Google may have conquered the ancient Japanese game of Go, but Microsoft has beaten humanity at Ms Pac Man. Maluuba, a Microsoft owned deep learning group, has managed to find a way to rapidly teach an AI how to work on tasks. Naturally, the first application is beating the world highscore at an arcade game.
The theoretical highscore for Ms Pac Man happens to be 999,900; after which the ticker rolls back to 0. Humans have never been able to reach this limit, and it’s one of those things that people assumed could never be achieved in our life time.
Maluuba solved the problem by breaking the problem down into smaller chunks. Instead of a single AI running at a time, the group used 150 agents running simultaneously. Each agent was then trained using a system of positive and negative reinforcement in order for it to determine what was best for achieving its ordered goal.
Pac Man was a proof of concept for the AI, as Microsoft has more practical aims in mind for the system. The company believes that this deep learning could be applied to making advances in natural language processing; a problem that still plagues voice recognition systems. It could also be used for management in planning sales targets and attempting to sort through company data.
AI playing games isn’t entirely new, but the platform has proven to be important for training the computers to react quickly and in real time. Google’s Go champion Deep Mind had time to think between actions due to the turn based nature of the game; but is also being trained to react quickly by subjecting it to classic games. Of course, it’s no secret that Google wants it to play Starcraft as well.
Maybe Microsoft should turn its attention to the classic RTS, and then we could see a proper showdown between two unbeatable super computers.
[Source: The Verge]