Slow Chess Blitz Classic
Moderator: Ras
-
Gabor Szots
- Posts: 1479
- Joined: Sat Jul 21, 2018 7:43 am
- Location: Budapest, Hungary
- Full name: Gabor Szots
Re: Slow Chess Blitz Classic
Thank you, Jon. I am very glad at your future plans as well.
Gabor Szots
CCRL testing group
CCRL testing group
-
jonkr
- Posts: 178
- Joined: Wed Nov 13, 2019 1:36 am
- Full name: Jonathan Kreuzer
Re: Slow Chess Blitz Classic
Also I'd like to thank everyone who tested so far up to version to 1.8 and the rating lists CCRL, CEGT, FGRL. Not sure if if there are any lists I missed / don't know about. It's fun to see Slow is now holding its own against opponents that used to be too strong. Senpai 2.0 and Wasp 3.75 were a couple I occasionally tested against since v1.5.
And I've grown fond of FRC so I'm glad there's a rating list for it, the games can be fun to watch, it seems to produce games that are varied and sharp (messy? complicated? not sure best term) more often which I like.
And I've grown fond of FRC so I'm glad there's a rating list for it, the games can be fun to watch, it seems to produce games that are varied and sharp (messy? complicated? not sure best term) more often which I like.
-
Modern Times
- Posts: 3759
- Joined: Thu Jun 07, 2012 11:02 pm
Re: Slow Chess Blitz Classic
Yes I'm a huge fan. No opening book, the engine has to think from move 1, and the different starting positions really challenge the engines and give some wild games !
-
Alayan
- Posts: 550
- Joined: Tue Nov 19, 2019 8:48 pm
- Full name: Alayan Feh
Re: Slow Chess Blitz Classic
You've made a great work over the last few months, gaining a lot of elo in a short amount of time while also adding functionality. Few programmers can boast such results.
It's very understandable that you want to take a break, but I hope in a few weeks you'll get back to working on your engine, it has big potential, and maybe you could find tricks that would be useful to other authors, too.
Some thoughts :
- Now that SlowChess supports SMP, it'd be nice to see it in engine competitions. But it might need some debugging to work properly on high-threads machines.
- A Linux version would be greatly appreciated. It would also be useful for the aforementioned competitions.
- Improving tuning is something that has massive elo potential but is not so easy to do. For example, Ethereal has a 7.4M positions set, but this set's accuracy could take big improvements, as could its size (covering more situations, and limiting sss in rarer situations) ; while the tuner itself can't handle non-linear values which is a big limitation. Generating a better tuning set that the few already available is entirely doable, the downside is that it requires significant hardware time which nobody yet has been willing or able to concentrate for such a task. There isn't great incentive to share the results of such a work, either.
- I hope you get some better testing resources. If your engine had been open source, I'd have suggested joining OpenBench like RubiChess did recently. In any case, I think you could overtake Andscacs in a few months.
It's very understandable that you want to take a break, but I hope in a few weeks you'll get back to working on your engine, it has big potential, and maybe you could find tricks that would be useful to other authors, too.
Some thoughts :
- Now that SlowChess supports SMP, it'd be nice to see it in engine competitions. But it might need some debugging to work properly on high-threads machines.
- A Linux version would be greatly appreciated. It would also be useful for the aforementioned competitions.
- Improving tuning is something that has massive elo potential but is not so easy to do. For example, Ethereal has a 7.4M positions set, but this set's accuracy could take big improvements, as could its size (covering more situations, and limiting sss in rarer situations) ; while the tuner itself can't handle non-linear values which is a big limitation. Generating a better tuning set that the few already available is entirely doable, the downside is that it requires significant hardware time which nobody yet has been willing or able to concentrate for such a task. There isn't great incentive to share the results of such a work, either.
- I hope you get some better testing resources. If your engine had been open source, I'd have suggested joining OpenBench like RubiChess did recently. In any case, I think you could overtake Andscacs in a few months.
-
jonkr
- Posts: 178
- Joined: Wed Nov 13, 2019 1:36 am
- Full name: Jonathan Kreuzer
Re: Slow Chess Blitz Classic
During the periods of fast progress I was working on it basically full time, so that's definitely large part of the reason there was so much progress. After taking a full break for a few weeks, I'm thinking about returning to it more as an occasional hobby (possibly with addition of dedicated testing computer though.) So I'm expecting at minimum months before any updates.
A Linux compile could be more straightforward if there are no big issues, just never was a Linux person so need to look into it first.
After every release I usually am left thinking I could probably get a 30-50 Elo gain if I keep working, then after that might get too hard to improve.
Some Tuning details for anyone curious :
I may be over-optimistic on possible improvement from learning/training because I was surprised at how reasonable the tuned values seemed for every value. I also enabled some terms I wasn't sure about before since they showed correlation. Then in testing after making the changes did measure an Elo gain. The only thing that seemed obviously weird about the tuning was the biggest error reduction would have been to set the side-to-move bonus to over 1/3 of a pawn (so delta from switching stm actually double that) and I figured it was unlikely that would improve play. But stuff like the 6 different file scores for rooks got tuned values in the order I expected (open, half-open weak pawn, either of previous with opponent minor outpost, half-open pawn-defended opponent pawn, closed, closed behind own blocked pawn.) Even on tactical stuff values seemed reasonable, so I uncommented some stuff like the small penalty for rook can be attacked by a safe knight move, etc. Also I could have fixed some passed pawn tuning issues earlier/easier if I had done the tuning before. (Undervaluing advanced passed pawns & immediate promotion threats, over valuing passed pawns still on own side of board, over valuing passed pawns over-extended near opponent king.)
A Linux compile could be more straightforward if there are no big issues, just never was a Linux person so need to look into it first.
After every release I usually am left thinking I could probably get a 30-50 Elo gain if I keep working, then after that might get too hard to improve.
Some Tuning details for anyone curious :
I may be over-optimistic on possible improvement from learning/training because I was surprised at how reasonable the tuned values seemed for every value. I also enabled some terms I wasn't sure about before since they showed correlation. Then in testing after making the changes did measure an Elo gain. The only thing that seemed obviously weird about the tuning was the biggest error reduction would have been to set the side-to-move bonus to over 1/3 of a pawn (so delta from switching stm actually double that) and I figured it was unlikely that would improve play. But stuff like the 6 different file scores for rooks got tuned values in the order I expected (open, half-open weak pawn, either of previous with opponent minor outpost, half-open pawn-defended opponent pawn, closed, closed behind own blocked pawn.) Even on tactical stuff values seemed reasonable, so I uncommented some stuff like the small penalty for rook can be attacked by a safe knight move, etc. Also I could have fixed some passed pawn tuning issues earlier/easier if I had done the tuning before. (Undervaluing advanced passed pawns & immediate promotion threats, over valuing passed pawns still on own side of board, over valuing passed pawns over-extended near opponent king.)
-
Modern Times
- Posts: 3759
- Joined: Thu Jun 07, 2012 11:02 pm