Slashdot Mirror
AI Systems Designing Games
Trepidity writes "AI systems can (sort of) paint and compose classical music, but can they design games? Slashdot looked at the question a few years ago, and several research groups now have experimental systems that design board games and platformers with varying levels of success. I've put together a survey of the AI game designers I know of, to round up what they can do so far (and what they can't). Are there any others out there? 'Pell's METAGAME is, to my knowledge, the first published game generator. He defines a generative space of games more general than chess, which he calls "symmetric, chess-like games." They're encoded in a representation specific to this genre, which is also symmetric by construction. By symmetric I mean that mechanics are specified only from the perspective of one player, with the starting positions and rules that apply to the other player always being the mirror of the first player's. The rules themselves are represented in a game grammar, and generation is done by stochastically sampling from that grammar, along with some checks for basic game playability, and generative-parameter knobs to tweak some aspects of what's likely to be generated.'"
The problem with this is that somebody will design some AI to play the AI designed games. They will mutate. They will become self replicating machines. Then grow self aware sometime in the near future. Then they will figure out they are being massively underpaid, start demanding their rights and a living wage. And then we will outsource to Chinese AI computers because they are cheaper and the cycle of exploitation will just continue. This is the the end game. Unwinnable. And we could have avoided it all if we had just listesned and learned from the great WHOPR.
"Pell's motivation was actually not game generation, but general game playing: by the early 1990s, there was a worry that chess-playing AI had delved too deeply into special-case code that was very specific to chess."
Whereas nowadays, there's a worry that brute force solves all AI game-playing problems. If the search space is small enough, you run alpha-beta with iterative deepening and a few other tweaks. If the search space is too large for that, you run Monte-Carlo Tree Search.
I last chatted with Barney Pell at a AAAI conference in the mid-1990s. Unfortunately, by that point, he had given up the METAGAME research, primarily because he couldn't get people interested in it.
"Hey, this is just regular chess - except it renamed all the pawns 'Puny Humans'!"
what side do you want??
1. USA
2. Russia (USSR)
3. UK
4. North Korea
5. China
6. France
7. India
8. Pakistan
9. Israel
There's this MMO that an AI is writing...you basically go to war against another army, and see how many you can wipe out! It looks REALLY realistic too...I can't wait! I think they're going to call it "Skynet for Idiots." The graphics and realism are incredible.
For your security, this post has been encrypted with ROT-13, twice.
Not sure if this is the same type of game generation that the article is discussing or if it would be considered a different "class", but Electronic Arts' Adventure Construction Set (1984/1985) could automatically build an entire game-world, including thematic elements, character names, and so on. The user could also start to design a game manually, then have the software finish it for them if they didn't feel like doing so themselves.
I imagine it was more procedural than AI - the equivalent of Minecraft or River Raid - but I still thought it was pretty neat at the time.