Wolfgang wrote: ↑Sat Jul 02, 2022 3:25 pmSenseless
I find it quite funny, that there are people out there, which truly believe, that mathematical laws are an opinion and that they can see a truth beyond math. Thats really insane.
Fun fact for these people: One engine tests its patches with 10sec+100ms games, followed by 60sec+600ms, singlethreading. Guess which engine this is... Starts with "S" and ends with "ish"...seems, this kind of bulletspeed testing and development works not so bad at all - what a huge surprise...
But I am out here. Discussing math is the perfect example of wasting time.
Don't you know that you cannot graft a new idea on a closed mind!
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
Wolfgang wrote: ↑Sat Jul 02, 2022 3:25 pmSenseless
I find it quite funny, that there are people out there, which truly believe, that mathematical laws are an opinion and that they can see a truth beyond math. Thats really insane.
Fun fact for these people: One engine tests its patches with 10sec+100ms games, followed by 60sec+600ms, singlethreading. Guess which engine this is... Starts with "S" and ends with "ish"...seems, this kind of bulletspeed testing and development works not so bad at all - what a huge surprise...
But I am out here. Discussing math is the perfect example of wasting time.
This engine also get a lot of machine time for testing(clearly more than other engines)
At every second I see thosands of cores to help stockfish to improve so there is no comparison with other methods of testing.
We do not know how much rating stockfish could get with the same computer time and using half of the testing time for longer time control than 60 sec+600 ms.
Note that I do not claim that it is bad to test at 10+100ms and 60sec+600ms but only that maybe it is better to use part of the testing time for clearly longer time control(of course you are going to test relatively less patches at longer time control so these patches do not use most of the testing time).
Wolfgang wrote: ↑Sat Jul 02, 2022 12:08 pm
Do you really believe what you say/write??
LOL (for you: laughing out loud)
What you write is simply wrong, not to say nonsense. With longer TC the difference between the programs may become smaller, but the sequence remains the same. Maybe not every single place in a rating list, but no program is e.g. No.3 with short TC and suddenly No. 10 or so with longer TC. That's simply not true
In theory it is possible that the engine has a bug that cause problems only at LTC so I think that it is better if the developer also test at long time control before the release.
Of course the developer cannot test 10000 games with longer time control but even having 300 games at 5+3 of the new version and the old version
before the release of the new version and reporting the result is better than nothing.
Of course the engine has been tested at longer time control even at 30:00+10 with 32 threads, but for a comparison with the older version I have decided for a 1:00+0.5 tc test that gives similar results to ccrl and cegt blitz.
Fabio Gobbato wrote: ↑Sat Jul 02, 2022 7:23 am
- Small improvement in chess960
All these changes results in about +80 elo compared to Revenge 2.0 measured in selfplay at 1:00+0.5 time control.
This new version compared to Revenge 2.0 is stronger in late middlegame.
In my chess960 testing, I got +98 Elo as measured by bayeselo. If Ordo is your preference, it is +118 Elo, which is a bigger difference than I normally see between those two.
Fabio Gobbato wrote: ↑Sat Jul 02, 2022 7:23 am
- Small improvement in chess960
All these changes results in about +80 elo compared to Revenge 2.0 measured in selfplay at 1:00+0.5 time control.
This new version compared to Revenge 2.0 is stronger in late middlegame.
In my chess960 testing, I got +98 Elo as measured by bayeselo. If Ordo is your preference, it is +118 Elo, which is a bigger difference than I normally see between those two.
+80 elo are measured with standard elo. Small improvement in chess960 is related to a change in the search that affect the play for some elo at most. It seems that the new neural network have improved a lot the play in chess960 more than standard chess probably because I have used more positions for the training.
