Super Mario AI Competition

http://julian.togelius.com/mariocompetition2009/

by fogus • 17 years ago

101 33 17 years ago

julian.togelius.com

I've long wondered why the DARPA challenge for automated cars is necessary when the code could be tested using video games for orders of magnitude less money.

CGamesPlay• 17 years ago

In theory there is no difference between theory and practice. In practice, however...

nazgulnarsil5• 17 years ago

crude simulations don't invalidate theories. i hate anti-scientific quips like that.

dmm• 17 years ago

Building the vehicles with the associated sensor and control systems were just as important as the AIs controlling I suppose.

stonemetal• 17 years ago

I have friend who works at NASA in sim and test. Let me assure you simulations don't have to cost an order of magnitude less when the gov is involved.

All joking aside you are talking about custom designed systems. How would you design a sim that works with whatever wonky sensor setup a team can come up with? Two isn't pre-challenge VandV up to the individual teams? They should have simed it long before it got to the competition if they were going to. The competition is to prove that it works in real life\bragging rights. It isn't nearly as cool to say my simulated robot is faster than your simulated robot.

maximilian• 17 years ago

Given perfect information, like from a video game (with access to the object model, etc), the DARPA challenges aren't really that difficult. You just have to plan a path and then stay on it.

Its the noise of the real world that is infinitely challenging. You have to incorporate probabilities into your decision making, which is much more difficult. Jumping a mario around is much more AI intensive because the amount of solutions is much higher. A car is fixed to roads and has a map. Mario might know where everything is, but knowing what to do is much more difficult.

rawr• 17 years ago

I like how a hefty fraction of the algorithms use neural networks so that they get better as they watch humans play.

The reason for this, of course, is that the developers just want a reason to play Mario :-)

rw• 17 years ago

Fine, yes; but you don't have to use neural networks to get that behavior.

naveensundar3• 17 years ago

AI contests like this and the DARPA Grand Challenge Race encourage solutions in a single domain. AFAIK in the DARPA Challenge there was overfitting and I doubt the generalization capability of the systems involved. We need more contests in the spirit of "General Game Playing" contests where the problem domain is unknown to the programmers. http://games.stanford.edu/. You only get a formal description before the game and a small amount of "thinking" time which the system has to use understand the game and formulate a strategy.

apu3• 17 years ago

There is a vast difference between AI for games (which are purely theoretical exercises) and AI for working in the real world. The fundamental characteristic of the real world is that there is uncertainty and ambiguity in every measurement and even your actions (did my car actually move when I pressed the accelerator?).

The kinds of systems you build to tackle these latter types of problems are very different from the kinds you need for the first type. Both have their place, of course, although IMHO, the latter type is more interesting, as it's a much messier area to work in.

naveensundar3• 17 years ago

Well, messiness is just one part of the problem. I usually see AI as a layered system that solves a general problem. The layers are broadly: lower level sensory systems(vision, speech recognition) and higher level symbolic reasoning systems. Recently, there has a been a lot of success in the lower layers due to Machine Learning and probabilistic systems. The messiness comes in the lower layers. When dealing with the higher layers you face much harder problems. Hard in the sense of Turing uncomputable. The only formal solution proposed for general AI is uncomputable http://www.hutter1.net/ai/uaibook.htm.

Also there is nothing which restricts real world problem solvers to a single domain. Most messy problems have a formal solution. The engineering part is what remains (but the engineering part is nevertheless hard!). Some non "messy" problems have no computable solution e.g. deciding whether an arbitrary argument in first order logic is true. They challenge the bounds on human thinking.

Also "General Game Playing" is not only about games. (Like Game Theory is not only about games).

psyklic• 17 years ago

Almost all AI these days (even academic research) is a "solution for a single domain," unfortuantely.

dkarl• 17 years ago

Oy, such a short timeline for a task like this! Submissions are due by August 14th or September 3.

SwellJoe• 17 years ago

On a related note, one of the more interesting lightning talks at YAPC this year was about the Tactical Amulet Extraction Bot, a framework for programmatically playing nethack. Here's a blog about it: http://taeb-blog.sartak.org/

And the code: http://search.cpan.org/~sartak/TAEB-0.03/lib/TAEB.pm

I think it's been discussed here at HN once or twice, as well.

apgwoz• 17 years ago

Programming video game AI is fun. I took a class at UPenn where we had to build a pacman playing agent with reenforcement learning. The gameplay was't standard pacman, but with just a few, general features and correct training it was easy to get pretty decent results. And it's not that hard to implement. Not sure how it'd work on Mario though. Anyone have any thoughts?

Super Mario AI Competition

Related Stories