MBCA 10x10: The strongest Grand Chess engine?

[Tord Romstad]

Since many years, I've been wanting to write an engine for Grand Chess, one of my favorite chess variants. Earlier this year, the inventor of the game – Christian Freeling – sadly passed away. This motivated me to finally implement the game, in his honor.

I originally wrote this as an Easter Egg game for the Marvelous Brass Chessplaying Automaton (the chess app I am developing for Apple platforms), but after it turned out to be much stronger than I expected, I decided to release it as a standalone UCI engine. The source code can be found here:

https://github.com/romstad/mbca-10x10

This wasn't written from scratch: I started with the normal 8x8 classical chess MBCA engine, extended the square sets (or bitboards, as many of you like to call them) to 128 bits, rewrote the move generator, ripped out the neural net, and trained a new small net with bullet, starting from random weights. Finally, after a few iterations of generating data and training stronger nets, I did an SPSA tune of the search parameters.

I'm surprised by how strong this thing is. I haven't been able to survive until the endgame in any of the games I have played against it. In my latest 2+1 blitz match against Fairy-Stockfish NNUE, MBCA 10x10 won by 199-1. Of course this is hardly a fair match: FSF can play tons of variants, and can't be expected to compete with an engine that is designed to play just a single variant.

The next steps will be to program some weaker human-friendly levels (much harder than writing a strong engine!) and implement the Easter egg game in my app. In the unlikely event that there is a significant amount of interest, I'll also consider writing a simple cross-platform Grand Chess app with human vs engine and engine vs engine play.

[hgm]

Very interesting. Fairy Stockfish shouldn't be exactly weak. Do you have any explanation for the huge strength discrepancy?

Did the FSF version you used use NNUE evaluation? (I am not sure anyone has trained a net for Grand Chess, but I understood it should be possible to do that with FSF.)

[Tord Romstad]

Very interesting. Fairy Stockfish shouldn't be exactly weak. Do you have any explanation for the huge strength discrepancy?

Did the FSF version you used use NNUE evaluation? (I am not sure anyone has trained a net for Grand Chess, but I understood it should be possible to do that with FSF.)

Yes, there is a FSF NNUE for Grand Chess, and I used that one. I suspect it's way too big (it's 159 MB) and trained on way too little data. By comparison, my own network is a simple (1600 -> 1408) x 2 -> 8 architecture, trained with bullet from about 1 billion positions, and weighing in at 4.6 MB.

I think the poor results of FSF are caused in part by the too big net, in part by a ton of small inefficiencies caused by having to be able to play almost any chess variant imaginable, and in part being based on the somewhat old Stockfish 14.

By the way, didn't you use to host online tournaments for the somewhat related variant of Gothic Chess? If that ever gets resurrected, I'd love to join. I think Grand Chess is a better game, but 10x8 variants are fun, too, and I think it would be very easy to adapt mbca-10x10 to handle the 10x8 board.

[Tord Romstad]

By the way, those of you with an iPhone or an iPad can now try out the Easter Egg game in the TestFlight beta of The Marvelous Brass Chessplaying Automaton. Long press on the app's logo to unlock the Easter Egg and play Grand Chess.

For now, the lower levels are just MultiPV with softmax move selection, which isn't very human-like and not much fun to play against. I hope to improve this soon, but I also have tons of boring but important database features and iPhone Duo support to work on, so I'm not sure when I'll find the time.