Google's DeepMind Artificial Intelligence recently beat the top Go player, Lee Sedol, with its latest AlphaGo program last year. Similarly, other games have been conquered by computers, including Checkers in 1994, Chess in 1997, and Jeopardy in 2011. Now, a new team backed by Microsoft has achieved the impossible and has acquired the perfect score in the classic arcade game, "Ms. Pac-Man."

A staggering challenge

Maluuba, the artificial intelligence team, acquired by Microsoft earlier in the year, has now showcased their new AI system that was able to achieve a 999,900 score on the Atari 2600 version of "Ms. Pac-Man." The score is the maximum achievable score in the classic game, which was previously believed to be unbeatable by a human or machine due to its intentional lack of predictability.


As of the moment, the highest score a human has ever reached in the game is 266,330 on the Atari 2600 version. Some players have achieved the top score, but only with the use of cheats, which mostly doesn't count. On the other hand, the record for the arcade version of the game is currently held by Abdner Ashman of Queens, New York, who beat the previous record held by Chris Arya of 920,310 points with his score of 921,360 points.

A different approach

According to the creator's of the new machine learning system, they uniquely approached the problem with "Ms.

Pac-Man." The team programmed a system that is comprised of a divide and conquer method and a reinforced learning system using different agents. Instead of trying to predict the movements of the ghosts and calculating the appropriate reaction, the team used 150 separate systems to handle different jobs.

As an example, one agent may be responsible for getting a specific pellet, while another would be responsible for tracking ghost movements. Essentially, all agents will then get a vote on the next appropriate move.


Their "suggested" moves would then be sent to a "top agent," which Microsoft referred to as a senior manager, that ultimately chooses the appropriate action based on all of the other agent's suggestions.

Further applications

The unique system only works because each agent is able to focus on one job, and all of their decisions are then combined into a single move that ultimately benefits the whole. According to the company, the Hybrid Reward Architecture process that is used with their system can be applied to different fields.

The AI learning system has a lot of potentials when it is applied to processes such as language processing, sales leads, and other practical applications.