-
Notifications
You must be signed in to change notification settings - Fork 528
Batchsize Node Collisions Testing
jjoshua2 edited this page Oct 19, 2018
·
2 revisions
1024 bs 256 nc (cccc 2 params) vs default so far.
Score of lc0 v18 1024 vs lc0 v18: 33 - 45 - 89 [0.464]
Elo difference: -25.01 +/- 36.08
Trying medium vs large batchsize. 1024 batchsize and 256 node cccc vs TCEC 512/32.
Score of lc0 v18 1024 vs lc0 v18 512: 20 - 29 - 65 [0.461]
Elo difference: -27.49 +/- 41.92
Trying 512/128 vs default.
Score of lc0 v18 512 vs lc0 v18: 87 - 86 - 155 [0.502]
Elo difference: 1.06 +/- 27.32
Trying lower node collisions than default.
Score of lc0 v18 512 16 vs lc0 v18 512: 41 - 48 - 109 [0.482]
Elo difference: -12.29 +/- 32.49
Trying larger than default.
Score of lc0 v18 512 vs lc0 v18 512 nc128: 46 - 37 - 95 [0.525]
Elo difference: 17.58 +/- 34.91
Navs will add more data to this.
TLDR. 1024 batchsize never wins. Non-default node collisions never wins.