Comments/Ratings for a Single Item

Retrofitting an existing high-quality engine to play variants is problematic.  It is not so bad for variants that are very close, like Fischer Random.  Even things like Siriwan that are played on 8 x 8 might not be tooo bad.  But adding the ability to change board size to a program which was created without that in mind would be very difficult.  And it probably would not be worth the time.  The really top-level engines like Stockfish are so good because they have lots and lots of code that very specifically designed and optimized to play Chess.  Most of this code is, at best, throw-away for playing variants, and at worst, downright harmful and hard to remove.

Best to create from scratch.  Want to dip your toes into these waters, I would recommend you download the source code for SjaakII, and see if you can improve it, or at least understand it.

Cheers :)

I think you are a bit too pessimistic on this. There actually exists a version of Stockfish that can play variants, some very different from Chess (Crazyhouse). And it blows away everything that was written before. So the general search strategy must count for a lot. And the Stockfish developmet system also includes advanced tuning methods.

HG & Greg both,

That was mostly about Alpha0 rather than Stockfish, which is a machine learning program :)! Totally diferent monster. They changed a program desinged to play go into a program that play chess and trained it for only 2 hours. I think it is obvious that the diffenece between Go and Chess are far greater than let's say between chess and Grand Chess or Omega chess. This is what I was asking about. I hope I have made myself a bit clearer :)!

It's certainly possible I'm too pessimistic.  Playing Crazyhouse is pretty impressive as Stockfish has no concept of drops, and that's a huge change.  That said, it's still an 8x8 game.  Stockfish has bitboards deeply embedded throughout.  That said, it is very object-oriented and template-driven, so replacing the bitboard class with one that uses more bits might not be too bad.  But I still think it actually would be harder to make it play Omega Chess than Crazyhouse.  Anyway, Stockfish's selectivity is incredible.  It makes sense to me to just design an engine with universality in mind and then copy the fantastic elements of Stockfish's search algorithm.

Regarding AlphaChess, apparently this is a completely different beast.  I should have read the article before commenting :)  I might have some thoughts after I read it.

AlphaZero is a neural net which learns by playing against itself, starting with random moves and working up to a rating of over 3300, iirc. It runs on some very fancy hardware, so  its learning time is misleading. It played millions of games against itself to learn Go. I'm curious about just what it learns when it's teaching itself a game. How dependent is it on the exact board geometry and each player moving only one (or a few) pieces per turn? If the game is a very large (~1000 - 10,000 squares or more) massively multi-move abstract strategy war game with a board that can change between games, does that sort of thing make any significant difference, or merely add some time to the AI learning process?