a killer was born last night.

[mwyoung]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1


DESKTOP-CORSAIR, Blitz 0min+10sec 0


1 Lc0 v0.23.3+git.29c15a4 +89 +5/=15/-0 62.50% 12.5/20
2 Stockfish 220220 64 POPCNT -89 +0/=15/-5 37.50% 7.5/20

[corres]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

[mwyoung]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

Default except for cs 2000000

Default for sf 32 threads 4gb ht
CPU 2950x vs 2080ti

[corres]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

Default except for cs 2000000
Default for sf 32 threads 4gb ht
CPU 2950x vs 2080ti

Thanks.
I run a short test with 200218Stockfish and I got a different result:
Stockfish : lc0-62464 = 1 : 0 19draw (20 games)
Parameters were as above.
I am afraid the series of 62 nets are still not a terrorist.

[mwyoung]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

Default except for cs 2000000
Default for sf 32 threads 4gb ht
CPU 2950x vs 2080ti

Thanks.
I run a short test with 200218Stockfish and I got a different result:
Stockfish : lc0-62464 = 1 : 0 19draw (20 games)
Parameters were as above.
I am afraid the series of 62 nets are still not a terrorist.

I am getting different results. Now playing the NN with Lc0 24.0. Score +2 =16 -0. That is a combined score of +7 =31 -0.
The new Lc0 24.0 has major changes from Lc0 23.3. But so far playing ok.

[corres]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

Default except for cs 2000000
Default for sf 32 threads 4gb ht
CPU 2950x vs 2080ti

Thanks.
I run a short test with 200218Stockfish and I got a different result:
Stockfish : lc0-62464 = 1 : 0 19draw (20 games)
Parameters were as above.
I am afraid the series of 62 nets are still not a terrorist.

I am getting different results. Now playing the NN with Lc0 24.0. Score +2 =16 -0. That is a combined score of +7 =31 -0.
The new Lc0 24.0 has major changes from Lc0 23.3. But so far playing ok.

As I see the main difference betrween lc0v0.23 and lc0v0.24 is in self game generation. Using v0.24 there is possibility to start games from opening book. This would enhance the knowledge of Leela about mid-games.
I think it is a good trick. I wrote about cudnn-fp16 backend,

[mwyoung]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

Default except for cs 2000000
Default for sf 32 threads 4gb ht
CPU 2950x vs 2080ti

Thanks.
I run a short test with 200218Stockfish and I got a different result:
Stockfish : lc0-62464 = 1 : 0 19draw (20 games)
Parameters were as above.
I am afraid the series of 62 nets are still not a terrorist.

I am getting different results. Now playing the NN with Lc0 24.0. Score +2 =16 -0. That is a combined score of +7 =31 -0.
The new Lc0 24.0 has major changes from Lc0 23.3. But so far playing ok.

As I see the main difference betrween lc0v0.23 and lc0v0.24 is in self game generation. Using v0.24 there is possibility to start games from opening book. This would enhance the knowledge of Leela about mid-games.
I think it is a good trick.

That was not what I was referring to in regards to Lc0 0.24.0. Look at the default settings for game play.

[corres]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

Default except for cs 2000000
Default for sf 32 threads 4gb ht
CPU 2950x vs 2080ti

Thanks.
I run a short test with 200218Stockfish and I got a different result:
Stockfish : lc0-62464 = 1 : 0 19draw (20 games)
Parameters were as above.
I am afraid the series of 62 nets are still not a terrorist.

I am getting different results. Now playing the NN with Lc0 24.0. Score +2 =16 -0. That is a combined score of +7 =31 -0.
The new Lc0 24.0 has major changes from Lc0 23.3. But so far playing ok.

I repeated the test with kuidee-parameters and policy temperature 2.10
The new result is:
62464 : Stockfisch = 4 : 3 (13 draw) 20 games.
So where is the terrorist?
Nowhere.

[mwyoung]

Getting really good early results with this NN 62464. https://training.lczero.org/
62464 is the high peak on the graph. Policy temp 2.1
...

What are the other parameters?

Default except for cs 2000000
Default for sf 32 threads 4gb ht
CPU 2950x vs 2080ti

Thanks.
I run a short test with 200218Stockfish and I got a different result:
Stockfish : lc0-62464 = 1 : 0 19draw (20 games)
Parameters were as above.
I am afraid the series of 62 nets are still not a terrorist.

I am getting different results. Now playing the NN with Lc0 24.0. Score +2 =16 -0. That is a combined score of +7 =31 -0.
The new Lc0 24.0 has major changes from Lc0 23.3. But so far playing ok.

I repeated the test with kuidee-parameters and policy temperature 2.10
The new result is:
62464 : Stockfisch = 4 : 3 (13 draw) 20 games.
So where is the terrorist?
Nowhere.

"So where is the terrorist?"

It is playing online. Now with a combined score of +9 =52 -1. Killer!

[Ovyron]

I can't find a release of 24.0, I guess it's still deemed as "unstable"?