A method that allows a computer to give advice and "train" another computer, in a way that refers to a "teacher" and "student", has been developed by researchers at Washington State University's School of Electrical Engineering and Computer Science.
The paper was published online in Connection Science, and the research was led by Matthew E. Taylor, a professor of Artificial Intelligence. As it became known, the researchers put their “agents” (as the virtual robots used in the study are characterized) to function as teacher-student pairs: the goal was to train the “students” in two computer games, Pac-Man and a version of Starcraft. As the research showed, the “student” was able to learn the games, even surpassing the teacher.
Many science fiction fans will immediately think of menacing digital forms like Skynet from The Terminator or the Cylons from Battlestar Galactica, but according to Taylor, robots aren’t going to take over the world anytime soon because “they’re just too stupid.” Even the most sophisticated robots, he points out, get confused easily—and when they get confused, they stop working. It often takes two to three times longer than he thinks it should to get a robot working, he adds.
Training computers in video games is an important part of robotics research. In this way, robots can train each other on new tasks without human intervention: for example, a cleaning robot can train its replacement.
According to Taylor, the best way to properly train a robot for new tasks is to transfer the “brain” of an “experienced” predecessor to it. However, problems arise when the hardware and software are not compatible with the new model.
In the study, the researchers programmed the “teachers” to focus on action advice—specifically, when the “teacher” tells the “student” to take action. “We designed algorithms to provide advice, and we’re trying to figure out when our advice has the greatest impact,” Taylor says.
Source: naftemporiki.gr
