Page 1 of 13

this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 3:28 pm
by MikeB

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 4:11 pm
by zullil
MikeB wrote:
Tue Aug 04, 2020 3:28 pm
https://tests.stockfishchess.org/tests/ ... 64f05e4c58

I think this qualifies for SF 12!!
That's relative to Stockfish-dev. About +120 Elo above Stockfish 11, I think.

So much for human programming of evaluation functions. :wink:

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 4:17 pm
by Leto
Is that at LTC?

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 4:27 pm
by lkaufman
Leto wrote:
Tue Aug 04, 2020 4:17 pm
Is that at LTC?
I'm getting a result that sfnnue (yesterday's 2108 net) on one thread is beating Stockfish 11 on four threads, at 2' + 1" by 9 elo after 200 games. This nnue is the real deal.

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 4:45 pm
by zullil
Leto wrote:
Tue Aug 04, 2020 4:17 pm
Is that at LTC?
tc (20+0.2 )s

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 4:56 pm
by Raphexon
zullil wrote:
Tue Aug 04, 2020 4:45 pm
Leto wrote:
Tue Aug 04, 2020 4:17 pm
Is that at LTC?
tc (20+0.2 )s
20+0.2 at 8 threads.

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 5:58 pm
by MikeB
Leto wrote:
Tue Aug 04, 2020 4:17 pm
Is that at LTC?
It scales very well.

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Tue Aug 04, 2020 6:01 pm
by Gian-Carlo Pascutto
There's still so much room for experimentation. And I suspect that trying and training new approaches is quite a bit faster on a 20M NNUE net compared to a Leela one...

Dream situation for computer chess.

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Wed Aug 05, 2020 10:41 am
by Ovyron
This is the breakthrough I've been predicting for years! It finally happened, and I'm just glad chess looks very healthy after it (draw rates have gone down! Some lines are proving to be playable again!)
Gian-Carlo Pascutto wrote:
Tue Aug 04, 2020 6:01 pm
There's still so much room for experimentation. And I suspect that trying and training new approaches is quite a bit faster on a 20M NNUE net compared to a Leela one...

Dream situation for computer chess.
I agree, a Stockfish 12 release could be premature. These are new unexplored terrains, the code is nowhere near being optimized, the nets are nowhere near properly tested, imagine if SF 12 is released and one month later there's another 50 elo jump, we want a SF 12 that lasts for long!

This makes Alpha Zero look like peanuts! Bring back Deep Mind!

Re: this will be the merge of a lifetime : SF 80 Elo+

Posted: Wed Aug 05, 2020 12:28 pm
by MikeB
Ovyron wrote:
Wed Aug 05, 2020 10:41 am
This is the breakthrough I've been predicting for years! It finally happened, and I'm just glad chess looks very healthy after it (draw rates have gone down! Some lines are proving to be playable again!)
Gian-Carlo Pascutto wrote:
Tue Aug 04, 2020 6:01 pm
There's still so much room for experimentation. And I suspect that trying and training new approaches is quite a bit faster on a 20M NNUE net compared to a Leela one...

Dream situation for computer chess.
I agree, a Stockfish 12 release could be premature. These are new unexplored terrains, the code is nowhere near being optimized, the nets are nowhere near properly tested, imagine if SF 12 is released and one month later there's another 50 elo jump, we want a SF 12 that lasts for long!

This makes Alpha Zero look like peanuts! Bring back Deep Mind!
To both - that is exactly why it should be #12 - set a marker in time from this point forward to measure progress. They always test regression against a prior release - what better time to set new marker once you merged a PR that is a 'significant" change - you will always want to measure progress from this point forward against the initial release.