Slashdot Mirror
GA Tech Researchers Train Computer To Create New "Mario Brothers" Levels
An anonymous reader writes with a Georgia Institute of Technology report that researchers there have created a computing system that views gameplay video from streaming services like YouTube or Twitch, analyzes the footage and then is able to create original new sections of a game. The team tested their discovery, the first of its kind, with the original Super Mario Brothers, a well-known two-dimensional platformer game that will allow the new automatic-level designer to replicate results across similar games.
Rather than the playable character himself, the Georgia Tech system focuses instead on the in-game terrain. "For example, pipes in the Mario games tend to stick out of the ground, so the system learns this and prevents any pipes from being flush with grassy surfaces. It also prevents "breaks" by using spatial analysis – e.g. no impossibly long jumps for the hero."
This isn't a discovery. It's an invention.
why is it posted here as news?
Red Alert 2 had that in 1999. Even smarter than that IMO.
Moderating "-1, Disagree" is simple censorship. Have the guts to post your opinion. -- Spazmania (174582)
Wait until their ai starts scanning Kaizo videos.
Finally, an intelligent level designer.
Too bad it will end up used by idiots. Look at how FPS map design devolved over the past 20 years ...
This sounds like a basic sequence generator. You can do them with n-grams or anything that gives you the probability of an item coming up next and you pick the next item by the probability or some other heuristic. Basically you have a list of all possible tiles and you perform a greedy traversal through them using some specific heuristic. There is absolute nothing new in any of this. It's decades old. Even using YouTube videos instead buying the games themselves and using a camera or directly scanning the game isn't new. The article claims their solution works on any game, but provides no examples. I'd suspect there's some hard coding in there that says not to make jump sequences too long. It doesn't sound like their heuristic is self-learning (and the normal way of doing that is genetic algorithms or any general regression algorithm).
I would expect this to be more of an Into to AI class project, not a research project. Those who don't learn history are doomed to repeat it. The students probably do a good job getting it working, but is it valid to call it new research? Anyone have a link to their actual work?
I hope they're not paying for education...
Level generation is not new. Having a level generator that poorly positions obstacles or creates impossible jumps is not a level generator, it's a broken level generator. There is literally nothing new here whatsoever. I made a flash maze game back in 2005 that did exactly this: give it a few rules to follow and hey presto! New level that's solvable.
Even a decade ago it was hardly ground breaking.
The level generator was tested with the original Super Mario? Has it been open sourced?
Or do we finally have an explanation for what sort of twisty brain wrote the "cat mario"/syobon levels?
This is the crowed who are on a massive ping sweep across the entire IPv4 address space, who refuse to blacklist addresses, and when you complain block you at their firewall from sending in complaints, but still manage to fuck about with your systems.
Try checking for these IP addresses:
128.61.240.31
130.207.203.2
143.215.130.15
143.215.130.239