New engine releases & news 2021

[amanjpro]

Rustic doesn't have a particularly simple approach. What do you mean with "When I do a stupid change"?

I mean, when I introduce a bug in Zahak, Rustic catches it fast. As it doesn't have much pruning, it is strong tactically and good at catching my stupidities.

I have noticed that Rustic is tactically stronger than most engines in its rating range; probably because it is faster, and can often see 1-2 ply deeper. It has a simple evaluation, but as I'm an OK chess player myself, I think my PST's are better than the general PST's floating around on the internet; so even though the evaluation is simple, it is not entirely stupid. The PST's do include some positional and strategical knowledge, to some extent. Tuning will make this better, probably.

I highly doubt it is because of the PSTs, after all PST only teaches the AI where to put pieces, not what to do with them. Rustic is good, because it does close to no pruning. Some tactical strength fades away when you introduce lots of pruning/reductions. It is a balance that is not super easy to find.

As said, Rustic isn't really simple per se, but it also doesn't have a lot of features yet. What I do know, is that the features it does have are (probably) bug-free. While testing the version with the TT, it did gain 105 Elo, but I wasn't satisfied because I thought, according to tests with other engines, the gain should be bigger; and in the end I found a bug (saving the wrong move in the TT in the beta-cutoff) that, after fixing it, gained another 60 Elo.

It is very easy to add a feature in the engine, see that it gains 50 Elo, and call it done... while the feature _could_ have gained 75 or 100 Elo. Do that a few times, and your engine will become stronger, but not as strong as it could have been. If you're particularly unlucky, a feature, if it has a serious bug, could even make the engine weaker.

I looked into Rustic and compared it feature wise to Zahak, while I have way more search functions, I don't have multi-stage move generations and magic bitboards, that puts Zahak at a huge disadvantage. I wrote the entire move generation first, then read the literature, my stupidity. Also, Go vs Rust is a huge advantage for Rustic too.

I tuned the eval function almost 100% from scratch, I picked PTSQ from somewhere, but touched it heavily that it doesn't resemble the original anymore. I wanted the engine to play more or less like me

Initial results of implementing magic bitboards adds around 100nps

[Gabor Szots]

Berserk is now at version 3.2.0.

[Guenther]

Bit-Genie 2 released
https://github.com/Aryan1508/Bit-Genie

New SF dev(s) released yesterday
https://abrok.eu/stockfish/

[Frank Quisinsky]

Hi there,

1. Last Demolito version is from March 29th, 2021
https://ci.appveyor.com/project/lucasar ... /artifacts

2. Gogobello 3.0
http://sasachess.altervista.org/gogobello/index.html

gogobello 3.0: (NNUE Revolution)
Status: released in April 2021
* - Pure NNUE Evaluation
* - Neural Network included in the binary

Best
Frank

[Guenther]

Hi there,

1. Last Demolito version is from March 29th, 2021
https://ci.appveyor.com/project/lucasar ... /artifacts

2. Gogobello 3.0
http://sasachess.altervista.org/gogobello/index.html

gogobello 3.0: (NNUE Revolution)
Status: released in April 2021
* - Pure NNUE Evaluation
* - Neural Network included in the binary

Best
Frank

Note that Gogobello has removed all HCE in this version and only uses the current SF NNUE, so it will not appear in my chronology
and that was the reason it wasn't mentioned.
May be it is an aprils fool, who knows. (The embedded NNUE doesn't work anyway and one must download the file from SF)

'Real' releases were:

Berserk 3.2.1
https://github.com/jhonnold/berserk

GreKo-PSQ 1.0 (experimental version)
http://greko.su/index_en.html

FoxSEE 6.3.1 and 6.3.2
https://github.com/redsalmon91/FoxSEE/releases

[NoChessNoLife]

New LcFiSh Neural Network has been released. It is based on Leela T60 training.
You can read about it here:
http://lczero.org/blog/2021/04/jumping- ... bandwagon/
Binaries can be found here:
https://github.com/borg323/lc0/releases/tag/lcfish

[Guenther]

New LcFiSh Neural Network has been released. It is based on Leela T60 training.
You can read about it here:
http://lczero.org/blog/2021/04/jumping- ... bandwagon/
Binaries can be found here:
https://github.com/borg323/lc0/releases/tag/lcfish

sounds like

Code: Select all

2021-04-01

[Gabor Szots]

For anyone interested, I have put a Megalodon 0.3.2 64-bit compile for Windows here:



I expect it to run on any 64-bit hardware.

[Guenther]

For anyone interested, I have put a Megalodon 0.3.2 64-bit compile for Windows here:



I expect it to run on any 64-bit hardware.

I have added of course two compilations too a few days ago. (see chronology)
0.3.2 can still make illegal moves from time to time though...and it has a stalemate bug.

[Gabor Szots]

I have added of course two compilations too a few days ago. (see chronology)
0.3.2 can still make illegal moves from time to time though...and it has a stalemate bug.

OK, I'm not going to pester any more.