RubiChess NNUE player implemented

[AndrewGrant]

I don't need 100x the same idea's in programs with different names.

That is the way the wind is blowing ...

[Frank Quisinsky]

Hi Andrew,

remember on all the different IPPs.
Today not longer a big topic.

The work the Fire programmer made was great or the ideas the GullChess programmer have (clearly better king safty ... Fruit / Rybka SAGA). All the 100-games IPP matches between same engines I can found here. Of course 1-Minute per game matches. What the people like to do, without to looking in the games ... not possible with such fast time controls. Now the same with LCO / NNUE / Stockfish ... a very boring storry!

So I am thinking different times comes and go.
For the moment it's NNUE and x programs used the same ideas.

FACT is we have a lot of wonderful and very own chess programs.
I am sure that not all of the programmers jump of the same train in speedy Gonzales manner.

To proud with the work they do in so many years of chess programming as to use such ideas.
I can life with it if I find Wasp on rating systems on place 200 if 180 others used NNUE or other same ideas!

I think to have NNUE is interesting for Stockfish and analyzes positions or complete games.
For me enough if Stockfish have it alone.

Best
Frank

[carldaman]

An engine suffers from a lack of eval tunability once it uses a net for its evaluation. You can't even set contempt to anything any more, as it won't recognize the setting (if there is one).

Another related problem is that SF + most nets, although very strong, display a tendency towards drawishness with the Black pieces against moderately weaker opponents that it should normally be able to beat more frequently based on pure strength differential. [Edit: primarily at longer TCs]

It is as if the engine plays not lose with Black, rather than win. This safety-first approach is a big problem as I see it, as a user interested in good games and analysis conducive to creative play. Whatever wins occur with Black from balanced openings, or with no book used, tend to be long, drawn-out grinds. Many wins with White are like that, too, although some nice attacks are not uncommon from the White side.

Considering the monstrous strength and clear superiority these nets hold over other engines, I have to be disappointed in this aversion to risk that can't be tuned away. All the tinkering would have to be done beforehand in the net training phase, to produce a net capable of playing all-around aggressive chess - something that's a lot easier said than done. Otherwise, the evaluation then becomes a black box and it's too late to tinker and adjust anything not related to search.

[Gabor Szots]

I won't upload binaries so you have to compile yourself.

Any compile available yet?

I have put one here: https://drive.google.com/drive/folders/ ... IreoNOB2xX

Requires BMI2 CPU.

Not compressed as that causes a warning that I will be denied Google services if I continue to spread malware.

For the same reason, I will delete it within a couple of hours or at most one day.

[chrisw]

The net is (almost) everything, the engine is just a tool to play the music.
In future rating lists and tournaments the creators of the net files will decide about winning or losing, not the engine author turning some screws for 2.5 Elo.

This is what I have been saying. Most people agree, barring a few special cases of brain-dead individuals.
Where is the interest? Where is the intrigue? Where is the novelty? Where is the nuance? Hell, where is the Chess?
If the inevitable happens, and all top engines go into pasting NNUEs, then I'm out. Its just not interesting, and I won't be apart of the scene.

Interesting comment. You know, well maybe not, the real days of comp chess as an exciting and interesting place we’re in the 1990’s and actually on rec games chess computer. The reason? Well, technologically we had broken the 5-6-7 ply limit, microcomputers existed, en sufficient masse, for there to be independent non-academic authors, and the evaluation function plus trying to have a directed smarter search was the key element. On the other side were the nps, soon to become the Elo is all merchants, and to some extent the programs reflected the (generally crazy and extreme) social/political nature of the programmers, do we had wild mixed chess “theory”, program “theory” and politics all blended into one giant argument. Plus some quite smart non-specialists joining in. I’m afraid that the nps, chess is tactics, Elo is everything clique won out, even though they were actually wrong. I got lifetime banned of course by this site. Nowadays, I’m finding it extremely hard to get “style” into my proto engine, it’s possible, but auto-tuning on result is a crippler, one loses control of style choice to an algorithm. As is Elo Elo Elo mantra.

[perejaslav]

Requires BMI2 CPU.

I have POPCNT support only. But thanks anyway

[dkappe]

It's OK for me but I hope that the group of programs using NNUE will be small because I don't need 100x the same idea's in programs with different names.

And for dkappe:
I remember you developed an own chess engines with around 2.200 Elo.
Not strong enough that you are using the main work by Fabien (Fruit - Toga)?

Hmmmm, 2200 elo? I did run Bad Gyal 9/Little Ender in A0lite, which is a python based mcts/nn engine, but is closer to 2850 ccrl. It did finish joint last in this season’s QL with 5.0 points. In scorpio, Allie or lc0 it would be closer to 3200 ccrl. Even in the pokey A0lite at 450 nps, it comfortably beats Fruit on 1 cpu.

I think you misunderstood me. One can generate nnue training data from a uci engine. I’ve done this with Toga II 4.0.2, Komodo, ICE and Bad Gyal, producing a different series of nets from each of them.

The good news for you is that, like the different engines that produced them, they all have very different styles of play.

[dkappe]

Another related problem is that SF + most nets, although very strong, display a tendency towards drawishness with the Black pieces against moderately weaker opponents that it should normally be able to beat more frequently based on pure strength differential. [Edit: primarily at longer TCs]

Might you be generalizing over all nnue’s from Stockfish’s play? Night Nurse showed quite the opposite tendency. Igel romped through TCEC L2 with many dynamic wins both as white and black, but started a string of draws once it ran into tougher competition in L1.

[dkappe]

Makes me wonder what a (depth=1) similarity test will reveal.

A perfect acid test for your similarity test. You have a number of my Toga and ICE based nets available to you. You can run them on a number of different NNUE supporting engines and compare them against the originals. I’d love to see the results of this test.

Based on various test suites, these nets seem more similar to their originator than other nnue based engines.

[Frank Quisinsky]

After all I know wrote a German programmer Matador (Knappe).
That's maybe a problem with all the Pseudo names here!

If you are't the programmer of Matador (I am thinking all the time) my sorry!

Best
Frank