Making Game AI a little bit dumber

[Jubal]

I don't think this said much that felt groundbreaking to me, but it was an interesting little read:
https://www.newscientist.com/article/2078134-cleverness-isnt-everything-for-a-gaming-artificial-intelligence/

Thoughts?

[Glaurung]

It makes sense: if the AI can always beat a human player, then no-one will want to play against it. Modelling human fallibility is going to be interesting, though, and we might even learn something from it

[Clockwork]

For starcraft at least, I think it depends on whether when programming an AI they account for every human move or if the AI can learn on the fly what to expect. I frequently watch the tournaments and follow the pro scene and the pro players make an educated guess on what style of play their opponent is going for based on sometimes the first building built. It takes a combination of knowing what the player has done before and which matchups they've been practicing. An AI would have to have a library of past games for their human opponent and be able to make a judgement call on when a timing attack would be most effective or who to rush against and who to turtle up against.

[Lizard]

For starcraft at least, I think it depends on whether when programming an AI they account for every human move or if the AI can learn on the fly what to expect. I frequently watch the tournaments and follow the pro scene and the pro players make an educated guess on what style of play their opponent is going for based on sometimes the first building built. It takes a combination of knowing what the player has done before and which matchups they've been practicing. An AI would have to have a library of past games for their human opponent and be able to make a judgement call on when a timing attack would be most effective or who to rush against and who to turtle up against.

I'm actually working on something that might be able to do this. Not for Starcraft because that armadillo's complicated, at least not yet anyway.


But yeah, it's a well-known fact (at least amongst AI specialists and games developers) that AI for things like FPS games start off near-perfect and you have to make them stupider until you get to a point where they're beatable.


As for things that learn from other players... they're only as good as the people they've played. And the efficiency of the algorithm used to train them/perform the move calculation.

[GUNINANRUNIN]

A video game AI that is fun to play against is a good AI. AI War for example. The enemy is fun to play against because it has a strong element of predictability, so it isn't necessarily smart in the most conventional sense, but it is really good because you can stack behaviors on top of behaviors until you need to pause the game for 30 minutes just to come up with a plan that accounts for every reaction the AI will produce.

[Jubal]

I think when I've done AI programming I tended to be quite good at having fallible AI, but that may just because I was bad at it