Slashdot Mirror
Robots Are Coming For Our Ms. Pac-Man High Scores (fastcompany.com)
A Microsoft-made AI system has achieved a perfect score of 999,990 points on the Atari 2600 version of the classic 'Ms. Pac-Man.' From a report: Researchers at the Microsoft-owned deep learning company Maluuba have used an AI system to break the all-time Ms. Pac-Man record. In a blog post, Microsoft wrote that, "using a divide-and-conquer method that could have broad implications for teaching AI agents to do complex tasks that augment human capabilities," Maluuba's AI was able to record a perfect Ms. Pac-Man score of 999,990 on the Atari 2600 version of the game, breaking the all-time record of 933,580.
When a computer beat the world champion at chess, I didn't care. When it happened with Go, I didn't care.
But now it's personal...
#DeleteChrome
More AI BS. Just stop it already. This isn't AI. Some idiots will be claiming Eliza is AI in their next funding cycle.
Seriously, it's just a pattern matching algorithm.
It's not able to do other things.
AI can both walk and chew gum at the same time.
Oh, wait, ok, maybe it's smarter than the Comrade-in-Chief, but that's still not AI.
-- Tigger warning: This post may contain tiggers! --
A quick Google search suggests that the limiting factor in Ms. Pacmac top high scores is luck. Fruits that give a varying number of points show up in each stage. There are a finite number of them and it's pure luck whether you get one worth a high number of points. While getting to the Kill screen (essentially the end of the game) takes skill, it's well within human ability. Doing it enough times so the stars align and you get 1000000+ points is not as easy. RNG manipulation might be possible but it seems like someone would have mentioned it if it was viable to be performed by human. This is for the arcade version which is the version the 933580 human world record was made on. I don't know if the Atari 2600 version has any important differences but if it does the initial comparison between scores was invalid to begin with.
This particular Atari game was one of the few games that resisted to Deep Q Learning (a form of Reinforcement Learning invented by DeepMind). Many researchers have tried over the last couple of years to solve it. This time, Microsoft found an ingenious solution to the problem, that combines experience from multiple agents and learns to form sub-goals. Their solution could mean that in the future it might be easier to apply reinforcement learning to other settings, such as robotics. The interesting part about reinforcement learning is that it learns dynamic behavior, as opposed to static classification. It learns to act intelligently. This kind of AI is invaluable.