Chess variants made with help from alpha zero article

[jmartus]

WIRED: AI Ruined Chess. Now, It's Making the Game Beautiful Again.
https://www.wired.com/story/ai-ruined-c ... -beautiful

No castling or self capture chess sounds the most interesting from the article.

[mmt]

Assessing Game Balance with AlphaZero: Exploring Alternative Rule Sets in Chess
https://arxiv.org/abs/2009.04374

I wish they would evaluate a no-black castling variant or other variants discussed here http://talkchess.com/forum3/viewtopic.php?t=71367 instead of the boring no-stalemate variant with a completely predictable result. It looks like they looked at small modifications (no stalemate or no castling in the first 10 moves) and very large modifications that make it a completely different game (self-capture). No-black castling would be a good middle ground with its complete elimination of draws. It would be a better paper if the authors started a topic here or on some other forum before choosing which variants to evaluate.

[Alayan]

Stalemate=win is by far the variant that's the most similar to classical chess out of all these tested in this paper.

The assessment of the Stalemate=win chess variant, as pro-vided by Vladimir Kramnik:

I was at first somewhat surprised that the decisivegame percentage in this variation was roughlyequal to that of classical chess, with similar lev-els of performance for White and Black. I waspersonally expecting the change to lead to moredecisive games and a higher winning percentagefor White.

Roughly equal ?????

13,94% decisive games vs 11,80% decisive games at 1s/move ; and 2,9% vs 2,1% at 1m/move, that's a very substantial increase. Chess is a drawish game and for a given strength NN self-play is extremely drawish.

In a tournament like TCEC, 25% decisive games instead of 20% would be a massive difference.

[mmt]

13,94% decisive games vs 11,80% decisive games at 1s/move ; and 2,9% vs 2,1% at 1m/move, that's a very substantial increase. Chess is a drawish game and for a given strength NN self-play is extremely drawish.

In a tournament like TCEC, 25% decisive games instead of 20% would be a massive difference.

I don't think it's enough of an increase to make it a variant that people or programmers would switch to unless there are other reasons. I think changing the rules to completely eliminate draws would be needed. If you can keep using 90% of your chess knowledge like with no-black castling variant it's also a big plus if you want current players to try it out. I hope chess sites will offer some variants - it seems easy to implement.

[Raphexon]

Stalemate=win is by far the variant that's the most similar to classical chess out of all these tested in this paper.

The assessment of the Stalemate=win chess variant, as pro-vided by Vladimir Kramnik:

I was at first somewhat surprised that the decisivegame percentage in this variation was roughlyequal to that of classical chess, with similar lev-els of performance for White and Black. I waspersonally expecting the change to lead to moredecisive games and a higher winning percentagefor White.

Roughly equal ?????

13,94% decisive games vs 11,80% decisive games at 1s/move ; and 2,9% vs 2,1% at 1m/move, that's a very substantial increase. Chess is a drawish game and for a given strength NN self-play is extremely drawish.

In a tournament like TCEC, 25% decisive games instead of 20% would be a massive difference.

Yes, 38% more wins at 1m/move is a nice start.

Anyway, big board chess would do the most to eliminate drawrate.
Big board also has the advantage of lessening first player advantage, while the variants studied with help of A0 exemplified the first player advantage.
I don't feel there's a point to a variant where white has a good chance of winning but black has to fight tooth and nail for a draw.

Kaufman has written about different variants in the late 90s.

http://www.shogi.net/shogi-l/Archive/1999/Nfeb07-06.txt

I personally think something between Grand Chess and Chu-Shogi is the sweet spot.
Grand Chess is enough like chess that even a beginner should be able to follow it.
It doesn't suffer from a significant opening advantage* and the depth it would add would make it safe from drawdeath for at least a decade I bet.

