Hacker News Reader

Comment by dannyz | Hacker News Reader

While computer engines that are significantly better than the best chess players are plentiful, computer engines that are human rated and play human like moves do not exist. From the definition of ELO you could make a 1500 rated chess engine that plays perfectly half of its games and plays random moves the other half of the games, and it would stay at 1500 ELO if it continued to play against 1500 ELO players. This is an extreme example but it is not too far off from what chess engines do to reduce their ELO.

View on HN

> computer engines that are human rated and play human like moves do not exist

I submit that you're wrong there: https://maiachess.com/

I have played a few hundred games against the various levels of Maia (https://lichess.org/@/maia1), and it has helped me improve from 1100 to 1425 blitz rating on lichess (along with playing against stronger players in the arena tournaments). It does seem to encapsulate the “average” playing style and common mistakes of the different ratings. It feels much fairer, more instructive, and more relaxing than playing against Stockfish, who will wildly blunder and then subsequently torture you.

However, even as a weak player, Maia is exploitable in a way that even novice humans are not. For example, it loves to give up obvious back-rank mates and almost never protects against it (e.g. with h3). It is easy for humans to spot if they are in danger of back-rank mate (especially if you’ve been burned before) and defend against it, so it does not happen much even at low levels. But Maia, game after game, gifts you an obvious mate in 1.

> computer engines that are human rated and play human like moves do not exist

Without knowing much about chess, I'd be inclined to say this can't possibly be true. Predicting human chess moves given an ELO rating sounds like a much easier task than training it to be good at the game. It is a different task alltogether, but until you reach relatively high elo scores, I'd expect a fake human ai would have a much easier time simply because it only needs to plan a few moves ahead to mimic the mindset of a lower level player.

The problem is that humans don't play that way, by evaluating a tree a few moves ahead and selecting from that. Humans use conscious heuristics and unconscious intuitions to make their moves that would be hard for a computer to mimic. It's easy to make a machine complete a task better than a human would, much harder to make it convincingly human-like in its behavior. Consider the task of picking apples from a tree. You could easily make a machine to do that. But what if you had to make a machine that to an onlooker (who couldn't see the robot's "face", let's say) would appear to be a human picking apples from a tree? And these people aren't just glancing at the apple-picking robot, they're spending their entire lives painstakingly analyzing the movements of this machine and trying their hardest to predict how they will move next. And the people doing this are self-selected to be the best performers in the world at predicting the moves of this apple-picking robot. Think you can make the apple-picking robot that will fool these people??

There are many chess AIs on chess.com specifically designed to play "like" a specific grandmaster or well known chess streamer. I don't think any titled player would not be able to guess they're playing a computer if they played a few games against the AI without being told. It's very well known that computer moves are very different from human moves, even the ones specifically designed to represent a human.

I think the proposal is to train neural networks on games played by human players at a given ELO.

Even if it were possible, I don't think there's a way to do it in a way that would make any sense. Would we set up an AI now and then forever use the same one? In that case people will learn any small quirks of the AI and optimize against those vs human play, fundamentally changing the game. If the proposal is to be constantly updating this AI, well even in that case I'd argue that people will over time identify any small difference between machine and human play, but in that case people will also be upset that someone got normed against the "easy" AI, or against the AI that didn't know about the dumdenmorph-joyce-allens countergambit yet, etc.

That kind of engine exists! It's called Maia, as mentioned in a sibling comment.

https://maiachess.com/

> The problem is that humans don't play that way, by evaluating a tree a few moves ahead and selecting from that.

As a modest 1400 blitz player, I definitely evaluate future states. That's not all I do, but it's certainly one thing I do. AlphaZero can also run just intuiting the next move without evaluating any additional states also fwiw. Though it is much better when it is allowed to do so.

> Consider the task of picking apples from a tree. You could easily make a machine to do that. But what if you had to make a machine that to an onlooker (who couldn't see the robot's "face", let's say) would appear to be a human picking apples from a tree? That would be much harder.

This is a very deep and nuanced task, made doubly difficult by obscure robotic hardware requirements. Not a great parallel.

> There are many chess AIs on chess.com, specifically designed to play "like" a specific grandmaster or well known chess streamer. I don't think any titled player would not be able to guess they're playing a computer if they played a few games against the AI without being told. It's very well known that computer moves are very different from human moves.

Maybe? A quick google suggested Maia is a close match to what I was suggesting. People are suggesting it does feel like a human in the thread. https://www.reddit.com/r/chess/comments/k4o6z1/introducing_m...

I don’t know that human calculation is so fundamentally different than computer evaluation, it’s just that it’s much, much slower. The human GM’s eval of any given position, without thinking ahead at all, is probably better than most engines. The problem is that strong evaluation isn’t worth much compared to a weaker evaluation that is nevertheless several candidate moves wider and many plies deeper (you might argue AlphaZero disproves this, but Stockfish was regularly beating the neural network engines even before its own neural network, for example).

Humans see fewer candidate moves. They regularly miss quiet moves that are the strongest. They often calculate to a certain depth (even a very shallow one) and stop based on gut instinct. But it’s still just trees of moves with some eval function.

I don’t think it’s fundamentally that hard to mimic, and it would actually be genuinely interesting for didactic purposes. But for fairly obvious reasons it’s not a priority outside of a couple of projects.

I haven’t followed the new wave of DL based engines but in the old Alpha-Beta algorithm days engines could be tuned by giving them a fixed processing budget per move (or a more sophisticated processing budget to allow for bursting).

Is that not a viable strategy on the new engines?

I think it's really hard to make it play like a human. Even GMs miss two-move tactics sometimes, wheras computers are pretty perfect at this.

On the other hand traditional algorithms might not be very good at positional chess when on a budget, wheras even weak human players can understand basic concepts like putting your rook on an open file. Then it gets even more complicated with the NNUE position evaluation models - maybe even with a one-move budget the neural net is still implicitly strong at understanding several moves ahead, and could still beat a strong human player.

Yeah exactly - it's easier to build an engine that plays as well as possible compared to an engine that plays reasonable/imperfect moves like a strong human player.

I believe this has not been true for some months now.

Reply on news.ycombinator.com