Microsoft Uses AI to Beat Ms. Pac-Man

Posted on June 14, 2017 by Paul Thurrott in Cloud with 11 Comments

Microsoft researchers have developed an artificial intelligence-based system that discovered how to achieve the maximum possible score on the 1980s arcade game classic Ms. Pac-Man.

This may sound trivial. But the firm says that this method could have broad implications for teaching AI agents to do complex tasks that augment human capabilities. And as it turns out, Ms. Pac-Man is apparently among the most difficult of video games to beat.

The researchers come from a Canadian company called Maluuba that Microsoft acquired earlier this year. They set out to use AI to beat the Atari 2600 version of Ms. Pac-Man, which allows a maximum possible score of 999,999. Why? Because the problem requires the AI to break the game solution down into small pieces and then assemble them in the correct order at the correct time. And this process is very similar to how the human brain is thought to work.

As it turns out, Ms. Pac-Man is particularly well-designed for this work. It was intentionally designed to be less predictable than its Pac-Man predecessor, so each game is different, and random.

Their solution is called the Hybrid Reward Architecture, and it utilizes over 150 individual agents that work in tandem to master the game. A “top agent” assembles suggestions from the others and decides where to move the on-screen character as the game progressed.

The researchers expect to channel this work into other AI research such as natural language processing. But maybe they could think about tackling Galaga first.

But seriously, you can learn more about this work on the Microsoft website. It’s interesting stuff.