Chu-Shogi with westernized pieces is sort of followable, but it is really at the far end of (chess) complexity that can be enjoyed by a human.
By now it has no active playerbase, but it was a historically popular variants that had a decent following well until the 80s.
http://history.chess.free.fr/images/sho ... gi-occ.jpg

Unlike Kaufman I don't like Shogi too much. In my opinion the drop-in rules stray it way too far from regular chess. (Which may be the point for others)
For a spectator who isn't a master in Shogi the tactics (by strong engines) are also going to be way too complex to be very enjoyable.
Hard (if not impossible) to keep track of what's going on.
Go has the same problem, you need to be very good to be able to follow a grandmaster. Let alone what an engine is capable of.


*In my test with Fairy-Stockfish I actually got the (almost paradoxical) result of black winning more than white.
White vs Black: +596 -843 =561 [0.438] 2000
I'll do a longer test with more threads in the future to see what happens with increased playing strength.

[mmt]

Yes, 38% more wins at 1m/move is a nice start.

Or looking at this from another perspective, you've reduced the number of draws by less than 1%.

[Andrew]

Seeing the various options with pawn moves altered reminded me of a variant I played years ago back in school, which I'd seen in some book covering variants. In it normal pawn moves were diagonal and captures directly ahead. No idea what en passant rules were! Does anyone remember this?

I really hope the alpha zero people can release to the public some of the networks (and possibly playing programs) for people to tinker with these variants on their own. Would also be nice for them or someone to develop a program/utility where you can define the rules yourself, press go and train a network! Some of the variants for 6x6 chess and other small board sizes would be especially interesting as an individual with a modest system could potentially train a strong network quickly!

It would also be interesting to see them train networks for dealing with KP endgames and KPR endgames (which also hopefully the Stockfish team will consider). No idea what this would gain in ELO, but for analysis it could be a big thing.

Andrew

[Raphexon]

Seeing the various options with pawn moves altered reminded me of a variant I played years ago back in school, which I'd seen in some book covering variants. In it normal pawn moves were diagonal and captures directly ahead. No idea what en passant rules were! Does anyone remember this?

I really hope the alpha zero people can release to the public some of the networks (and possibly playing programs) for people to tinker with these variants on their own. Would also be nice for them or someone to develop a program/utility where you can define the rules yourself, press go and train a network! Some of the variants for 6x6 chess and other small board sizes would be especially interesting as an individual with a modest system could potentially train a strong network quickly!

It would also be interesting to see them train networks for dealing with KP endgames and KPR endgames (which also hopefully the Stockfish team will consider). No idea what this would gain in ELO, but for analysis it could be a big thing.

Andrew

Yes, that would make it really easy to prototype and test variants.
Something like MuZero may even be nicer for it.

Is this the book you're thinking of?



There's also



But may be a little too old..

[Andrew]

Seeing the various options with pawn moves altered reminded me of a variant I played years ago back in school, which I'd seen in some book covering variants. In it normal pawn moves were diagonal and captures directly ahead. No idea what en passant rules were! Does anyone remember this?

I really hope the alpha zero people can release to the public some of the networks (and possibly playing programs) for people to tinker with these variants on their own. Would also be nice for them or someone to develop a program/utility where you can define the rules yourself, press go and train a network! Some of the variants for 6x6 chess and other small board sizes would be especially interesting as an individual with a modest system could potentially train a strong network quickly!

It would also be interesting to see them train networks for dealing with KP endgames and KPR endgames (which also hopefully the Stockfish team will consider). No idea what this would gain in ELO, but for analysis it could be a big thing.

Andrew

Yes, that would make it really easy to prototype and test variants.
Something like MuZero may even be nicer for it.

Is this the book you're thinking of?



There's also



But may be a little too old..

Thanks, I think it was either the David Pritchard book, or possibly the "The Oxford Companion to Chess" book

Andrew

[hgm]

The alternative Pawns are known as Berolina Pawns. The e.p. rules are similar to those of orthodox Chess: Pawns have a double push by which they move to the second square on a forward diagonal, and an enemy Pawn that attacked the square the moved through with its capture move can then move to that square, making the doubly pushed Pawn disappear. What is remarkable here is that an enemy Pawn might also be able to make a normal non-capture move to the skipped square. In which case nothing special happens. And e.p. rights cannot be unambiguously encoded by a single square; you need to know both the e.p. square and the Pawn that moved through it.