Just found a 55 elo bug in SlowChess 2.1

[jonkr]

Didn't notice so much extra discussion, I'll go back and read the technical discussion.

Just wanted to say that the approximately 55 elo from the history fix has held over a total of 30000+ bullet games against a variety of engines (not pure test since I started testing with other changes too, but getting somewhat stronger results after the other changes instead of weaker.) History with its interaction with LMR / LMP is worth a lot of elo.

I'll try to release new version within couple weeks, seemed like good time to get in some minor fixes and try some changes that otherwise I wouldn't want to put out another new version for. Also I just finished the eval tuning/learning write-up to release with it.

As far as I know the 2.1 Linux version on the website works fine as multi-threaded engine but will try some more Linux testing to make sure.

[jonkr]

To be specific it was testing at +55 elo self-play, +46 elo against other engines. I just meant that it wasn't something that at first appears to be +10 elo in self-play, then shows no gain in later gauntlet testing which happens surprisingly often, although my error bars are pretty wide so tough to say if noise or matchup specific. (Also I assume would be less at long time control.)

[Alayan]

I got annoyed by the importance of history being dismissed, so I had a look at Ethereal's history code, and got an elo gainer by disabling history updates when depth=1 and the 1st quiet tried failed-high.

http://chess.grantnet.us/test/6265/

Code: Select all

ELO   | 6.19 +- 3.83 (95%)
SPRT  | 60.0+0.6s Threads=1 Hash=64MB
LLR   | 2.96 (-2.94, 2.94) [0.00, 5.00]
Games | N: 8081 W: 1109 L: 965 D: 6007

Probably +2 elo gain. History heuristics are an elo goldmine.

[Gary Internet]

As far as I know the 2.1 Linux version on the website works fine as multi-threaded engine but will try some more Linux testing to make sure.

That's great news. I'd love to see version 2.2 playing at TCEC in Season 19 which won't start until some time in July, so you have plenty of time to fix bugs, do Linux testing, tuning etc. TCEC desperately needs more actively developed engines at the strength level that SlowChess 2.2 will come in at i.e. a solid League One level competitor (like Booot 6.4) that has a chance of getting into the Premier Division and will at least be competitive with the likes of Rofchade and Fire, so they don't just become dominant in League One.

[cucumber]

The rate at which you’re gaining elo is incredible. Have you considered putting your engine onto OpenBench? It lets the community contribute CPU power to testing and helps make sure some eyes glance over your code for these sorts of bugs.

[jonkr]

Alayan : I actually am not updating history at all on depth 1, it's one of those things that I never was able to get any testing results on suggesting a measurable difference either way, but next time I'm working on search will try out some more different history stuff. I was previously able to make more history related adjustments to gain a few more elo after my fix.

Gary Internet : I'm not sure how they decide program roster but I do think would be fun to be in those tournaments. My hope would be to do well enough someone makes a youtube analysis video of one of the games, probably titled something like "Leela brilliantly dismantles Slow Chess Blitz"
Will release a Linux version of 2.2 too when I update but not Linux expert, only know it worked when I booted from my computed from an Ubuntu USB to test it.

Cucumber :
Going public source on Github and Open Bench is a future possibility. Not really sure. Extra testing capability may at some point make working on it interesting if I have too much trouble finding anything that improves it. Right now just a personal project to tinker with and make quick changes to try stuff out.

[Alayan]

TCEC and CCC give out invitations to new promising engines. Running on Linux, not having crash issues, and having SMP support are the main requirements.

I think they'd be very happy to have SlowChess on board. I can always tell them to look at SlowChess if needed, just make sure you have a working Linux compile. TCEC would most likely be happy to let SlowChess play some bonus games on their hardware to help test it for crashes or performance issues at high thread count.