Researchers at Microsoft have created a system of artificial intelligence that learned how to reach the maximum possible score of 999,990 points in the Atari 2600 version of Ms. Pac-Man.
Developed by Maluuba, a startup acquired by Microsoft last January, the Hybrid Reward Architecture is a method that breaks down the game into simple tasks and divided among over 150 AI agents. These agents each had a specific task, such as collecting pellets or avoiding ghosts, while a top level agent was in charge of deciding where to move Ms. Pac-Man based on the agents’ suggestions. Check out how it works in the video below.
Maluuba says Ms. Pac-Man was chosen for this experiment because it was less predictable than regular Pac-Man, which made it perfect for teaching AI agents how to respond to random environments. Reinforcement learning was used to achieve the perfect game, where the agents learned over time how to achieve rewards and lessen negative responses for every action tried. The company believes this method could be used for natural language processing in AI research or help a company’s sales by predicting potential customers at specific times.
CHECK OUT MORE
- Insurgency Sandstorm Brings Story to the Realistic Military Shooter
- Overwatch’s Horizon Lunar Colony Map Arrives on June 20
- GTA Online’s Gunrunning Update Now Live, Patch Notes Available
- The Pokemon Company is Working on a Main Line Pokemon Game for the Switch
About the author
