Note that engines like Lc0 and SF12 at 1 second per move already perform at ~3200 elo and so it's not a big suprise they are finding more and more similar moves with increasing time controls. The result is better than I expected.
Great!
My opinion about the convergence on Sim I observed too between SF12 and Lc0 at longer and longer TC is that in mostly quiet positions like Sim ones, Lc0 is incredibly strong and SF12 approaches it with longer and longer TC. One can check that by doing the Sim of Lc0 at 100ms vs Lc0 2500ms (above 70%) and SF12 at 100ms vs 2500ms (below 60%).
Still not sure about the meaning of the convergence. Do they converge to play "perfect" positional chess at longer TC? Maybe they have the same move ordering, that's why they converge? And even the perfect chess often probably has multiple perfect moves in quiet positions, so there is not necessarily a convergence even if they play perfect moves.
I think it's the same relationship with the increase of draws with increased time control.
There are some differences. Perfect chess is most probably a draw, so draw rate naturally increases towards 100%. Perfect move in a position is not necessarily unique. In fact knowing the rating of the random player and the rating of the perfect player, one can roughly estimate that there are on average 1.3 - 1.4 perfect moves for each position, so the Sim test of two different perfect players won't show 100%, but 70-80-90%
Note that engines like Lc0 and SF12 at 1 second per move already perform at ~3200 elo and so it's not a big suprise they are finding more and more similar moves with increasing time controls. The result is better than I expected.
Great!
My opinion about the convergence on Sim I observed too between SF12 and Lc0 at longer and longer TC is that in mostly quiet positions like Sim ones, Lc0 is incredibly strong and SF12 approaches it with longer and longer TC. One can check that by doing the Sim of Lc0 at 100ms vs Lc0 2500ms (above 70%) and SF12 at 100ms vs 2500ms (below 60%).
Still not sure about the meaning of the convergence. Do they converge to play "perfect" positional chess at longer TC? Maybe they have the same move ordering, that's why they converge? And even the perfect chess often probably has multiple perfect moves in quiet positions, so there is not necessarily a convergence even if they play perfect moves.
I think it's the same relationship with the increase of draws with increased time control.
There are some differences. Perfect chess is most probably a draw, so draw rate naturally increases towards 100%. Perfect move in a position is not necessarily unique. In fact knowing the rating of the random player and the rating of the perfect player, one can roughly estimate that there are on average 1.3 - 1.4 perfect moves for each position, so the Sim test of two different perfect players won't show 100%, but 70-80-90%
Ofc perfect move is not unique. There are 3 possible outcomes for each move, and most possible moves in a given position don't change previous outcome. However, on average there is still probably not more than 2 optimal (calling them perfect is kind of too pretentious). Problem is that in couple of decisive moments (and sometimes could be even more than 10 such occasions) in a typical game there is only one optimal move. However, a similarity tool can only measure average values and not values in those critical points of the game.
RubiChess wrote: ↑Thu Sep 10, 2020 9:02 am
Can you also test "regular" Rubi-1.8? Would be interesting to see how good (bad) it is compared to the other non-NNUE engines.
Dann Corbit wrote: ↑Wed Sep 09, 2020 7:15 pm
You would have to turn off the book or the test would be meaningless.
The same could be said for programs that communicate with Bojun Guo's chess database.
It would be fun to see the result of that though - Weiss can do it with NoobBook = true if you want to try @Rebel (expect about 0.7 seconds per move or so, though).
Also slightly off-topic, but it would be cool to see an update to SRL - many engines have had a new version since the last update
I considered the SRL a closed project, but the NNUE developments awakened my curiosity again, will add NNUE engines only.
90% of coding is debugging, the other 10% is writing bugs.
Rebel wrote: ↑Sat Sep 12, 2020 7:16 am
I considered the SRL a closed project, but the NNUE developments awakened my curiosity again, will add NNUE engines only.
How well do the new engines Seer and Komodo Dragon, as well as the updated Halogen 8, Nemorino 6, Ethereal 12.75, Igel 2.8, and Minic 3 perform when given low ply openings?