LCZero: Progress and Scaling. Relation to CCRL Elo

[Dann Corbit]

I meant software which uses neural networks.is it giraffe ?

Giraffe, but it had a very different kind of learning and search.

I think if Giraffe were converted to Cuda, we might all get our eyes opened.
Oh, wait. They already did that.

[peter]

I meant software which uses neural networks.is it giraffe ?

Giraffe, but it had a very different kind of learning and search.

I think if Giraffe were converted to Cuda, we might all get our eyes opened.
Oh, wait. They already did that.

[Nay Lin Tun]

If Leela were running in A0 hardware, Leela would be beating stockfish 8 on 32 cores. (64 thread)

[George Tsavdaris]

I noticed yesterday that surprisingly Leela "solves" a good position.

White to play 1.Nd5!! and win. 1.Re4 just draws....
[d]8/2p2p1p/3k1p2/pn6/PNp4p/r3Rp1P/2P2P2/BK6 w - - 0 1

It finds Nd5!! A superb move! In a superb position(mate in 178).
But don't get too excited, as of course she finds it for wrong reasons and doesn't see it is winning. But she finds it and for someone that tries to solve a position the fact that Nd5 has so many visits and it's the preferable move of Leela, should ring a bell when analyzing a position that this move has its merits.
To be fair to other engines, Leela underestimates Re4 and thinks black is better, which is not the case as white puts much pressure to black with Re4, but can't win.



Good old [d] doesn't wok as it seems.

[Milos]

If Leela were running in A0 hardware, Leela would be beating stockfish 8 on 32 cores. (64 thread)

Ofc it would, if those 32 cores were ARM cores running at 800MHz

[JJJ]

been watching a 1'1 match between leela & SF8 on one core, & it's transparently clear that, while tactics are an obvious issue, leela's woeful misevaluation of endgames is bleeding a lot of wins (& draws). i don't really understand how, after 15 million games, it could be so bad, but she has repeatedly dissolved into drawn endgames she thinks are +2 or +3. KRB vs KR is the last example off the top of my ahead, & i think the game eval was something like +2.50.

is there any concrete evidence this is improving?

Yeah, I see the exact same thing with Leela. But I guess with more training she will overcome that as well anyway.

[Albert Silver]

been watching a 1'1 match between leela & SF8 on one core, & it's transparently clear that, while tactics are an obvious issue, leela's woeful misevaluation of endgames is bleeding a lot of wins (& draws). i don't really understand how, after 15 million games, it could be so bad, but she has repeatedly dissolved into drawn endgames she thinks are +2 or +3. KRB vs KR is the last example off the top of my ahead, & i think the game eval was something like +2.50.

is there any concrete evidence this is improving?

Well, for this to be a provable issue (compared to other top engines) you would need to show that other engines did better in the exact same time. Share a position where she made a mistake, entering the wrong endgame, and that SF or Komodo find the solution in that same time control.

You are aware that KRB vs KR is quite winnable, and depends very much on the defender having the right position and not making any mistakes even then? There are excellent practical chances to win it.

[Milos]

been watching a 1'1 match between leela & SF8 on one core, & it's transparently clear that, while tactics are an obvious issue, leela's woeful misevaluation of endgames is bleeding a lot of wins (& draws). i don't really understand how, after 15 million games, it could be so bad, but she has repeatedly dissolved into drawn endgames she thinks are +2 or +3. KRB vs KR is the last example off the top of my ahead, & i think the game eval was something like +2.50.

is there any concrete evidence this is improving?

Well, for this to be a provable issue (compared to other top engines) you would need to show that other engines did better in the exact same time. Share a position where she made a mistake, entering the wrong endgame, and that SF or Komodo find the solution in that same time control.

You are aware that KRB vs KR is quite winnable, and depends very much on the defender having the right position and not making any mistakes even then? There are excellent practical chances to win it.

Lol, here is a test for you. Play against your strongest LC0 10s per move or even more if you wish. You with naked king and give LC0 bishop and knight. You can try couple of times. And please tell us the result.
Than take any below 2500 Elo engine without TBs and repeat. And tell us the result.

[Albert Silver]

been watching a 1'1 match between leela & SF8 on one core, & it's transparently clear that, while tactics are an obvious issue, leela's woeful misevaluation of endgames is bleeding a lot of wins (& draws). i don't really understand how, after 15 million games, it could be so bad, but she has repeatedly dissolved into drawn endgames she thinks are +2 or +3. KRB vs KR is the last example off the top of my ahead, & i think the game eval was something like +2.50.

is there any concrete evidence this is improving?

Well, for this to be a provable issue (compared to other top engines) you would need to show that other engines did better in the exact same time. Share a position where she made a mistake, entering the wrong endgame, and that SF or Komodo find the solution in that same time control.

You are aware that KRB vs KR is quite winnable, and depends very much on the defender having the right position and not making any mistakes even then? There are excellent practical chances to win it.

Lol, here is a test for you. Play against your strongest LC0 10s per move or even more if you wish. You with naked king and give LC0 bishop and knight. You can try couple of times. And please tell us the result.
Than take any below 2500 Elo engine without TBs and repeat. And tell us the result.

No need. It used to have trouble until it learned how to win them. All this has been tracked in detail:

https://docs.google.com/spreadsheets/d/ ... edit#gid=0

[duncan]

Well, for this to be a provable issue (compared to other top engines) you would need to show that other engines did better in the exact same time. Share a position where she made a mistake, entering the wrong endgame, and that SF or Komodo find the solution in that same time control.

You are aware that KRB vs KR is quite winnable, and depends very much on the defender having the right position and not making any mistakes even then? There are excellent practical chances to win it.

Lol, here is a test for you. Play against your strongest LC0 10s per move or even more if you wish. You with naked king and give LC0 bishop and knight. You can try couple of times. And please tell us the result.
Than take any below 2500 Elo engine without TBs and repeat. And tell us the result.

No need. It used to have trouble until it learned how to win them. All this has been tracked in detail:

https://docs.google.com/spreadsheets/d/ ... edit#gid=0

do you know who can do it in least amount of moves stockfish or lc0?