Chess engines are optimizing for the wrong thing!
Chess is a game. In order for something to be called a game, it must be fun, it must be tailored to the level of the players and sometimes, especially nowadays, it needs to be exciting to an audience.
Now, chess engines are fantastic in respecting the rules of chess and mating the king in the quickest possible way, but it's not a game anymore, it's a process. Occasionally people watch what computer engines are doing and notice the beauty in some of the ideas, but that beauty is coincidental, it has no value to the machine and "sparks no joy".
I've been advocating for a while training chess engines on other values, like beauty or excitement, but those are hard to quantify. So here is a list of values that I thought would be great to train chess engines on:
- player rating
- which is great because it's constrained in time, so if someone is a GM, but completely drunk and haven't been sleeping for a week, the engine would adapt for their play at that time
- I know that engines have a manual level configuration, but I doubt it was ever correctly modelled as an input. Most of the time, a random move is chosen from the list of best moves, which is not what I am suggesting here at all
- value and risk of a move
- I know this sounds like what engines are doing now, but they are actually minimizing risk, not maximizing value
- We also have the player rating to take into account now, so the calculation changes with the player! A move that would be negative because another perfect computer chess engine would take advantage of a minute flaw means nothing now, because there is no way an 1800 rated human will see it. And if they do, what a boost in confidence when they win and what pleasure in witnessing the moment!
- balance risk with the probability of winning
- this is the best part. Riskier moves are more fun, but can cause one to lose. Allow a probability of the other player missing a move, based on what we have calculated above. We are actually adding a value of disrespect from the engine. It attempts to win despite the moves it makes, not because of them.
What I am modelling here is not a computer that plays perfect chess, but a chess streamer. They gambit, they try weird stuff, they do moves that look good because they can think of what the other player is or the audience are going to feel. They are min-maxing entertainment!
A chess streamer is usually a guy around 2500 showing mercy and teaching when playing against lower rated players and trying entertaining strategies against equal or even better rated ones. They rate the level of a move, which is an essential metric on what strategies they are going to employ and what moves they are going to play. In other words, they are never considering a move without taking context into account.
Imagine a normal chess engine, using min-max or neural networks to determine how to win the game. Against another computer, the valuation function is extremely important, since it limits the number of possible moves to one or two. Against a human noob, there are a lot of moves that will lead to a win. It is obvious that another metric is necessary to filter them out. That's how humans play!
Short story shorter: use opponent rating to broaden the list of winning candidate moves, then filter them with a second metric that maximizes entertainment value.
Be the first to post a comment