Microsoft boasts artificial intelligence that can 'beat' Ms. Pac-Man


Move over, Billy Mitchell. There's a new 1980s arcade video game master, and his name is...well, he doesn't have a name. And it's also not a "he" -- it's Microsoft's latest demonstration of artificial intelligence.

The team at Maluuba, a Canadian artificial intelligence firm acquired by Microsoft earlier this year, has managed to program a system that can play the Atari 2600 version of Ms. Pac-Man indefinitely without ever being hit by a ghost, allowing it to rack up a score of 999,990 before being reset to zero.

The highest score ever recorded by a human player was 266,330.

Detailed in an official Microsoft blog post, the process is highly complex due to the chaos and unpredictability built into the game. There have been many previous attempts to code an infallible AI player, but inevitably the game would always win due to random conditions that force Pac-Man in between two ghosts.

The solution Maluuba came up with was to effectively have 150 different "agents" that would each independently and simultaneously choose which direction to head at all times, and one agent designated as the leader would use that data to make the ultimate decision. This method is called Hybrid Reward Architecture.

Microsoft is quick to reinforce this as progress for the worlds of machine learning and artificial intelligence as a whole. The idea of solving something that can never be solved formulaically is the goal of all these types of technology, like IBM's Watson computer's famous mastery of the quiz show Jeopardy.

But the analysis on the other side is just as swift and just as salient. The Ms. Pac-Man "algorithm," to use a loose term, does in fact assign hard values to the various items and obstacles in the game. There are only "always avoid this" and "always approach this" determinations. Finding an equivalent real-world problem to apply this kind of logic to is almost a pointless exercise.

Nevertheless, it is a definite and optimistic step forward in the always emerging and endlessly interesting field of AI and machine learning.

© 1998-2018 BetaNews, Inc. All Rights Reserved. Privacy Policy - Cookie Policy.