This is super cool:
Using deep learning to develop programs that can defeat video games isn't a new feat, but this accomplishment is notable for several reasons.
First of all, it's notable because of the type of game chosen. The old 1980s arcade games weren't designed to be beaten - they were designed to keep people pumping in quarters. And when Ms. Pac-Man was developed, it was actually programmed to be less predictable than the original Pac-Man, so that it would be tougher for players to beat it.
The second and perhaps most notable aspect of this accomplishment, though, is the approach that the researchers took to solve Ms. Pac-Man. Rather than develop a single intelligent agent to learn the game, as other researchers have done, this team instead used a number of simpler intelligent agents to learn a single aspect of the game. For example, there are agents learning about ghost behavior, about fruit behavior, about pellet behavior, etc.
Each individual agent (there's over 100), develops a course of action it thinks Ms. Pac-Man should follow based on the small part of the game it's focused on. Those decisions are then aggregated, and the program moves Ms. Pac-Man based on the weighted average of preferences from the individual agents....
