Researchers at Washington State University have developed a method which enables computers to teach skills and give advice to another computer, based on playing the Pac-Man video game.
Up to now human beings have enjoyed a significant advantage over computers and robots: they have the ability to teach one another and also to teach machines to work. Now however computers have shown that they are capable of getting their fellow-computers to actually learn skills. A method developed by researchers at the School of Electrical Engineering and Computer Science at Washington State University (WSU) enables a computer to give advice and pass on skills to another computer in a way that mimics how a human teacher and student might interact, not simply by removing the ‘brains’ of one and inserting them into another. Trying out a new approach to ‘machine learning’ – a process whereby machines learn automatically by means of algorithms – they succeeded in empowering one computer to show another how to play Pac-Man. The project, which was part-funded by the National Science Foundation, an independent US government agency, was unveiled on 27 March in a paper written by Matthew E. Taylor, who holds the Allred Distinguished Professorship in Artificial Intelligence at WSU.
Teaching that mirrors human behavior
In their study, the researchers programmed their teaching ‘agent’ computer to focus on action advice or telling a student ‘agent’ when to act. Using the new approach, the ‘teaching’ computer was able to impart advice to a ‘student’ computer in the same way that one person would teach another, apart from the algorithm that controls the frequency with which the teaching computer proffers advice. Results showed that if advice is given too often, the student computer does not manage to learn for itself, but that with too little advice the ‘student’ takes far too long to pick up the skills. Explained Matthew E. Taylor, who is leading the project: “We designed algorithms for advice-giving, and we’re trying to figure out when the advice makes the biggest difference.” With the right timing, the teaching agent is able to guide a student machine to learn to play Pac-Man and also a version of the StarCraft video game. The researchers were able to demonstrate that the student agent not only learned the games but in the end actually surpassed the teacher.
Applications in many fields
The implications of these research results are huge and could revolutionize the fields of robotics, machine learning and human-computer learning and interaction. The range of applications of the skill-learning approach goes way beyond games situations and could be applied to a wide variety of actions. In future this advanced machine learning could for instance enable obsolete factory robots to teach their replacements to do the job without any downtime. The team’s longer term goal is to create robots which can teach people. Meanwhile Taylor is also keen to reassure people who are worried about robots taking over the world with their enhanced artificial intelligence. “They’re actually very dumb,” he underlines, pointing out that “even the most advanced robots are easily confused. And when they get confused, they stop working.”