CEGT - rating lists March 28th 2021

[Wolfgang]

Hi Larry,

interesting that you figure out the OS as possible explanation. Yesterday I started a match with 5'+3" and Ponder under Linux Mint and Arena 3.10beta. After 30 games, of course a very small sample the Performance vs. Stockfish 13 was much better than previousely.
I will make some matches @40/3 without during the next days for comparison. AVX2 compile on i5-4440 Haswell, the only machine where Linux is installed here (my working PC).

BTW I'm pretty sure that Stefan uses Windows.

[Werner]

As the engine running inside command prompt ist about 40% faster here in nps as in a GUI like Arena or Shredder,
it is possible, a uci parameter is not correct set and misinterpreted by the GUI ??

[lkaufman]

As the engine running inside command prompt ist about 40% faster here in nps as in a GUI like Arena or Shredder,
it is possible, a uci parameter is not correct set and misinterpreted by the GUI ??

We discovered that something was set wrong in MCTS mode with the Windows compile, and fixing it is showing +68 elo after 300 games. So our next Dragon release will show a large elo gain for MCTS in Windows even if the gain in Linux is small.

[pohl4711]

We discovered that something was set wrong in MCTS mode with the Windows compile, and fixing it is showing +68 elo after 300 games. So our next Dragon release will show a large elo gain for MCTS in Windows even if the gain in Linux is small.

This matches perfeclty to my testings...

SPCC ratinglist:
AB-versions distance (Komodo 14 / Dragon): +204 Elo
MCTS-versions distance (Komodo 14 MCTS / Dragon MCTS) now: +141 Elo -> adding +68 Elo for the Windows Binary fix would give: +209 Elo

So, my results match perfectly (same Elo-increase of AB and MCTS versions (both around +200 Elo better)). And the CEGT/FGRL-results are still very strange (Dragon MCTS scoring is still much too bad, even though if +68 Elo for the windows-bugfix are added).