Researchers at Washington State University in the School of Electrical Engineering and Computer Science have combined two of my favorite things, robots and video games, to develop an experiment that is cooler than the sum of its parts.
The robots in question aren’t physical machines but rather, virtual robots, or agents, within a computer program, and they’re teaching each other to play “Pac-Man” and “StarCraft” with some pretty astounding results. The research focuses on action advice, or telling an agent when to act. Initially the student agent performed very poorly but over time they did learn to play the game and actually surpass the teacher agent in skill.
After the initial wow factor of the headline wears off, you might ask yourself what utility there is in teaching primitive artificial intelligences leisure activities and there is actually a pretty good answer. One of the major problems in robots and AI stems from the inability of one machine or intelligence to pass information on to another. Due to hardware or software problems, information can be lost when transferring from one machine to the next, allowing them to teach one another overcomes that concern. It also allows for the possibility of human and robot student/teacher relationships. This becomes especially impressive when considering that while “Pac-Man” may be relatively simple, “StarCraft” is fairly complex.
Imagine a robot that could do household chores or cook dinner to your specifications and tastes. Imagine that a few years later you upgrade your robot; you don’t want to have to go through that calibration process again, so the initial agent could teach its own replacement.
Current research is focusing on the optimal times to give action advice. Matthew E. Taylor, WSU’s Allred Distinguished Professor in AI and the author of the paper stated: ‘As anyone with teenagers knows, the trick is in knowing when the robot should give advice. If it gives no advice, the robot is not teaching. But if it always gives advice, the student gets annoyed and doesn’t learn to outperform the teacher.’
Robots appear to have similar problems with learning. Identifying the best times to give advice, and best times to hold your synthetic metal alloy tongue is an important task.
As computer technology continues to increase, the reality if thinking robots becomes more and more apparent. Teaching them how to learn seems like a no-brainer (pun intended) and specifically teaching them to hold off hoards of invading insects will only help us in the long run. The practical applications of eating dots and chasing ghosts are less obvious.
Source: E! Science News