TalkChess.com

found on http://www.chessdom.com/breaking-leela- ... season-12/.

Nice promo for the project!

maschmdt wrote:found on http://www.chessdom.com/breaking-leela- ... season-12/.

Nice promo for the project!

A lot of nonsense there e.g.

Tests have shown that LC0 performs as much as 300 Elo better when paired with 6-man Syzygy tablebases

Guenther wrote:

maschmdt wrote:found on http://www.chessdom.com/breaking-leela- ... season-12/.

Nice promo for the project!

A lot of nonsense there e.g.

Tests have shown that LC0 performs as much as 300 Elo better when paired with 6-man Syzygy tablebases

Did you do tests that show that it is not the case?

I do not know about tests that showed 300 elo improvement for LC0 by using tablebases in the search but it does not prove that it is not the case.

If TCEC used 4 TPUs, LeelaZero could already maybe be on AlphaZero's level. However, if TCEC keeps on using the 44 cores cpu machine, I am pretty sure it would never ever beat Stockfish even in TCEC 100. The 10x128 network size as is may already be the optimal size with the 44 cores hardware in mind; A0 calculated 4 TPU's would be necessary for their 40x256 net i.e. 8x larger and maybe 8x slower.

I am considering entering the weakest version of Scorpio-MCTS (scorpio-mcts-min) that I estimate 2300-2400 elo -- it might be 2700 elos there too like leela. The evaluation of scorpio-mcts-min (and ofcourse this is hand coded evaluation) is about 100x faster than leela's. So if LeelaZero doesn't beat, it means the NN effort is useless on that hardware.

Give a thumbs up here if you like this idea -- not many hours left till it starts.

Daniel

Daniel Shawul wrote:If TCEC used 4 TPUs, LeelaZero could already maybe be on AlphaZero's level. However, if TCEC keeps on using the 44 cores cpu machine, I am pretty sure it would never ever beat Stockfish even in TCEC 100. The 10x128 network size as is may already be the optimal size with the 44 cores hardware in mind; A0 calculated 4 TPU's would be necessary for their 40x256 net i.e. 8x larger and maybe 8x slower.

I am considering entering the weakest version of Scorpio-MCTS (scorpio-mcts-min) that I estimate 2300-2400 elo -- it might be 2700 elos there too like leela. The evaluation of scorpio-mcts-min (and ofcourse this is hand coded evaluation) is about 100x faster than leela's. So if LeelaZero doesn't beat, it means the NN effort is useless on that hardware.

Give a thumbs up here if you like this idea -- not many hours left till it starts.

Daniel

please go ahead.

It's playing now against the Baron. So hurry to TCEC. 24:00 too late for Europeans.

Pretty miserable performance so far, only one draw against The Baron. Archive is down, how have the games been?

IanO wrote:Pretty miserable performance so far, only one draw against The Baron. Archive is down, how have the games been?

If you consider "miserable" that a 1 month old project is already at level to play TCEC division 4 and somehow getting some eventual draws with 3000 elo engines while playing with a bugged network, well, we definitely have a different concept of "miserable".

I would say expected performance.

LCzero is already +185 elo in selfplay it seems compared to LCzero id 125 :
http://lczero.org/

In other hand, I m not sure tablebase give LCzero 300 elo

Disappointing results so far. Only 1/2 a point out of 5.