Can AI Evolve to Master Super Mario? A Look at NEAT and Lua Programming

[Drakkith]

Here's a neat video I found of an evolving AI that plays the video game Super Mario. The AI is based off of something called NEAT, or Neuro Evolution of Augmenting Topologies (paper here), and was programmed in Lua (source code here). I thought it was pretty cool, so I just figured I'd share.

 

[Borg]

It's fascinating how artifical learning algorithms are advancing.

 

[Drakkith]

It's fascinating how artifical learning algorithms are advancing.

Indeed. I actually created a really simple simulation of some evolving cells about a year ago, but it was nothing like this. There was no learning involved, the cells just adapted over time without any "intelligence".

 

[Borg]

This month's Scientific American happens to have several articles on AI. It also refers to neural networks similar to what is mentioned in the video.

 

[Borek]

Makes me think about http://www.bpp.com.pl/bitozoa2/bitozoa2.html

 

[Hornbein]

I've heard that numerous games are played successfully this way. It was a key step on the way to the Go championship.

 

[Drakkith]

I've heard that numerous games are played successfully this way. It was a key step on the way to the Go championship.

The "Go" championship?

 

[Ygggdrasil]

The "Go" championship?

Earlier this year, an algorithm designed by Google defeated the top-rated Go player in the world. Go is an ancient Chinese board game which has simple rules, but is very complex because of the number of possible moves (there are many more possible games of Go than there are of chess, for example).

In the final game of their historic match, Google’s artificially intelligent Go-playing computer system has defeated Korean grandmaster Lee Sedol, finishing the best-of-five series with four wins and one loss.

http://www.wired.com/2016/03/googles-ai-wins-fifth-final-game-go-genius-lee-sedol/

Here's a link to the academic paper on the algorithm: http://www.nature.com/nature/journal/v529/n7587/full/nature16961.html

 

[collinsmark]

Several times I've created Artificial Neural Networks (ANN) for fun and kicks (various languages: Matlab, C++ and C#). Thus far they were only more conventional feed-forward networks that were trained using back-propagation with well defined training sets.

I think this NEAT algorithm might be quite useful for applications where well defined training sets are not practical.

 

[Hornbein]

I did not think Go would be solved in my lifetime. It was an epoch-making event, IMO.