Microsoft researchers built AI that used a divide and conquer approach to beat 'Ms. Pac-Man'

The Maluuba team managed to get 999,990 points in Ms. Pac-Man by using an approach they call Hybrid Reward Architecture.

The team from Maluuba, a Canadian AI startup that Microsoft acquired early this year have built an AI that got the highest possible score on the Atari 2600 version of Ms. Pac-Man. Building the capabilities of AI agents to tackle video games is a test of the AI systems, but Ms. Pac-Man has proven a tough nut to crack among AI researchers. The Maluuba team, however, managed to get 999,990 points by using an approach they call the Hybrid Reward Architecture.

A number of discrete agents were used for the task. Each agent was made to focus on a very narrow goal and was rewarded for meeting very specific objectives, such as avoiding a ghost or gobbling up a particular pellet on the screen. The 150 agents were controlled by a top agent, which processed all the inputs from the agent into the final command. While the number of votes for a particular action was counted, there was also a weight given to the intensity with which an agent wanted to make a move. For example, avoiding a ghost was more important than grabbing a pellet.

The researchers managed to crack the game by dividing it into a series of very small problems. The overall architecture of the AI is similar to how the brain is believed to function. Rahul Mehrotra, a program manager at Maluuba said, "A lot of companies working on AI use games to build intelligent algorithms because there’s a lot of human-like intelligence capabilities that you need to beat the games." The research has been published at Arxiv and is available to the public.

also see