An artificial intelligence system created by Microsoft achieved the perfect score in the 1980s arcade game Ms. Pac-Man, a goal that has long eluded gamers.
The feat was done by Maluuba, a deep learning startup that Microsoft purchased early this year. The team used a divide-and-conquer method that may have significant implications in how artificial intelligence systems can be taught to carry out complex tasks in place of humans.
Microsoft AI Scores 999,990 In 'Ms. Pac-Man'
The artificial intelligence system that Microsoft researchers from Maluuba created was able to play the Atari 2600 version of Ms. Pac-Man flawlessly, achieving the maximum score of 999,990.
The feat has not been achieved by human gamers, even after so many years of trying. The highest score that a human has ever reached in the Atari 2600 version of Ms. Pac-Man is 266,330, with the perfect score of 999,990 only obtainable by using cheat codes.
According to McGill University associate professor of computer science Doina Precup, the perfect score is a significant achievement in the field of artificial intelligence because Ms. Pac-Man is a very difficult game for the technology. Ms. Pac-Man was programmed to be even less predictable than the original Pac-Man, designed in such a manner so that players will keep dropping their quarters to beat high scores.
How Did AI Beat 'Ms. Pac-Man'?
The road to the Ms. Pac-Man perfect score required researchers to divide the task of mastering the arcade game into small pieces that were then distributed to more than 150 artificial intelligence agents.
This method, named by the Maluuba team as Hybrid Reward Architecture, had the agents working together to achieve the perfect score in Ms. Pac-Man. While some agents received rewards for finding specific pellets, others focused on keeping distance from the ghosts. The team then created a top agent, which received suggestions from all other agents and decided where Ms. Pac-Man should move.
The top agent considered the number of agents pushing forward a suggestion, but also made its decision based on the intensity of the suggestions. One given example was when 100 agents wanted to go right to pick up a pellet, there were three agents that wanted to head in another direction to escape a ghost. The top agent would give more weight to the three agents so that Ms. Pac-Man would not get caught.
Video Games To Train AI System
Using video games to train artificial intelligence systems is not new.
In February 2015, a report revealed that Google's DeepMind artificial intelligence system was able to learn how to play old games on the Atari 2600 just by watching humans play. Deep Q-Network, the AI gamer powered by DeepMind technology, was able to play 49 different titles and beat the top scores of skilled human gamers in more than 20 games.
In November 2016, after defeating South Korean world champion Lee Sedol in the ancient board game Go, DeepMind moved on to a bigger challenge of trying to master StarCraft II.
Ultimately, the goal in the creation of artificial intelligence systems is to assist humans in their tasks, and in some cases replace them so that humans can focus on other activities. Who knew that Ms. Pac-Man would somehow be involved in such an endeavor?