34 years later, Microsoft researchers conquer Ms. Pac-Man’s Atari 2600 port – Polygon

Microsoft researchers testing artificial intelligence and machine learning created an AI system that can hit the theoretical maximum score on the Atari 2600 adaptation of Ms. Pac-Man. No human has ever done this.

The AI was built by Maluuba, a startup Microsoft bought earlier this year, and it was testing a type of AI called reinforcement learning.

Simply described, about 150 AIs were created and went into the game fixated on an individual goal, such as finding the nearest pellet or avoiding a ghost. These decisions were collated and then managed by a top-level AI, which effectively took those decisions as suggestions and made the best choice from them.

There had to be a weighting of decisions, too. Obviously, hitting a ghost (while un-supercharged) is a fail state, so even if 100 AIs said to turn right and snag the next pellet with a ghost oncoming, the three saying don’t do that would have their advice weighted more.

Ultimately, the AI was able to rack up the maximum score of 999,990. The researchers say this is the highest score for this version ever achieved — by four times that over the nearest human score. It’s unclear what record they’re citing. Twin Galaxies, which is Guinness World Records’ source for all-time high scores, does not list a point total record for Ms. Pac-Man on the 2600. This archive of high scores lists 211,480 as the all-time high for the cartridge.

The white paper discussing the team’s findings and achievements says Ms. Pac-Man was chosen because its ghost behavior was more complex and unpredictable than Pac-Man’s. Indeed, elite Pac-Man players had to master patterns, often published in the back of video game fanzines, to get to world-class high scores. Ms. Pac-Man in the arcade was built to defeat such strategies. But the paper doesn’t explain how this translated to the home console version, or why that particular version was used and not a ROM of the arcade version.

What does this accomplish? Researchers theorize that their methodology could help a sales organization determine which potential customers to target at a specific time or specific day, freeing a sales executive to focus on the pitch instead of sorting through leads. It could also have applications in language processing.

Comments

Write a Reply or Comment:

Your email address will not be published.*