,0.80),bd(1171,-0.76,0.41,-0.84),kd(476,0.75,0.48,0.76),kf(472,-0.75,0.91,-0.81),if(90,0.76,0.16,0.80),kf(50,-0.76,0.50,-0.84),eq(14,0.76,0.22,0.75),jc(4,-0.72,0.19,-0.73),gf(1,0.70,0.25,0.70),op(0,-nan,0.00,nan) 
I0121 17:55:30.734592 30129 mcts_debugger.cc:139] kd: N=7163, W=-5490.75, Q=-0.766544, p=0.14197, v=-0.850747
I0121 17:55:30.734598 30129 mcts_debugger.cc:139] ch: N=1374, W=-1131.4, Q=-0.823437, p=0.338758, v=-0.843087
I0121 17:55:30.734604 30129 mcts_debugger.cc:139] di: N=394, W=-325.479, Q=-0.826088, p=0.10128, v=-0.769427
I0121 17:55:30.734617 30129 mcts_debugger.cc:139] bg: N=293, W=-243.426, Q=-0.830806, p=0.0814862, v=-0.790699
I0121 17:55:30.734627 30129 mcts_debugger.cc:139] bp: N=261, W=-220.618, Q=-0.84528, p=0.0871321, v=-0.829451
I0121 17:55:30.734632 30129 mcts_debugger.cc:139] bh: N=219, W=-191.466, Q=-0.874273, p=0.0987152, v=-0.857132
I0121 17:55:30.734642 30129 mcts_debugger.cc:139] bi: N=73, W=-63.9489, Q=-0.876012, p=0.033503, v=-0.802584
I0121 17:55:30.734650 30129 mcts_debugger.cc:139] eh: N=66, W=-52.9326, Q=-0.802009, p=0.0106665, v=-0.865476
I0121 17:55:30.734656 30129 mcts_debugger.cc:139] cp: N=64, W=-52.9554, Q=-0.827427, p=0.0165746, v=-0.828988
I0121 17:55:30.734663 30129 mcts_debugger.cc:139] cq: N=21, W=-18.6035, Q=-0.885883, p=0.0106966, v=-0.872695
I0121 17:55:31.220618 30129 mcts_debugger.cc:48] model global step: 639200
I0121 17:55:31.220625 30129 mcts_debugger.cc:49] ========== debug info for 54th move(w) end   ==========

we can consider showing it for forth move path, fifth move path, etc... up to tenth move path, but i am not sure that this is very accurate (and probably not very needed) because last moves usually have low simulations

to sum up, i think it would be great to add second move path and third move path for additional analysis information

should we add it to the existing v=1 ? or should we create a new verbose level to do that ?

@wodesuck @pnprog

wodesuck commented 5 years ago

support in this commit.

wonderingabout commented 5 years ago

@pnprog for GoreviewPartner, see this

@wodesuck big thanks again :100:

wonderingabout commented 5 years ago

i think we can close this issue now, thanks again, really :1st_place_medal: