It's a cool concept but the first naive thought that comes to mind for me is that white could just easily dominate by taking advantage of the fact that taking a queen is almost always the top engine move. So just by maintaining relative king safety on white's side, you just open your queen early and make sure that every move there onwards hangs your queen in some fashion. It's very easy to hang a queen. You just have an invincible juggernaut for the first half of the game until you've demolished enough pieces to make it hard to find ways to hang your queen, and by that point the material advantage is such that the opponent might just resign.
The title "Human Chess is a chess variant where playing the top engine move is forbidden" kinda suggests that the top move just isn't available to the player.
However, the thought you had, and similar ones, are very much the intentional side-effect of the rules. The only way to win the game, as stated, is forcing your opponent to make the top engine move. Or, of course, correctly claim that your opponent made such a move (even though it wasn't forced). Or, having your opponent make the incorrect claim about your move.
So, it isn't necessarily "playing good chess". Though, I must say, I'm not qualified to have any good idea of what it would mean to be good at this game. It definitely helps to be good at chess, and have a good command of what are the correct engine moves. Especially since you lose if you incorrectly claim a position and opponent move was "the top engine line".
I suppose most would reduce this to leaving the opponent to only one legal move. In which case, the problem is is trivial. But, after move 2? You need to know most opening lines, and probably play intentionally bad in many situations.
Hm, this is cooler the more I think about it.
Imagine intentionally setting up material sacrifice with a resulting choice of multiple moves for the opponent to capitalize. If you can correctly evaluate the best computer move, you have a strong advantage. If it is not obvious, then the opponent might not dare to gamble the challenge.
Has Hikaru tried this in one of his streams? I'm certain he would have a blast.
The only thing about this that doesn't "spark joy" are the ambiguous practical implementations.
- Which engine? This matters a lot.
- How do you define the computational cut-off? CPU-minutes? Move depth? Etc. Not necessarily a simple problem.
- The rule "When multiple moves have the top score, they are all top moves". Needs a specific score evaluation delta for grouping "top moves".
All of these could rather simply be resolved if it isn't very important... might even add some uncertainty to it, for fun. Like, say: 1. Stockfish 15. 2. Allow the computer whatever resources it has available, 1 minute, and play some drum roll sample. 3. Pawn-evlauation of 0.05.
I'd also be interested in a variant where the engine-recommended move is displayed to the players. It would take out the uncertainty of whether a move is illegal, but would allow for different win conditions (e.g. approach checkmate without ever using an optimal move).
Not wasting time with claims would be another very practical advantage.
A further variant for fairness purposes: let each player bring their preferred chess engine, instead of arguing about the choice of only one; and have each player run both engines for mutual anti-cheating verification.
Then either the two engines agree on the best move (likely case if they are both strong) or all moves that either engine considers better than the other engine's best move (at least 2, usually not too many) can be interdicted.
This is what I thought it would be from the title. Top players say that they don’t necessarily see a top engine move but can immediately identify one when it is played on them.
It’s an exaggeration but with some truth to it. Outside of openings and end games the top move for a chess engine is often different from the top move for a player. That said, grandmasters do of course often play the top engine move.
Human players are dealing with both human limitations and human limitations which really changes the game. So a grandmaster can for example benefit from playing a slightly weaker but less well known opening that they have recently studied in depth with the assumption that their opponent hasn’t done the same.
Actually no, the top choice would be to force your opponent into a position where the series of second best moves literally destroys them.
You're looking for traps where only a particular move can save you.
A minimally modified engine lookup wins here.
There's a whole bunch of openings that ensure it for white, this game is rigged even more than playing the best move, even if you do enforce a random opening.
I wonder — at the top end the chess engines clearly compete with each other and produce different results, that’s how one can be said to be better than the other, right? But against us puny humans, especially novices, do they produce very different outputs? Or is it just like, the moves to crush a silly meat-brain are just super obvious, no need for creativity.
Especially in this game, the humans will be trying to play badly.
Some games from the most recent chess engine tournament look whacky as hell. Like if you showed the games to a grandmaster they would probably estimate it was 2 completely new 400 elo players.
I think some of the moves the engine suggests in human games look whacky to human grandmasters.
Hikaru Nakamura is a top GM that does commentary streams — and frequently comments things like “but what do I know about chess?” when responding to AI suggestions. It’ll suggest a weird move that seems to leave a mess on the board while insisting that everything is great. And obviously if you actually tried to play against Stockfish from that position, it would crush you.
Looks like 400 ELO, but hey — what do we know about chess?
One of the big issues with engines is that they don't assume they're playing against human players. Of course in an engine v engine tournament they're not. But it makes it hard to use them to study things, especially below elo 2000. Many times the chess.com "best move" seems utterly ridiculous, and when I click for it to follow through on the suggested move line, it shows the opponent making a response no human would ever ever make. I think we need to have some sort of division between "this engine will win any game ever" and "this engine will always win against a human."
There's been times when the chess.com calls my move a mistake, and suggests something nonsense that leads to a guaranteed mate in 15, when I got a (non-guaranteed but real) mate in 5 from my move, because I know I'm playing against a human. The engine move is more guaranteed, but very illogical unless your brain can do the equivalent of the 30 turns of minmax.
I had a similar naive thought, but it doesnt resolve easily. For starters, Black can do the same thing just a move behind.
Second, if you ever hang your queen two ways at once - one of them could be a less optimal take (-5 is not as good as -8)
Third, whoever is a move ahead in a race of taking pieces will be the first to run out of weak pieces to take. Their available move pool is shrinking faster. Not sure how it would play out, black would need to cater to it by removing defenders and hanging pieces of their own, etc.
That said, first move advantage does seem strong still due to how forcing a queen can be. An example would be 1.e4..e5 2.Qh5..d5 3.Qxf7#
The best way to find out the best strategy for Human Chess is to train AlphaZero to play it, and learn from its example. Then we can make a Human Human Chess variant where you lose the game by playing the top move suggested by this newly trained engine.
It is an interesting theoretical question whether we can have Aleph Zero Human Chess where Human(Human(...(Chess))) is applied infinitely, approaching Aleph Zero trainings of AlphaZero, or we get a redundant variant after some application where further application of Human() no longer produces a new variant.
The answer is no: every time you apply Human(), you reduce by one the number of legal moves at every given game state. After some finite number of iterations, there are no longer any legal opening moves.
With this variant, since there are a finite number of legal board positions in chess, and Human^n chess is simply chess with a finite number of moves prohibited at each board position (claiming is the same as resigning, in this formulation), there are a finite number of distinct Human^n chess, so the mapping Human^n-1 chess -> Human^n chess must eventually reach a cycle (potentially of length 1, i.e. a fixed point).
I’m not sure about a straight-up cycle. I think it requires determinism, whereas the engine is not necessarily deterministic (eg. NN based ones like AlphaZero).
I think it’s possible that at least 1 position (and probably a lot more) will have more than 1 optimal solution. If the engine is not fully deterministic, then it’s possible instead of a normal cycle, there are a fixed set of strategies at each n that form a cycle, but no single sequence that repeats.
The other thing that would compound this is that no current chess engine solves the game fully. There would be even more positions that have multiple “optimal” solutions if the engine only looks ahead to bounded x.
That's a good point about non-determinism, although I wonder if there are known convergence/stability results in the ML literature that allow you to effectively ignore that detail (i.e. it seems plausible that you could get something like "human^n chess always stabilizes at some fixed amount of training time/computational power"). You can also just fix the randomization seed, but that's obviously a less satisfying result.
> The other thing that would compound this is that no current chess engine solves the game fully. There would be even more positions that have multiple “optimal” solutions if the engine only looks ahead to bounded x.
I'm not sure this is an obstacle; we're explicitly excluding the computer's preferred move, rather than the necessarily optimal move, after all. You could easily play human chess with the engine from (e.g.) Battle Chess, which honestly is sort of an interesting idea in its own right.
There will be opening theory soon, and it will be essential in many cases.
For instance after 1.e3 e6 2.Qh5, White threatens Qxf7+ which would force black to play the top engine move. Then 2...g6 3.Qxg6 is one idea -- but there are two recaptures, fxg6 and hxg6, and only one of them can be the top engine move (hxg6, I'm guessing). So 3...fxg6 probably refutes this idea. But are you sure enough as white to try to claim a win if black goes hxg6?
And after say 1.e3 e6 2.Qh5 g6 3.Qxh7 (avoiding that line and going for material), not only does black not have to care about their rook (white can't take it, it would be the best move), black actually has 3...Qh4 winning -- he threatens 4...Qxf2+, white can't play 4.Qxh4 as that's the best engine move, and white's queen is threatened twice, so black will be able to take it -- provided he checked this line before the game to know which piece to take with.
Edit: it doesn't actually win, white has 4.Qf5 to defend f2... what a strange game.
Edit 2: once a piece is _en prise_ somewhere, the game can otherwise become somewhat normal as taking it would be the best move and so would moving it to a safe spot - so other moves can be played as usual. But would they be good?
The article states that if there are multiple moves the engine recommends, that they all count as the "optimal" move, even if there's an indication of preference by the engine.
I'm approaching 40 years old, have played chess off and on since I was a kid, and I'm not sure I've ever played a match that didn't include several blunders. Like, on the off chance I'm playing someone who doesn't blunder often, I'll certainly pick up their slack.
My game quality is measured in how many times I say "fuck!" right after moving a piece. A very good game for me is about a two-fuck game.
The fun thing about chess, you don’t seem to think you are good, but I can’t imagine only
making two obvious mistakes in a game! It grows with you, haha.
