Computers are besting us at everything these days, including our beloved childhood video games. MarI/O, built by programmer Seth Bling, may be 25 years late to Super Mario World, but the bot caught itself up to speed rather fast, thanks to artificial neural networks that evolve in hours rather than millions of years.
As you can see in the video, MarI/O starts out pretty dumb — at first, he mostly stands still, unsure which direction he should run. (Clearly, computer programs lack our left to right cognitive bias). But, as the training session evolves, combinations of buttons that move MarI/O further along in less time increase the bot’s fitness score. Like beneficial genetic traits in a population, neural connections that produce high fitness are selected to be “bred” into the next generation. Over time, this process causes MarI/O’s brain to evolve a network of complex connections capable of making incredibly nuanced decisions.
Its like watching ten years of childhood Super Mario mastery compressed into minutes. If that’s the power of artificial neural evolution, then I’m afraid the machines may have already won.
Code for the bot can be found here. For more background on the machine learning algorithms that built MarI/O, you can check out this paper.
[YouTube via FlowingData]