Your favorite engines?

[Madeleine Birchfield]

One side effect of the really big nets like Dragon 2.5 and SF14.1 is that you now need at least 10b positions to properly train them. SF can rely on its leela farm team for the bulk of it’s data, and Dragon goes through at least 20b so far for its reinforcement learning (RL). But few of the other NNUE engines have made that jump. It’s just too expensive. That’s why you won’t see any more free nets from me — time and cost are prohibitive.

You might see some more commercial engines to underwrite net training, I suspect. That or data center operators with deep pockets will have to donate lots more resources.

Understandable, Ethereal and Revenge have already become commercial, and Igel set up a Patreon for donations last year to pay off the cost for training. Mark Jordan has locked up his Stein (for Leela) and StockfiNN (for Stockfish) nets behind his Patreon page. I wouldn't be surprised if other engines/net trainers went that route, but I probably wouldn't use their engines if they went down that route.

I think Albert Silver could have done fairly well if he partnered up with Frank Schneider, implemented NNUE in Fritz 18 and trained the Fat Fritz 2 net for the NNUE version of Fritz. Now that Fritz 18 with the Fat Fritz 2 net would have been a good engine to go with the Fritz GUI, and Albert Silver, Frank Schneider, and Chessbase would likely have been praised for their accomplishments in the computer chess community rather than derided.

[Frank Quisinsky]

Hi Madeleine,

my two main points in computer-chess:

1.
I try to find out the most aggressive engines with perfect king safty with many pieces on board.
Easy, I like the openings and to create opening databases with equal positions.
A very interesting work.

After my still running FCP-Tourney-2022 the next FEOBOS project will be start.

I need strong statistics for identify the strongest 8-10 engines for opening analyzes.
This 8-10 engines will build a team for ... find the equal A00-E99 opening positions, 4 moves after ECO code formed.
FEOBOS v2 I will start with Klaus Wlotzka next year (Excel expert).

2.
All the time I have most fun with engines can play amazing king attacks.

Example to older Winboard times:
ETChess, Anaconda, Aristarch, Phalanx, Glaurung, AnMon
All the commercials are not the king attacker (not Shredder, not the last Fritz versions by Frans Morsch and not Gambit-Tiger). Hiarcs by Mark Uniacke are all the time interesting or Gandalf ... not 100% attacker but aggressive engines!

In chess-computer times ...
More often I saw that from Novag as from Fidelity.
Nigel Short by Ed Schröder ist great with 10Mhz only (tuned).
TheKing with his special mate-search was great.

Most computer-chess people I know have most fun with aggressive chess programs (during watching the games), I also.

Today we have a lot interesting engines.
Spark 1.0 for an example.
Have around the same strength TOP-10 humans have and can play human-like-style, is a great king attacker.

It must not be a TOP-50 engine!
Means, I can have fun with AnMon (2400 Elo, 1 core), with Spark (2775 Elo, 1 core) and of course with my favorite engines: Wasp, SlowChess and Booot.

Best
Frank

[Rubinus]

From the older ones: Junior - interesting game with different material, pity that the development seems to have stopped. King - the configurability is amazing and there is no need to invent anything, it can be copied from Chessmaster. Good to play for me. Smarthink and Gandalf, used to be commercial, now they're also downloadable somewhere and still a pretty interesting game. Although Gandalf has tactical gaps at times, too little depth. Of the very old free ones, I enjoyed Amyan and Zeus. The latter seemed to me to be very positional, but the depth was at 90s levels and it worked out tactically accordingly, plus it's only WB so not exactly stable in a ChessBase environment. Of the newer ones, maybe Andscacs - quite a peculiar positional game, but capable of attacking hard. For analysis I now use Stockfish and sometimes LC0, but I don't have the best graphics for the latter, so it has to be checked with Stockfish for tactics. I'm running a themed king gambit tournament on my old computer now, and ShashChess looks very good. Otherwise I enjoy supporting weaker commercial engines, the competition is always good, so Ethereal, Revenge, Chiron and Dragon, although the last one seems to play almost like Stockfish, only a bit worse. I used to like Hiarcs, but with his latest releases he seemed to have stagnated gameplay-wise and benefited more from his own really good library. Moreover, for an engine presented as intelligent he usually did very badly in gambits, for example once he lost I think to Toga White after 1.d4 e5?!.

[Madeleine Birchfield]

Dietrich, I thought Leela's data were free, and I see no reason why other authors could not use them. You yourself made noticeably different nets using the same data -- I think someone could make even better use of these data and create an original, different engine.

Unfortunately, the current set of authors (Andrew Grant, Connor McMonigle, et cetera) would rather all generate their own data from their own engines rather than use other engines' data, as they consider using external data sources for training to result in an engine not unique enough.

[matejst]

I think Albert Silver could have done fairly well if he partnered up with Frank Schneider, implemented NNUE in Fritz 18 and trained the Fat Fritz 2 net for the NNUE version of Fritz. Now that Fritz 18 with the Fat Fritz 2 net would have been a good engine to go with the Fritz GUI, and Albert Silver, Frank Schneider, and Chessbase would likely have been praised for their accomplishments in the computer chess community rather than derided.

