Do you think that AI in competitive vidya will ever get better at simulating playing against a human opponent? Is that even something that should be strived for? I’ve been getting into fighting games and am really realizing how entirely different it is playing against the CPU and other players.
That’s tough. I think it depends on the game. For fighting games, I think it’s hard to say if it will ever reach that point.
Sethbling did an interesting proof of concept here with neural networks being trained to copy his style of play, with the fitness criteria being how closely it matched the actions he’d take in a given scenario, so you end up with an AI that plays very similarly to a human player.
Of course this is a lot easier because Mario Kart is what I call an “Efficiency Race”. It’s a game about efficiency, more than a game about Rock Paper Scissors. There’s a tiny bit of RPS in the use of items and passing opponents, but not much.
I doubt the ability of AI to play like a human in that they can really play RPS in a complex game with multiple RPS loops and we the same way a human would. Or to be able to learn and adapt to a player over the course of a few games. I feel more like it would just have an expert policy and do the best thing each time, rather than play with the mentality of a human player. Also it’s hard to say if AI trained on imitating human behavior would really be capable of being “read” the same way as a human.
It would probably work pretty okay for FPS games, racing games, maybe RTS?
It’s hard to really say how this stuff will turn out, whether AI will develop patterns that can be read, and attempt to read your patterns in turn, or whether it could be programmed to play with a personality, like intentionally repeating the same move that the opponent loses to a lot.
Then again, there was Killer Instinct, which gave this a real shot with its Shadow AI System, which I haven’t personally tried, despite owning the game, but apparently it worked really well.
This isn’t a pure machine learning solution, a lot of the structure of the AI, and the tweaking of various variables was done by hand (essentially, the structure the AI learns by was designed by people who knew what it should look like), but still, it’s really impressive the types of results they got. It took very observant players to know all the tendencies they should copy for this. If we got this in more games, if these replaced the standard fighting game AI, then I think fighting games would be a lot better for it. I don’t know how hard that is, but this sounds from the talk like it was thrown into KI unceremoniously, as a cheap side-mode instead of a major developmental focus.
AI like this could theoretically be deployed to train new players. Imagine training an AI to do nothing but move forward and sweep at different ranges to teach players to punish on block? Or only throw fireballs and anti-air to teach them to beat that? Or only jump in to teach them to anti-air?
There’s a big potential here, but it’s hard to say if we’ll ever see a feature like this ever again.