SethBling wrote a program called MarI/O that learned how to play Super Mario World. The video embedded below shows, and describes how the computer used neural networks 1, and genetic algorithms to complete levels. At first the computer started out with not knowing anything about the game, or the controls involved. It learned everything by itself using NEAT 2, or Evolving Neural Networks through Augmenting Topologies. NEAT is a paper by Kenneth O. Stanley, and Risto Miikkulainen from the Department of Computer Sciences, The University of Texas 2.
The algorithm is based on actual biological evolution. The video shows the results of a 24-hour evolutionary learning session by the computer. It took about 34 generations, using evolution based on a parameter, for the computer to be able to finish a level.
If you’ve five minutes today, and want to learn something about Neural Networks, then I highly recommend watching this video.
Machine Learns To Play Super Mario World Using NEAT.
In other Mario related news….
A programmer spends 800 hours (six years) making a crocheted Super Mario Bros. 3 Map. Read More about it.
-
Neural Networks: In machine learning and cognitive science, artificial neural networks (ANNs) are a family of statistical learning models inspired by biological neural networks (the central nervous systems of animals, in particular the brain) and are used to estimate or approximate functions that can depend on a large number of inputs and are generally unknown. ↩
-
NEAT: Evolving Neural Networks through Augmenting Topologies - Paper. ↩ ↩2
