[ml] AI competition: play Mario Brothers

Geoff Schmidt gschmidt at gschmidt.org
Wed Aug 5 22:35:53 UTC 2009


On Aug 5, 2009, at 2:37 PM, Josh Myer wrote:
> For the people who have done some AI/ML work before: there's a
> competition on to build systems to play a version of Super Mario
> Brothers.

Once upon a time, I tried to train/evolve a neural-network-like-thing  
to play Super Mario Brothers 1. I started with a very simple  
architecture to see what would happen. It had some fascinating  
behavior.. I didn't give it any way to store state or measuring time  
(it had to look at the sprites and tiles on the screen and compute  
what buttons should be held down for the frame) so it ended up  
encoding state in Mario's velocity and position. After a few thousand  
generations, its play strategy involved creating a series of loops  
(Mario bashing himself across a pipe and falling back, for example)  
that were each slightly unstable and would eventually decay into a new  
loop, in the process "accidently" stomping goombas and grabbing  
mushrooms. It was hilarious but ridiculously overfit. It was  
interesting to see how limitations in its play arose from limitations  
in its cognitive architecture. It made me wonder how insects see the  
world.

I doubt I have anything useful to contribute here :)

geoff



More information about the ml mailing list