Most of them it's more like four or five "oh my god I hope they don't see that thing I spotted the second I took my hand off" moments—and that's just the ones I notice before they're exploited. I'm sure I make tons of moves that anyone half-decent would call blunders but that simply go unnoticed by both players at the board.
I'm so very bad at spotting diagonal attacks, especially. Anyone who can open up their bishops then play for time will eventually see me put my queen in some dumbshit situation that lets them take it free or cheap in a single move, for instance, not even any multi-move planning required.
Lichess has a variant called "Antichess". If you can take a piece, you have to. No checks/checkmate rules. First person to have zero remaining pieces wins.
You basically want to "blunder" into giving your opponent long chains of captures while avoiding any positions that allow your opponent to hang a piece.
That was my thought too. Does the engine know that the other player is forbidden to play the top engine move on the next turn? Then you can't just do something like c3 Nf6 Qa4 e5 Qe4!? to hang the queen in the center, knowing that Nxe4 is prohibited, because the strongest move for white if Nxe4 is prohibited would have been Qe4!
Any move the human chess engine makes would be losing by definition if it is also used as the bench mark.
Therefor, such an engine can only hang in computation - being unable to produce a top move because if it were to make a suggestion then the actual best move changes to avoid it. Since the engine is unable to produce a move - there is no top engine move which makes every move legal.
A normal game of chess is played while the engine locks up on the sideline
You forgot that the engine user may cheat and even if they declare the engine there's no real way to detect if they're truthful and not using a certain of the engine tuned for Human Chess specifically.
My first naive counterthought is that if you try to do that then black can ignore your queen and use their queen to take your pieces. You're a move ahead but you don't get a snowballing advantage.
I bet if you hang your queen 5-6 times, for one of those, taking the queen won't be the top move. Just think about if your queen would _still_ be hanging after a check that maybe captures some other piece first. And then that strategy has lost you your queen.
I like this strategy, but I don't think it's necessarily clear cut: while taking the queen is forbidden, the opponent also has the opportunity of putting their queen en prise.
So you end up in this scenario where both players are taking one another's pieces while leaving their queens en prise the whole time. Is it a draw, or is there some clever way to break this loop?
giving checkmate is a loss, the win condition for human chess is giving check in such a way that there is only one legal move. So hanging your queen might be good offensively, but its not good defense.
Moreover if you hang your queen in more than one way, your opponent can still take it in whichever way the computer evaluates as worse. Which is often easy to guess. The weird part of this will come from the fact that accuracy of engines diverges very rapidly off of the critical path. Once you're down a queen, you're basically free to play however you like.
You could have two engines: a strong engine and a weak engine. If the weak engine suggests a move then it's allowed. But if the strong engines suggests a different move than it's forbidden.
But you'd need to fine-tune the stupidity of the weak engine to have decent but not too good moves: an extra chore that would amply offset any increase of fun.
There must be a threshold above which making an engine move is allowed. Because if there's a checkmate move, it will be the best engine move and needs to be allowed.
No; checkmates are not needed. You win the game by almost checkmating the opponent: by leaving them only one legal move, which is thus necessarily the top engine move.
As others have mentioned, that is accounted for in the rules.
My immediate first reaction was also that it would be interesting to have a variant that is the same except you are allowed to checkmate, except then I realized the recursive nature of how board positions are evaluated makes that problematic. For instance, if there's a mate in 2, the first move of the mate in two is now certainly the "best move". Creeping up on a checkmate without ever making the "best move" until the very last one might actually be harder than the win condition based on strangling the opponent described in the current rules.
It just doesn't make sense though for that reason. You can't sneak up on an engine. There's not a single engine out there that wouldn't recognize a mate in 2 moves. Unless the opponent blunders (which actually might be forced if the best defensive move is blocked).
It just seems like you're changing the objective of the game entirely to the point where it's only slightly related to chess.
The article had an illustration of this: the player put their opponent into check where there was only one move to get out of check. The opponent would have to make that move, whether they were human or machine, so the player who made the original move wins the game.
It is an interesting variation on chess given the current state of tournament play, yet it isn't really a solution to the cheating problem since it is effectively a new game with a new end-state. But you are probably right about there needing to be some sort of threshold. While there the rules of the variation says that any move with equal scores is considered equivalent, I would imagine the players would need a very intimate knowledge of how the engine scores moves in certain scenarios.
If you deliver checkmate, you lose. You have to instead force your opponent into a situation where they only have one legal move (which is trivially the best engine move).
You can force the opponent into a position where the only move that saves them is a top engine move. Since they cannot play that move, the other option is to surrender.
So essentially this converts most mates in two into mates in one, but some become ties by repetition.
I don’t see how one way of checkmating could be worse than another way of checkmating? Do some engines give different scores for different check-mating moves? (Different moves from the same position I mean)
