So, I wrote a 3 layer feed-forward neural network (Pretty much identical with what's in AI For Game Developers, with much cleaner code) and I'm trying to think of a good use for it, like a "game" that it can play and you can see it learn. Right now I have it set up with a simple text-based "game" inspired by the shop interface in TES3: Morrowind. You give it the data on an item (price, condition, and color) and it decides if it wants to buy it, and then it asks you if its answer was correct. And it works good, the network quickly figures out that it should buy ones with high value and high condition, and color is irrelevant. Now, I want to move onto something slightly more complex, perhaps something graphical. Any ideas?

Some simple OCR(optical character recognition) is a good test for supervised learning. Maybe something like this applet: http://alrecenk.cjb.net:7790/java/show/ai/math/test8.html

The inputs are just a set of pixels and the output is the character corresponding to the highest output. Also it helps to scale down your characters (in my applet I scale to 10 by 10) and apply a little blur to assist the generalizing. Another good test is direct function approximation. It's not nearly as interesting, but it makes it obvious whether or not everything is working like it is supposed to. example applet of functions:
http://alrecenk.cjb.net:7790/java/show/ai/point/test.html

Hmmm, you could also play around with lunar landers:
http://alrecenk.cjb.net:7790/java/show/ai/point/test2.html

[Edited by - Alrecenk on November 11, 2006 10:33:34 AM]

I'm going to mess around with the Lunar Lander idea, that sounds like a good idea. Thanks.

Try teaching a ragdoll how to stand, run, jump, crawl, walk, swim, etc.

Like this guy:
http://video.google.com/videoplay?docid=1002062030982551847&q=genetic+algorithm

Although this would involve other aspects of AI such as genetic algorithms. Looks fun though.

Oh, just so you know the lunar lander applet I posted wasn't actually a neural net. I never got my neural nets to do supervised learning that was that complex with that much accurracy that fast. Sure they would land it, but usually scored well below the player and I had to run alot of extra training cycles and apply some other techniques to get a decent landing.