Testing was done at 6400 nodes per move, controlled openings:
Score of Test58 58470 vs Leela ld2: 249 - 183 - 568 [0.533]
Elo difference: 22.96 +/- 14.13
1000 of 1000 games finished.
T58 passing LD2
Moderators: hgm, Rebel, chrisw
-
- Posts: 3019
- Joined: Wed Mar 08, 2006 9:57 pm
- Location: Rio de Janeiro, Brazil
Re: T58 passing LD2
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."
-
- Posts: 10948
- Joined: Wed Jul 26, 2006 10:21 pm
- Full name: Kai Laskos
Re: T58 passing LD2
At fixed time 15s + 0.25s on RTX 2070 OC from unbalanced openings to reduce the draw rate, I got a similar result:Albert Silver wrote: ↑Tue Dec 03, 2019 5:46 am Testing was done at 6400 nodes per move, controlled openings:
Score of Test58 58470 vs Leela ld2: 249 - 183 - 568 [0.533]
Elo difference: 22.96 +/- 14.13
1000 of 1000 games finished.
Score of lc0_58470 vs lc0_LD2: 141 - 103 - 156 [0.547] 400
Elo difference: 33.11 +/- 26.65
Finished match
The real, pentanomial error margins are about 13 Elo points, similar to your errors.
-
- Posts: 17
- Joined: Thu May 07, 2020 9:32 pm
- Full name: Kooper Moris
Re: T58 passing LD2
Another setting for 10x128 net on slow GPU.
[OPTIONS]
DisplayCacheUsage=true
MinibatchSize=256
MaxPrefetch=32
LogitQ=false
CPuct=1.200000
CPuctAtRoot=1.200000
CPuctBase=15996.000000
CPuctBaseAtRoot=15996.000000
CpuctFactor=1.000000
CpuctFactorAtRoot=1.000000
TempEndGame=2.000000
TempVisitOffset=1.000000
FpuValue=0.110000
PolicyTemperature=1.450000
ShortSightedness=0.030000
MinimumKLDGainPerNode=0.000005
SmartPruningFactor=2.330000
Lc0 v25.1 runs slower with low gpu, so use Lc0 v24.1.
Current best 10x128 net for low gpu: net701500.
[OPTIONS]
DisplayCacheUsage=true
MinibatchSize=256
MaxPrefetch=32
LogitQ=false
CPuct=1.200000
CPuctAtRoot=1.200000
CPuctBase=15996.000000
CPuctBaseAtRoot=15996.000000
CpuctFactor=1.000000
CpuctFactorAtRoot=1.000000
TempEndGame=2.000000
TempVisitOffset=1.000000
FpuValue=0.110000
PolicyTemperature=1.450000
ShortSightedness=0.030000
MinimumKLDGainPerNode=0.000005
SmartPruningFactor=2.330000
Lc0 v25.1 runs slower with low gpu, so use Lc0 v24.1.
Current best 10x128 net for low gpu: net701500.
-
- Posts: 4556
- Joined: Tue Jul 03, 2007 4:30 am
Re: T58 passing LD2
What's the current best for slow CPU?
-
- Posts: 1631
- Joined: Tue Aug 21, 2018 7:52 pm
- Full name: Dietrich Kappe
Re: T58 passing LD2
Slow CPU? Maybe the old 11258-112x9-se distilled net? Depends on how slow slow is.
Fat Titz by Stockfish, the engine with the bodaciously big net. Remember: size matters. If you want to learn more about this engine just google for "Fat Titz".
-
- Posts: 4556
- Joined: Tue Jul 03, 2007 4:30 am
Re: T58 passing LD2
Slow as in "Leela can't get more than 23 nps on it". I found acceptable performance with the 256x20-mod-swa net because it gets decent depth in short time, allowing to extrapolate what it'd show with more nodes.
I was worried that it's a 6 months old net, but if the alternative is something old, maybe I already got what I was looking for.
-
- Posts: 17
- Joined: Thu May 07, 2020 9:32 pm
- Full name: Kooper Moris
Re: T58 passing LD2
Update.
Current 10x128 best net is Sergio's t60-2-2990 from https://www.comp.nus.edu.sg/~sergio-v/new/
Current 10x128 best net is Sergio's t60-2-2990 from https://www.comp.nus.edu.sg/~sergio-v/new/
-
- Posts: 1346
- Joined: Sat Apr 19, 2014 1:47 pm
Re: T58 passing LD2
is it a net good at knight handicap ? what would be the best net at knight handicap ?Kooper wrote: ↑Mon Jun 15, 2020 10:04 pm Update.
Current 10x128 best net is Sergio's t60-2-2990 from https://www.comp.nus.edu.sg/~sergio-v/new/