SPCC: Testrun of Lc0 610226 finished

[pohl4711]

NN-testrun of Lc0 0.28.0 610226 finished.

https://www.sp-cc.de

(Perhaps you have to clear your browsercache or reload the website

[dkappe]

Since Dragon now trains against lc0, I look forward to its inclusion in your tests.

[mehmet123]

I'm starting to lose hope that Lc0 will be the most powerful chess engine at TCEC conditions. Lc0 610226 is only +2 elo stronger than Lc0 68695.
There are 532 nets between net 610226 and net 68695. After net 68695, although Lc0 plays 17-18 millions games by itself, there is no significant difference.

[dkappe]

I'm starting to lose hope that Lc0 will be the most powerful chess engine at TCEC conditions. Lc0 610226 is only +2 elo stronger than Lc0 68695.
There are 532 nets between net 610226 and net 68695. After net 68695, although Lc0 plays 17-18 millions games by itself, there is no significant difference.

What the training hasn’t achieved, the hardware change at TCEC may deliver. I fully expect lc0 to get a comfortable 2nd in the cup. Anything less would be a severe disappointment.

[mehmet123]

What the training hasn’t achieved, the hardware change at TCEC may deliver. I fully expect lc0 to get a comfortable 2nd in the cup. Anything less would be a severe disappointment.

It will be a surprise for me now that Lc0 will dominate Dragon from now on. Especially after the difference between Dragon and Stockfish Dev dropped to 15 elo in my 1 core/ 10 minute tests.

[Ozymandias]

I'm starting to lose hope that Lc0 will be the most powerful chess engine at TCEC conditions. Lc0 610226 is only +2 elo stronger than Lc0 68695.
There are 532 nets between net 610226 and net 68695. After net 68695, although Lc0 plays 17-18 millions games by itself, there is no significant difference.

Not 532 nets, but 1532. 68695 should read 608695.

[mehmet123]

I'm starting to lose hope that Lc0 will be the most powerful chess engine at TCEC conditions. Lc0 610226 is only +2 elo stronger than Lc0 68695.
There are 532 nets between net 610226 and net 68695. After net 68695, although Lc0 plays 17-18 millions games by itself, there is no significant difference.

Not 532 nets, but 1532. 68695 should read 608695.

Yes you are right. 1532 nets =~50 million self play games. Now I'm more pessimistic about LcZero

[Pedro]

I'm starting to lose hope that Lc0 will be the most powerful chess engine at TCEC conditions. Lc0 610226 is only +2 elo stronger than Lc0 68695.
There are 532 nets between net 610226 and net 68695. After net 68695, although Lc0 plays 17-18 millions games by itself, there is no significant difference.

What the training hasn’t achieved, the hardware change at TCEC may deliver. I fully expect lc0 to get a comfortable 2nd in the cup. Anything less would be a severe disappointment.

If DeepMind took Leela's latest T60 network and trained it using Google's available resources, how far would Lc0's strength go?

[mehmet123]

If DeepMind took Leela's latest T60 network and trained it using Google's available resources, how far would Lc0's strength go?

Therefore, we should not expect any significant change. AlphaZero trained by 44 million games , Lc0 T60 has been trained by ~200 million games. According to the feature of the chess engine, the increase in elo comes to a halt after a certain number of games.

[pohl4711]

Since Dragon now trains against lc0, I look forward to its inclusion in your tests.

Because Dragon runs on CPU not on GPU, it is tested for my ratinglist on my main-site (https://www.sp-cc.de/). But, when Dragon 2.5 was released, I did a "SuFi for the poor"- testrun vs. the latest Stockfish-Dev, using TCEC-openings and longer thinking-time and all CPU-cores. The result was quite disappointing and only a little better than Lc0-result was. You find it on the "NN vs SF testing"- site, just scroll down to the bottom of the site, all played games can be downloaded, too:
https://www.sp-cc.de/nn-vs-sf-testing.htm

So, I think, the next SuFi will be won by Stockfish, again. If nothing very unexpected happens...