A group of engineers from Microsoft’s deep learning subsidiary Maluuba wrote an AI Application for the Atari 2600 version of Ms. Pac-Man, tearing up the notoriously challenging game into separate components for their algorithms to solve individually. Together, the AI components reached the maximum possible score of 999,990, Microsoft announced on Wednesday.
Maluuba’s AI architecture simulates the organizational structure of a company: there’s a “top AI agent” who takes suggestions from all the other algorithms. It ranks those suggestions based not only on which direction they recommend Ms. Pac-Man to move, but also how certain the algorithms feel in making their decisions.
“For example, if 100 agents wanted to go right since that was the best path to their pellet, but three wanted to go left because there was a fatal ghost to the right, it would give more weight to the ones who had noticed the ghost and move left,” Microsoft explained in a blog post.
Ms. Pac-Man is a more complicated game than the original Pac-Man, and no individual has ever achieved its peak possible score without cheating, according to The Verge. Of course, cutting up the game into components and feeding it to multiple AI algorithms might itself be deemed cheating, so the monumentality of Microsoft’s achievement comes down to whether or not you accept the researchers’ methods.
Still, it’s clear that beating Ms. Pac-Man is the latest in a string of game achievements for the AI sector of Silicon Valley giants. Google’s AlphaGo algorithm beat the human world champion titleholder of the ancient Chinese game Go in a match last month, the second time it has stolen the Go crown from a human player.