Clover 8.1

Hello again, new version is here! From last release, I have tried to massively improve the readability of my code, it's still WIP but I think it's getting there. The gains should mostly come from search and bug fixes.

Although I didn't want to release now, I have less and less free time nowadays, so I figured I could release and worry about improving Clover some other time.

As per usual, have fun and let me know about any potential bugs!

Clover 8.0.2

This is a bug fix release, for Mac users. If you aren't one, you should ignore this release.
Stack size has been increased so that there is no more crashes.

Thank you Dariusz for reporting! :)

Clover 8.0

New version is here! This time with a new arch, using 7 king buckets. This scheme is eye-balled based on the distribution of the king square. As per usual, there is also many non-network changes, such as improving code quality. I really like how Clover looks like right now, but I know it can be improved.

I know there has been quite some talk about the so called SIMEX, a very valuable tool, which detected clones in the past (up to 5 years ago). I've had someone run some SIMEX tests on Clover and some other engines and well the results speak for themselves.

                      Clover 7.1-dev   Motor 0.7.0 dev       Obsidian 12.0  Viridithas 12.0.0
   Clover 7.1-dev            100.00%             51.82%             60.04%             59.01%
 Motor 0.7.0 dev              51.82%            100.00%             57.74%             48.52%
    Obsidian 12.0             60.04%             57.74%            100.00%             57.96%
Viridithas 12.0.0             59.01%             48.52%             57.96%            100.00% 

There definitely is going on some kind of... net sharing... like clearly, Clover uses self play data, Obsidian Leela... ugh wait this doesn't make sense... aha, it must be network architecture... ugh wait they are clearly different... then it must be training procedure! Ugh, but Obsidian uses multi step procedure... I have no idea, it mush be the shared ideas on Discord!! Although the test is done at DEPTH=1, this totally makes sense!!

Thanks to everyone who has aided me directly (Ciekce, first person to PR gainers to Clover :) ) or indirectly.
Have fun and report any bugs!

P.S.: Thanks Ipman for the avx512 compile!

Clover 7.1

New release is here, quicker than ever! Expected gains around the 15-20 Elo LTC, majorly due to LTC tune.
A big part of the code has also been refactored into more C++ code (at least I would like to think so).

As always, have fun and report bugs!

Ty Cie, for avx512 again! :)

Clover 7.0

Hello world again, Clover 7.0 is here! This time with a new net architecture:

• Using SCReLU instead of ReLU
• 1280 Hidden Layer size instead of 1024

but also a lot of search changes! Sadly, Ponder mode is removed, there was no interest in it and it was also bugged.

Expect around 25-30 Elo, this time scaling as you use longer time controls.

Have fun and report bugs! :)

Ty Cie for avx512 build!

Clover 6.2

After 6 long months of working occasionally on Clover, v6.2 is here. I expect around 40-45 LTC elo gains. Nothing out of the ordinary, just new search ideas being added and new net trained.
Expect next version to be huge, I will be working on new net arch :).

Have fun and report any bugs if possible!

Clover 6.1

New release is here, slower than ever! I expect 40 elo STC, 45 elo LTC, 95 elo DFRC STC (hype). Ponder is definetly more normal than last release, maybe not perfect, I would give it a try nonetheless.

Changes:

• Net arch is unchanged suprisingly, but I have discovered an interesting training pipeline, which I will continue to use until I change arch. Sad thing is it takes 2 days to train a new net 😩
• Various search changes, attempt at LTC scaling. I observed that Clover scales pretty bad, hence I am still trying to figure out way. Hopefully fixed in upcoming versions

Thanks to everyone that helped during development and... have fun using Clover :)

Clover 6.0

New release is here! Progress is slowing down a bit and it will continue to do so, as uni time is coming soon :).

Nevertheless, Elo gains are decent in standard chess, 35-40 self-play Elo, which hopefully translates nice vs other engines.
In (D)FRC, gains are massive, around 145-150 Elo!!
Aaaand, as requested, Clover now supports Ponder! :) Although, I am 95% sure it works, I'd rather get some external feedback on it ;)

Some notes about changes:

• New network arch with 1024x2 neurons for L2, trained on DFRC data, which is the main reason the DFRC gains are so massive.
• Various search changes and some cleanups

That's it for now, have fun trying Clover and report any bugs! ;)

Clover 5.0

New release is here! I expect again something around 40-50 elo stronger than last version.
Network arch is the same sadly, since other net arches require more data, which I can't really generate due to being out of space.
Next release might come after a long time, since I am at the same time starting uni in september.

This release also supports Fischer Random Chess! I have yet to train on DFRC games for a stronger FRC performance. :)

Anyway, have fun and report any bugs ;)

Edit: fixes illegal castle moves in Arena

Clover 4.1

After 1 month, new release is here. I expect it to be around 30-50 elo stronger than previous version. It has many compilation fixes, especially for Mac.
Besides the usual search tweaks, Clover now has a 5 king buckets arch, which gained pretty nice (15-20 elo).

Other than that, Clover is now part of the Engine Programming OB instance, which has massively sped up its progress. Many many thanks to these guys: @archishou, @crippa1337, @dede1751, @spamdrew128, @Alex2262, @cosmobobak, @Adam-Kulju, @GediminasMasaitis, @uwuplant, @Ciekce, @JouanDeag.

And of course, don't forget to report any possible bugs! ;)