[jessriedel]

Now that computers can be cheaply trained to play chess at greater than human skill level without any pre-existing knowledge of the game, has anyone looked at the question of whether the rules of chess are optimal or non-optimal under some measure of interestingness and simplicity? Like, there was some evolutionary pressure on the rules of chess to be simple and yet yield interesting games in the early days before the rules had solidified. It's now possible to take an arbitrary chess rules variant and cheaply train a machine to be (with high confidence) at least as good as a grand master would be in the universe where those rules had persisted for a thousand years. Are there better rules out there?

[chongli]

There was a story on here a few weeks ago [1] where retired grand master Vladimir Kramnik advocated for getting rid of castling. He worked with the Alpha Zero team to investigate the ramifications. They re-trained the network from scratch on the rule change and Kramnik claims it’ll lead to way more interesting games at the top level. It remains to be seen whether this plays out in practice, given that humans are more risk-averse (particularly in sharp positions) than Alpha Zero.

[1] https://news.ycombinator.com/item?id=21789224

[elfexec]

Not sure if this is what you are asking, but Bobby Fischer thought the current chess rules led to boring games so he invented Fischer Random Chess.

https://en.wikipedia.org/wiki/Fischer_random_chess

It makes the games far more complex and dynamic since the pieces are placed somewhat randomly. There are other variants which remove castling which makes the game even more dynamic as it prevent castling to secure your king from the middle of the rank.

[galaxyLogic]

How about this simple rule, has this been tried?

To begin with each player puts one piece of their color on the board on their turn where-ever they want.

When all pieces are on table, the player who placed the last piece gets the first move, to compensate for the fact that the other player had the advantage to place the first piece on the board where-ever they wanted to.

[jeffdavis]

It seems like you'd end up with a lot of pawns placed near the end of the board, quickly (instantly?) turning into queens.

[torgian]

Then add a rule where pawns are limited to the middle rows of the board, amd/or reduce the amount of pawns

[galaxyLogic]

That would seem to work.

Another approach could be that yes you must fill the first two rows of your side initially, but players would take turns in doing that piece by piece. You would see how your opponent places their army piece by piece on their side and you would adjust your placements accordingly.

[galaxyLogic]

I see. So I guess it wouldn't make much sense, without adding more new rules.

[jessriedel]

Thanks, I'm aware of many chess rule variants. The question is whether these have ever been studied systematically by training human-level computers on each rule set and then analyzing the games played.

[otabdeveloper4]

> cheaply

Citation needed. AlphaZero was 'cheap' only if you have Google-scale resources at hand.

[jessriedel]

AlphaZero beat Stockfish, which is substantially more powerful than humans. Much fewer resources would be needed for human level performance.

[lazyant]

I think moving the pawn two squares forward is one such rule innovation (as in the movement was not there in old time chess, also because of this en passant had to be implemented as well)

[emmelaich]

There is chess960 among others.

https://en.wikipedia.org/wiki/Fischer_random_chess