SlowChess Blitz Classic 2.0

[Damir]

Congratulations on such a nice elo improvement. Let's hope you continue in finding ways on improving your engine, as it is very difficuilt to do.

[Modern Times]

Slow Chess Blitz Classic 2.2 is now released.

With 100 self-play elo increase I'd hope for aroud +70 elo in blitz rating lists.

After the first 500 FRC games I've played it is +77 Elo (Ordo = +86 Elo)

I have had three disconnects which isn't a lot be any means but I don't recall getting any with the last couple of versions.

I have not played against Allie yet, and v2.1 performed well against that, so there could potentially be further improvement to come.

[jonkr]

Yeah I'm expecting to hit a wall at some point for elo gain. Was having trouble finding measurable gain for a week until everything starting falling into place for this release quicker than I thought possible, mainly from history bug.

My immediate plan is a couple week break then I'll probably try just continuing what I've been doing first. If that becomes uninteresting because of inability to make measurable improvement, decide what I want to do next. Either try open source and open bench, take a long complete break from chess, or work on features (or experiments) unrelated to gaining elo, or some combination.

More about my testing results :
Before release I tested most against Xiphos 0.6 and by the end it was beating it consistently (between +1 to +12 elo) under my test conditions, but given amount of possible variables like opening book, time control, hardware, tablebases, opponents, error bars, I don't really expect it to be higher in rating lists just somewhere close.

4 Cpu testing vs 4 cpu Xiphos was the opposite however, it never beat Xiphos in a match and scored an average of 15 elo less than it did in the 1-cpu matches. I only ran a couple matches given how much longer matches take using 4 cores, but multi-cpu performance definitely seems like something I need to look into and see if there is anything that needs fixing.

Modern Times :
Thanks for testing, that sounds like a good elo gain, it was always stronger in FRC in my tests than standard, and I did more FRC training games this time around too.
Hmm the disconnects are worrisome, and single core even, I don't remember getting any but will look into it.

[Graham Banks]

Impressive progress, Jonathan.
I genuinely hope that you don't lose interest again.

[jonkr]

Thanks Graham, currently I'm pretty confident I will at least do something more on Slow Chess.

After the first 500 FRC games I've played it is +77 Elo (Ordo = +86 Elo)

I have had three disconnects which isn't a lot be any means but I don't recall getting any with the last couple of versions.

I have not played against Allie yet, and v2.1 performed well against that, so there could potentially be further improvement to come.

For the disconnects I did try a 2000 game FRC gauntlet and got 1 disconnect, my guess is it's tablebase related in some way since the last eval was a won tablebase evaluation, and I didn't find any disconnects in my old pgns and don't usually test with tablebases on. The strength gain from 5-piece tablebases seemed hard to measure when I first tested, so for now I would suggest not setting the syzygyPath since crashing isn't good. (Unless you still get disconnects without tablebases then I need to keep looking.)

[AndrewGrant]

Thanks Graham, currently I'm pretty confident I will at least do something more on Slow Chess.

After the first 500 FRC games I've played it is +77 Elo (Ordo = +86 Elo)

I have had three disconnects which isn't a lot be any means but I don't recall getting any with the last couple of versions.

I have not played against Allie yet, and v2.1 performed well against that, so there could potentially be further improvement to come.

For the disconnects I did try a 2000 game FRC gauntlet and got 1 disconnect, my guess is it's tablebase related in some way since the last eval was a won tablebase evaluation, and I didn't find any disconnects in my old pgns and don't usually test with tablebases on. The strength gain from 5-piece tablebases seemed hard to measure when I first tested, so for now I would suggest not setting the syzygyPath since crashing isn't good. (Unless you still get disconnects without tablebases then I need to keep looking.)

Code: Select all

Started game 117 of 128 (Ethereal12.25 vs SlowChess2.2)
Warning: Illegal PV move e2d1 from SlowChess2.2 (1)
Warning: PV: hxg2 Nxc7 Qc6 Nxe8 Qxe8 Kxg2 Ne6 h4 Qc6 Nf3 Bb7 Ng5 Nxg5 hxg5 Qxc2 Rd2 Qxb3 Qb1 Qb4 Qxb2 Qxe4+ Kf1 Qg2+ Ke2 Qf3+ Ke1 Qh1+ Ke2 Qf3+ e2d1

[jonkr]

Code: Select all

Started game 117 of 128 (Ethereal12.25 vs SlowChess2.2)
Warning: Illegal PV move e2d1 from SlowChess2.2 (1)
Warning: PV: hxg2 Nxc7 Qc6 Nxe8 Qxe8 Kxg2 Ne6 h4 Qc6 Nf3 Bb7 Ng5 Nxg5 hxg5 Qxc2 Rd2 Qxb3 Qb1 Qb4 Qxb2 Qxe4+ Kf1 Qg2+ Ke2 Qf3+ Ke1 Qh1+ Ke2 Qf3+ e2d1

Were there any issues other than the warning? I've noticed occasional illegal PV move warnings since I started using cutechess-cli to test after 1.9, but has been low on my list to track down since until now I haven't encountered any issues completing the test games.

[AndrewGrant]

Were there any issues other than the warning? I've noticed occasional illegal PV move warnings since I started using cutechess-cli to test after 1.9, but has been low on my list to track down since until now I haven't encountered any issues completing the test games.

A few other examples of Illegal PVs, but no other issues. I'm going to post all the games in the match section of the form.

Also, in case you are not aware yet, your Linux release is compiled with symbols. I'm not sure whether you are trying to hold onto trade secrets or anything, but your current release is easily decompiled by someone motivated.

[jonkr]

No trade secrets, I'm realistic that there are top engines that are open source and way stronger, and I try to openly share any ideas that are interesting. Still I would probably recompile without symbols but don't have easy update process and never installed Linux off the thumb drive, so too lazy to do it. For now though I still like the idea of a personal engine. (And someone did once alter a public source program I made and release it commercially but not in the US and I doubt it made much money.)

[Modern Times]

For the disconnects I did try a 2000 game FRC gauntlet and got 1 disconnect, my guess is it's tablebase related in some way since the last eval was a won tablebase evaluation, and I didn't find any disconnects in my old pgns and don't usually test with tablebases on. The strength gain from 5-piece tablebases seemed hard to measure when I first tested, so for now I would suggest not setting the syzygyPath since crashing isn't good. (Unless you still get disconnects without tablebases then I need to keep looking.)

Still just the three disconnects in more than 900 games now, so it is a very minor issue.