Microsoft AI Gets Highest Possible Score For Ms. Pac-Man On Atari, 4X The Human High Score

June 15, 2017


This is a video demonstration of Microsoft's Maluuba artificial intelligence program playing Ms. Pac-Man on the Atari 2600, and getting the highest score possible, with 999,990 points. For reference the human high score on the same console is 266,360. That's quite an improvement. Granted it's still not as good as I could do, but I'm a L337 gamer and would find a way to score those impossible nine extra points, possibly by trying to choke my brother out with the controller cord (I'm pretty sure that worked with the NES version of Dr. Mario). The man narrating the video mentions that this AI form of complex decision making could be used in corporate settings including "sales funnels and financial models", although using it for anything besides beating old video games sounds like a waste of time to me.

Keep going for the video, which provides a relatively easy-to-understand description of how the reinforcement learning process work.

Thanks to Clark, who informed me if any artificial intelligence program beats his Jungle Hunt high score he'll break its little binary fingers off.

  • captaindash

    The 2600 was my shit. I'd love to get one of those again.

  • Jenness

    They don't know no one else got 999,999...I did. I might have hacked something but pretty sure that happened. Screw that AI. :p

  • GeneralDisorder

    If you put that game into an emulator and let the AI hack it they could increase the number of bits that store the score or rewrite the game a dozen times a second or... Really, the sky's the limit. Robots are scary-better than humans at just about everything.

  • Draco Basileus

    We're doomed...

  • Meh

    Or most likely the AI would continue to rack up points but the basic programming of the game is limited to 999990.

  • GeneralDisorder

    Bejeweled had a maximum score and that was released in 2004.

  • Bling Nye
  • Jenness

    This is perfect.

blog comments powered by Disqus
Previous Post
Next Post