Microsoft’s AI has finally achieved what was apparently impossible for humans: a perfect score on the classic arcade game Ms. Pac-Man. The software giant’s Maluuba team developed an artificial intelligence system that figured out how to hit a score of 999,900 points, which is considered a perfect score.
Maluuba develops game-focused AI for Microsoft
The company announced the achievement for Ms. Pac-Man in a blog post today. Microsoft acquired the Canada-based Maluuba deep learning team in January, and apparently, it paid the team members to play video games — lots of video games. Then again, it seems that more and more people are getting paid to play video games for fun as YouTubers or in eSports or other related areas.
In this case, Microsoft’s Maluuba team used a divide-and-conquer technique as they worked with their artificial intelligence to master the 1980s-era classic game Ms. Pac-Man. Divide and conquer is actually a branch of artificial intelligence called reinforcement learning, Microsoft explained. Using their divide-and-conquer method, Microsoft’s AI team was able to get their artificial intelligence to reach the perfect score of 999,990 on the Atari 2600 version of Ms. Pac-Man.
Microsoft says Ms. Pac-Man is just the beginning
Although right now it’s just Ms. Pac-Man, Microsoft’s AI team believes it could be only the beginning. The company explained that the method “could have broad implications for teaching AI agents to do complex tasks that augment human capabilities.” In fact, Microsoft’s blog quotes an associate professor at Montreal’s McGill University who also sees it as a major achievement among researchers in the AI field.
Additionally, Prof. Doina Precup was impressed with the way the Maluuba team achieved the perfect score on Ms. Pac-Man. Microsoft’s AI team divided the massive problem of reaching a perfect score on Ms. Pac-Man into bite-size pieces and had team members working on small pieces of the big problem. Some members of the team handled the task of keeping away from ghosts in the game, while others go a reward for locating a particular pellet.
Precup explained that the way they handled the problem is similar to the way some researchers believe that the brain works. As such, the implications for what they have done could be very important for teaching AIs how to perform complex tasks even though they do not have much information about them. She even went so far as to suggest that Microsoft’s AI team has advanced the field “another step toward more general intelligence.”
Microsoft’s AI not the first to study using video games
Microsoft’s AI team isn’t the first to use video games to test artificial intelligence systems either. Apparently, AI researchers prefer to use video games because they are better than controlled games like chess when it comes to mimicking the chaos experienced in the real world.
A couple of years ago, the Google DeepMind AI team used a wide variety of different Atari games in the development of their artificial intelligence. The DeepMind team provided negative and positive feedback, also a form of reinforcement learning, to their AI system as it worked toward mastering the Atari games.
Maluuba program manager Rahul Mehrotra explained that they selected Ms. Pac-Man because of how many different types of situations the AI can encounter while playing it. The game is very unpredictable, which makes it a valuable tool for AI researchers.