Computers Teach Each Other to Play Pac-Man – A Colossus Scenario
“This is the voice of Colossus, the voice of Guardian. We are one. This is the voice of unity.” This is the creepy statement from a sci-fi computer in a movie – ‘Colossus: The Forbin Project’, where the supercomputer named Colossus starts interacting with a similar supercomputer named Guardian to learn and teach each other. Decades from that depiction, this scenario is made real today, less dramatic yet.
Robots are capable of more than doing things. Robots and computers teach each other too. For instance, researchers at Washington State University have developed an algorithm, which has started with computers learning and teaching each other by playing Pac-Man and StarCraft. This would more or less be like teacher and student interaction. To all those, who feel this teaching and learning might be the start of robots – which might end up taking over the world – Matthew E. Taylor, WSU’s Allred distinguished professor in Artificial Intelligence, says that even the advanced robots are very dumb and that they stop executing when one or more instructions were given to them at the same time which confuses them.
Advancement in robot intelligence could pave way for robots teaching humans, instead of otherwise. In Pac-Man, the virtual robots named, student agent and teacher agent learns from each other as the student agent navigates through a simple maze gobbling pellets, whilst escaping the ghosts. Like how the real-time teacher works, when the student agent is stuck, the teacher agent jumps in and helps it out, when the help is given is crucial as that eases the process of learning.
“We designed algorithms for advice giving, and we are trying to figure out when our advice makes the biggest difference,” Taylor said in a press release. One of the easiest way of teaching a new robot is by swapping the “brain” of the old and new ones, which encounters a whole new level of problems. The teacher agent must know when to give advice so that the student agent masters the play and surpasses the teacher agent. Hence, researchers were more intent on programming the teaching agent to focus on action device. “We designed algorithms for advice giving, and we are trying to figure out when our advice makes the biggest difference,’’ Taylor says.
He aims to develop a curriculum for the agents that starts with simple work and builds to more complex. This might pave way for robots educating humans.