Race is on to evolve the ultimate Mario
Programmers have been challenged to build an artificial intelligence that learns how to master a version of Super Mario
AI researchers have been set the target of mastering another complex game deeply rooted in human culture: the Nintendo classic Super Mario Bros.
As well as providing a chance for teams to engage in a bit of healthy competition, the contest could help make future computer games more interesting, co-organiser Julian Togelius at the IT University of Copenhagen, Denmark, told New Scientist.
“I’m interested in creating learning or evolving software that can make games more fun, or easier to produce,” he says.
The mod father
Togelius and colleague Sergey Karakovskiy’s contest challenges entrants to develop software that can learn to master Mario by playing it over and over; much as generations of gamers have learned to play the game in bedrooms the world over.
The pair hope it will bring together advocates of different ideas about how to make software learn; those who mutate and evolve their code to improve it and those who prefer more established, mathematical techniques to spot and learn relationships. “Having a direct shoot-out like this is a good way to compare them,” says Togelius.
Evolving a better Mario would typically involve creating a population of programs each able to play the game, but in ways that differ slightly from each other. Software would pit them all against the game and combine elements of the most successful to “father” a new generation of Marios, each with some random mutations of their own included.
Many rounds of this automated process can create programs well adapted to a particular task, just like a species evolving to better fit and exploit its environment.
Two submissions (work-in-progress):
Posted on the competition’s Google Group:
[More at NewScientist]