The virtual robots, nicknamed "student agent" and "teacher agent," learn from each other as the student agent navigates through a dizzying maze trying to outrun four colorful ghosts all while gobbling up pellets. When the teacher agent detects its disciple is in trouble, it jumps in to give advice.
Knowing the precise moment to give advice is a skill most teachers and parents have crashed and burned trying to master. Unlike its human counterpart, this new algorithm developed by Taylor’s team recognizes the fine line between inattentiveness and micromanagement. With this refined version of action advice, researchers found that several "student agent" players soon outshone their teachers because they learned to think independently. This was illustrated when student agents played against imperfect teachers and won.
"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.
According to Taylor, the easiest way to teach a robot new skills is to swap out its old 'brain' for a new one. But, it's not as easy as it seems because hardware and software isn't always compatible with the new version. Robots’ ability to teach each other is an important area of research, because someday robots could be put in charge of training successors, saving humans from the arduous task.
Beyond the more immediate goal of computers teaching each other, Taylor says his ultimate goal is to have a robot teach humans. “Anything we can do to speed up human learning will be great for education, for training,” he says.
