You will never be the best on „Ms. Pac-Man” anymore
All of us knows this old 1980s arcade game – Ms. Pac-Man. And nobody could ever get the perfect 999,990 score on it. Now, it has been changed. Deep learning startup Maluuba has developed an AI system, which has achieved a perfect score. For the first time in Ms. Pac-Man history.
It could sound silly because actually using deep learning to develop programs that can win video games isn’t a new trick. Notwithstanding this fact, Maluuba accomplishment is a big deal for some reasons.
Ms. Pac-Man can look simple, but from algorithm point of view, the game is actually quite complicated. It is much less predictable than the original Pac-Man, which makes is tougher for players to beat it.
That’s why Maluuba took another approach to solve Ms. Pan-Man. Instead of developing one single intelligent agent to learn the game, Maluuba used 163 individual agents, to learn a single aspect of the game (e.g. 8 agents for ghosts, 1 for fruit, 154 for pellets behavior).
Each individual agent is focusing on the small part of the game and developing a course of action it thinks Ms.Pac -Man should follow. All these courses are then aggregated and weighted average of preferences from each agent. The program moves Ms. Pac-Man according to this decisions.
What is so revolutionary in this approach?
First of all, breaking up complex problems into smaller ones can make it much easier for deep learning system and allows to handle more complex behavior.
The researchers believe that this method could be implemented in real-world tasks in the future. As one of the researchers in a video explains „this advancement in reinforcement learning, can improve decision making in complex settings such as sales funnels, financial models or robotics.”
You might also like
Robots Designed For Life on Mars!
A team of researchers over at Airbus in Stevenage have completed development on a robot built around the idea of searching for life or any sign of habitability on the
PR2
The PR2 is a mobile manipulation platform built by Willow Garage. The PR2 software system is written entirely in ROS. As such, all PR2 capabilities are available via ROS interfaces.
Gentle as robot’s hand?
It sounds like an oxymoron, but maybe – thanks to Xiaodong Chen form Nanyang Technological University in Singapore – robotic hands will be able to be gentle.