I agree. I tried the free FF and I thought it was a good net, and that Albert did some good work.

About commercial engines: in general, I would prefer complete packages: a GUI AND and engine, and I would be ready to pay for both, but not for an engine only when there are so many excellent engines. Then, above 3200 Elos it simply makes no sense: in tactical positions the quality of the analysis is the same, and in quiet positions the difference for a human is almost irrelevant.

[dkappe]

One side effect of the really big nets like Dragon 2.5 and SF14.1 is that you now need at least 10b positions to properly train them. SF can rely on its leela farm team for the bulk of it’s data, and Dragon goes through at least 20b so far for its reinforcement learning (RL). But few of the other NNUE engines have made that jump. It’s just too expensive. That’s why you won’t see any more free nets from me — time and cost are prohibitive.

You might see some more commercial engines to underwrite net training, I suspect. That or data center operators with deep pockets will have to donate lots more resources.

Dietrich, I thought Leela's data were free, and I see no reason why other authors could not use them. You yourself made noticeably different nets using the same data -- I think someone could make even better use of these data and create an original, different engine.

I never used stockfish or leela data. I did use Bad Gyal data, which is very different, and data generated from other engines or sources.

How one would produce something different when using the same data as everyone else is an interesting question.

[matejst]

From the older ones: Junior - interesting game with different material, pity that the development seems to have stopped. King - the configurability is amazing and there is no need to invent anything, it can be copied from Chessmaster. Good to play for me. Smarthink and Gandalf, used to be commercial, now they're also downloadable somewhere and still a pretty interesting game. Although Gandalf has tactical gaps at times, too little depth. Of the very old free ones, I enjoyed Amyan and Zeus. The latter seemed to me to be very positional, but the depth was at 90s levels and it worked out tactically accordingly, plus it's only WB so not exactly stable in a ChessBase environment. Of the newer ones, maybe Andscacs - quite a peculiar positional game, but capable of attacking hard. For analysis I now use Stockfish and sometimes LC0, but I don't have the best graphics for the latter, so it has to be checked with Stockfish for tactics. I'm running a themed king gambit tournament on my old computer now, and ShashChess looks very good. Otherwise I enjoy supporting weaker commercial engines, the competition is always good, so Ethereal, Revenge, Chiron and Dragon, although the last one seems to play almost like Stockfish, only a bit worse. I used to like Hiarcs, but with his latest releases he seemed to have stagnated gameplay-wise and benefited more from his own really good library. Moreover, for an engine presented as intelligent he usually did very badly in gambits, for example once he lost I think to Toga White after 1.d4 e5?!.

I cannot use Smarthink afther the 1.90 version -- did not try too hard, though. I probably lack some dll on my computer. I also liked a lot of old engines -- I still have Gandalf 7.0, Zarkov and Frenzee installed, but the quality of play has improved too much with the neural nets. I did not buy HIARCS since CB's h10, so I would not know. I liked it very much back then.

[matejst]

I never used stockfish or leela data. I did use Bad Gyal data, which is very different, and data generated from other engines or sources.

How one would produce something different when using the same data as everyone else is an interesting question.

I meant that you made TWR and DH from roughly the same data, not that you used Leela's data.

[Frank Quisinsky]

We build the legend with Fritz??!

Fritz 5.32 is a very aggressive engine but comes with a catastrophic endgame.
Same for MMV and many other earlier chess computers.

Back to Fritz ...
From version to version Fritz is stronger and lost more and more the aggressiveness.

Today people are thinking ...
If chess programs are stronger (NN file for a modern example) the programs lost his face and lost the nice playing style.

Complete nonsense in case of Neural Network.
In case of Fritz it was right. The endgame are better with newer versions but the engine lost the own face (version 5.32). And to the end of Fritz development by Frans Morsch the engine have the hightest draw quote from all TOP-20 available programs. Very rarely are king attacks. A complete other program if we compare Fritz 13 with Fritz 5.

Legend build with Fritz, people have such an opinion because most thinking on Fritz development.
Stronger = Stronger in endgame only and the nice playing style is lost.

Rybka ...
Thinking on all the grandmasters and the great opinions by Grandmasters about Rybka. An Engine with 3000 Elo (today not in TOP-40). Three TOP-50 engines today have such a catastrophic king safty with many pieces on board Rybka have. But in times Rykba is on position 1 we all can't see that. Today with statistic and much stronger opponents very easy to see.

Rybka is very strong in endgames.
Endgames, most of club-players don't like.
But the club-players like Rykba because it was the number 1.

Playing styles of engines ...
A big topic and not easy for us and for Grandmasters (if I am thinking on opinions to Rybka).

Best
Frank

[Madeleine Birchfield]

Frank,

We seem to have different tastes in our engines. I very much prefer an engine that could understand the opening (especially important for opening theory research and preparation), play well positionally and able to grind out a slow win in a closed position, and engines using neural networks are better at that than engines which do not use them at all.

Agreed to you about endgames not really mattering much, which is why I don't like the later versions of Fritz (as they are terrible in the opening and middlegame where it actually matters), and I personally prefer Koivisto and Berserk over Slowchess, as Koivisto and Berserk are stronger than Slowchess in the opening and middlegame while Slowchess is extremely strong in the endgames.