Fuseki.net

Chess Rules Irregularities

What are the consequences of smoothing out some of the irregularities of the rules of chess?

Metrics for each change

Each change will be evaluated on a few axes:

Future

To actually get real answers would require setting up a training run for an AlphaZero-like engine.

Explanation of strategic depth:

What does the rating structure look like? i.e. tic-tac-toe doesn't have a very deep skill level tree, but chess does. ELO is in some sense an absolute scale; if you define random play as 1000, the higher the best known player's ELO is in this world, the deeper the game. re: this amazing youtube exploration of weird chess algorithms: https://www.youtube.com/watch?v=DpXy041BIlA

It's not clear what the result of this is for chess since ELO 1000's definition is not clear, and seems to be linked to the default skill of someone who only knows the rules. This is already a pre-selected group though, so the real depth of standard chess isn't known.

Another way to think about this: how many skill levels can a game have, where level 1 is random play, and level N+1 beats level N X% of the time, X>50%). Question: as X varies is it all proportional? Or does choice of X matter? Default to X=75. The existence of draws complicates this, too. Perhaps a better system would be that level X+1 on average receives N points per game against level X, 1>N>0.5. This system is game-able to artificially introduce depth (at the worst, various AIs could communicate their level to each other through hard-coded move patterns, and then just let whoever had the highest level win. But there may still be some value in this as long as the engines aren't trying to artificially introduce levels.

Main categories of rule variations

I will not explore variations involving more complexity to the rules, new pieces, etc. This comes out of realizing that for the game of go there are almost no meaningful ways to simplify the best rule-sets; this process is already done there.

Conceptual changes

What is king nature?

Kings have something special - if they die the army dies. But what if you changed king-nature and gave it to someone else?

Multiple King nature

Win by complete army destruction, not king capture - no King Nature

In this variation, the King is just another piece; you win when the other player has no pieces left. This changes stalemate conditions quite a bit. I'd expect a lot of opening changes too since there isn't as much emphasis on king safety. Standard chess is a subset of this, where the game just stops early. What would the king's value be here? According to wikipedia, as a piece its value is approximately 4 due to its strong local area of control. However, that's still based on checkmate rules; in this variation the King's value would be higher.

Self-capture

Allow capturing your own army

An interesting metric in general is the size of the gap between "board position value" and "material comparison". In chess for most positions by an engine, these two are relatively close. But the more dramatic positional advantages available, the greater the gap. It seems that if it is common to be winning, but behind in material, the game would be better.

Allow zero-length moves

Many chess pieces have "infinite" distance capability. If you put a bishop on a 20x20 board, everyone would assume that it could still move corner to corner. So, it can move any number of squares. But why can't it just move zero squares? Why out of all numbers is only zero excluded?

Knight movement irregularities

Do knights move 2 squares in one direction, then 1 perpendicular, or the other way around? In regular chess it doesn't matter since they fly over, rather than move through squares. But this flying power is irregular. There are three variations:

Remove Knight flying ability

Are there differences among these options? Because they can now be blocked, it matters. In say the 2,1 situation, it's also possible for my knight to attack my opponent's knight, but not be attacked, if there is a piece blocking it. This breaks the normal situation where B,N,R,K,Q are all guaranteed to be mutually attacking.

Promotion Changes

The rules around promotion add a lot of complexity to the game; you theoretically need 8 extra queens and every other promotable piece of each color on hand to be truly ready to play any possible game, for example. So most tournaments and live play situations are vulnerable to failing to actually be able to play a game.

There is also an issue with time around promotion - because it's rare and involves moving around off the board, in time-pressure situations it can be hard to resolve. Do you stop the clock while searching for the 3rd queen you just promoted?

This suggests a few simplifications:

Disallow promotion

Other forced promotion types

If the mythology of chess supported it, pawns may automatically have promoted to kings, or rooks, or other pieces, exclusively. Rooks wouldn't change things too much, but minor pieces would greatly increase draws and lessen the value of passed pawn defense. Some of these changes make having multiple bishops of the same color much more likely, allowing B-B batteries, which nearly never happens in standard chess.

