Games, man — they’re not just for people. A.I. researchers have recently kickstarted a new trend in harnessing the power of old Atari games to train artificial intelligence to into developing better problems solving skills.
There’s always been one game that’s acted as an impenetrable fortress for A.I.: Ms. Pac-Man. Finally, a deep-learning company owned by Microsoft has successfully built an A.I. system capable of beating the 36-year-old game, Microsoft announced.
How did the A.I. beat Ms. Pac-Man? Well, it wasn’t really an ending with a bang. The team from the deep learning startup called Maluuba, based in Montreal, Canada, watched as its robotic player ticked off to the maximum score of 999,990 points, which instantly reset back down to zero seconds later.
In a new paper uploaded to the arXiv repository, the research team details how the program learned to beat Ms. Pac-Man. Basically, the A.I. learned to defeat the game — one of the most difficult Atari games researchers are using to advance A.I. learning — by dividing up a large problem into smaller pieces, which A.I. agents individually began to tackle. A top agent acted as a leader and managed the individuals as a whole.
Reinforcement learning — in which actions are associated with a positive or negative response — was used to get the A.I. agents to learn through trial and error. More than 800 million frames of the game led to victory at the end.
Google’s DeepMind has been able to beat over 50 Atari games since 2015, but it never managed to unravel the complex difficulty of Ms Pac-Man.
“It’s approachable,” says Steve Golson, one of the co-creators of the arcade version of the game. “And yet it has this amazing complexity to it because of the randomness in the gameplay.”
The divide-and-conquer strategy employed by the Maluuba team worked extremely well. Breaking down a bigger problem into smaller pieces is exactly how humans intuitively proceed.