Page 1 of 4
Stockfish 10 was released 29.11.2018
Posted: Sun Dec 01, 2019 1:52 pm
by Jouni
After one year we have again +50 ELO gain! +46 single and +54 multicore in framework test. How much has Lc0 gained in last year
.
Re: Stockfish 10 was released 29.11.2018
Posted: Sun Dec 01, 2019 3:26 pm
by Modern Times
Yes we show a good increase at blitz:
http://ccrl.chessdom.com/ccrl/404/cgi/c ... +opponents
So is Stockfish 11 due for release soon ?
Re: Stockfish 10 was released 29.11.2018
Posted: Mon Dec 02, 2019 10:29 pm
by Jouni
NCM testing shows still only +30 ELO. Difficult to understand why. Because they test mp and use same book 8moves_v3.pgn! I think SF11 comes in 2020, when selftest has about +60 ELO gain.
Re: Stockfish 10 was released 29.11.2018
Posted: Mon Dec 09, 2019 6:08 pm
by Jouni
Surprisingly now FishCooking forum has Preparing for Stockfish 11 thread
. Something under Christmas three?
Re: Stockfish 10 was released 29.11.2018
Posted: Mon Dec 09, 2019 7:37 pm
by Modern Times
Jouni wrote: ↑Mon Dec 09, 2019 6:08 pm
Surprisingly now FishCooking forum has Preparing for Stockfish 11 thread
. Something under Christmas three?
Great news if so
Re: Stockfish 10 was released 29.11.2018
Posted: Mon Dec 09, 2019 7:40 pm
by Zenmastur
Modern Times wrote: ↑Mon Dec 09, 2019 7:37 pm
Jouni wrote: ↑Mon Dec 09, 2019 6:08 pm
Surprisingly now FishCooking forum has Preparing for Stockfish 11 thread
. Something under Christmas three?
Great news if so
IF you ask me SF has a problem. Enough of their recent simplifications seem to be regressions to cause a net regression. As long as they allow this to continue their progress will be dead slow. Just like it has been for the past several months. I think they need to reevaluate their simplification testing bounds.
Regards,
Zenmastur
Re: Stockfish 10 was released 29.11.2018
Posted: Mon Dec 09, 2019 8:06 pm
by Modern Times
I did have something weird with Stockfish recently. I chose the latest build at the time for 4040 testing, and after 100 games it was -30 Elo to Stockfish 10. Yes I know that is a tiny sample, but the chances of it recovering that 30 Elo and gaining a further +30-40 Elo over the next couple of hundred games looked remote. So I went back to the Oct 9th build that gave me such good results at blitz, and it is performing well. After I've finished that I'll take the latest build again and see how it goes, to see if the previous result a was just an aberration, or wait for SF11.
Re: Stockfish 10 was released 29.11.2018
Posted: Mon Dec 09, 2019 8:10 pm
by Zenmastur
Modern Times wrote: ↑Mon Dec 09, 2019 8:06 pm
I did have something weird with Stockfish recently. I chose the latest build at the time for 4040 testing, and after 100 games it was -30 Elo to Stockfish 10. Yes I know that is a tiny sample, but the chances of it recovering that 30 Elo and gaining a further +30-40 Elo over the next couple of hundred games looked remote. So I went back to the Oct 9th build that gave me such good results at blitz, and it is performing well. After I've finished that I'll take the latest build again and see how it goes, to see if the previous result a was just an aberration, or wait for SF11.
I'm not surprised. I used to always use the latest but so many are regressions it's better to wait. Unfortunately there are so many simplification regressions that they cancel out any progress made in the past several months.
Regards,
Zenmastur
Re: Stockfish 10 was released 29.11.2018
Posted: Tue Dec 10, 2019 5:19 am
by MikeB
Zenmastur wrote: ↑Mon Dec 09, 2019 8:10 pm
Modern Times wrote: ↑Mon Dec 09, 2019 8:06 pm
I did have something weird with Stockfish recently. I chose the latest build at the time for 4040 testing, and after 100 games it was -30 Elo to Stockfish 10. Yes I know that is a tiny sample, but the chances of it recovering that 30 Elo and gaining a further +30-40 Elo over the next couple of hundred games looked remote. So I went back to the Oct 9th build that gave me such good results at blitz, and it is performing well. After I've finished that I'll take the latest build again and see how it goes, to see if the previous result a was just an aberration, or wait for SF11.
I'm not surprised. I used to always use the latest but so many are regressions it's better to wait. Unfortunately there are so many simplification regressions that they cancel out any progress made in the past several months.
Regards,
Zenmastur
Code pruning is necessary as Stockfish constantly evolves, over time , some code is no longer provides the benefit it once did. It has been tried, tested and true for the last 10 years. The alternative is to do a complete rewrite every 3 or 4 years, which could be far worse. It can be frustrating at times, but the process works. Also consider the coders, they might put in a quite a few hours and a year or 18 months later, see there code get simplified away. It's part of the process. We're looking at 50 Elo gain since SF 10. Not too shabby.
Re: Stockfish 10 was released 29.11.2018
Posted: Tue Dec 10, 2019 8:01 am
by Uri Blass
Zenmastur wrote: ↑Mon Dec 09, 2019 8:10 pm
Modern Times wrote: ↑Mon Dec 09, 2019 8:06 pm
I did have something weird with Stockfish recently. I chose the latest build at the time for 4040 testing, and after 100 games it was -30 Elo to Stockfish 10. Yes I know that is a tiny sample, but the chances of it recovering that 30 Elo and gaining a further +30-40 Elo over the next couple of hundred games looked remote. So I went back to the Oct 9th build that gave me such good results at blitz, and it is performing well. After I've finished that I'll take the latest build again and see how it goes, to see if the previous result a was just an aberration, or wait for SF11.
I'm not surprised. I used to always use the latest but so many are regressions it's better to wait. Unfortunately there are so many simplification regressions that they cancel out any progress made in the past several months.
Regards,
Zenmastur
How do you know that there are many simplification regressions?
usually simplifications pass with more than 50%.
Can you release a significantly stronger stockfish if you delete the simplifications?
I would like to see a version of stockfish that show at least 5 elo improvement based on tests with 40000 games when to achieve this target you simply "delete" simplifications that of course mean adding code.