In Chess's historical context, it seems more natural for pawns to want to become kings than queens. But what do multiple kings mean? Do they compete with the other king? Or are they supporters of his war against the opponent? Can they be independently checkmated? The most natural interpretation seems to be that you lose when all of your kings are captured / capturable. You'd have to allow actual king capture within a continuing game. This is similar to how a campaigning general is effectively the king of his army; but if he dies, the king back in the capital still has power. Loss only happens when all such kings are dead.

Option to promote to king
Forced promotion to queen
Forced promotion to rook
Promotion options are only R, N, or B
Promotion options: N or B
Force promotion to B
Force promotion to N

With bishop promotion, you'd see more the previously extremely rare bishop battery.

Legacy rules

En Passant

Remove en passant entirely

Some violation of legacy expectations. The en passant rule seems like the result of a ratcheting process - X% of people will argue for it, X will eventually get high enough to be in effect, and once in effect it can't be undone.

Castling

Castling involves a bunch of violations of the rules other pieces follow:

Remove Castling entirely

There have been tournaments run with this variation.

Make Castling symmetrical

Official castling is arguably as logically complex as another variant, where on the queenside the king moves 3 steps instead of 2. The rule is "king moves til it's next to the rook, rook jumps over it".

Redefine castling as "king jumps sideways until next to its rook in the corner; rook jumps over the king

One-step pawn moves

Remove special-case two-step pawn moves

Board Statefulness

Knowing what is legal sometimes requires knowing past moves. That is, metadata about board history is required to know whether certain moves are currently legal.

En passant

What happens if you remove the statefulness rule of enpassant?

En passant is no longer limited to immediately after the opponent's pawns 2-step move; as long as a pawn might have just moved en passant, it can be taken en passant.

Castling & Statefulness

There is lots of state here: has the king ever moved? Has the rook ever moved? Do we really gain much by imposing these restrictions? They seem mostly symbolic/historical. i.e. a king can run away easily if his attackers haven't come close to breaching his hideaways yet. Castling requirements which depend on board orientation will be discussed later.

Remove statefulness from castling. You can castle whenever your king and rook are in the appropriate positions.

Board State memory

3-fold repeats

This is hard to remove, since infinite repeats wouldn't be detectable.

“meaningful move history"

I.e. the "50 moves without a capture or pawn move" rule also seems hard to remove, since it opens up infinite repeats or continued play to cause your opponent to flag.

Board Orientation

Some rules depend on knowing the board orientation. There are two kinds of board symmetry we could remove knowledge of:

Remove board orientation requirements from castling

This means that you can castle on either end of the board. The concept is that the ability to castle is a feature of the board, like a tunnel - from specific starting squares, any pair of K & R can castle. The extension of this, allowing cross-color castling, is infeasible because both before and after castling, the rook always attacks the king.

The only requirement for castling is piece position

Pawn movement

Out of all pieces, pawns are the only ones where you need to know which way the board is facing to know how they're allowed to move. Legacy chess variants retain this property even for other pieces; i.e. shogi/xiangqi have pieces which require knowing board orientation.

So removing it would imply pawns could move and take both up and down, and could even promote on either end of the board.

Changing this would probably make a huge difference to the game, since you'd be able to move your knight out of the way and then immediately move a pawn backwards to promote.

An even greater change would be to remove the requirement to track which part of the board are the top/bottom vs which are the sides. This would allow pawns to move one square in any direction orthogonally, and attack any direction diagonally; and when they were on a square which could be considered a "2nd rank" they can move two squares towards the center. The problem with this is that the a and h pawns should have already promoted since they start in what could be considered a promotion zone. There are some somewhat consistent rules making promotion optional and available as an instantaneous move during your turn.

Remove the concept of vertical orientation of the board
Remove all board orientation-specific rules

Next steps

Actually test these

You'd have to write a fast board evaluation function for each rule variation, then train an alphazero like engine. Goal would be to have lots of sample games for each variation, as well as an average of (summed material => win percentages) from random states within those games to reveal piece variation. For openings, you could just compare common variations openings

Evaluation of the criteria per variation:

Related: