Learning by teaching: the teachable agent

The computer lab at the Redwood City elementary school is jumping with fourth graders as they settle in front of their bright blue iMacs. In most school computer labs the remedial lessons would begin within minutes.

But unlike traditional computer lab learning models in which students are taught new concepts and then tested, this group of fourth graders is engaged in learning by doing the teaching.

The 9 and 10 year-olds strategize against the evil Moby, their computer enemy, to figure out a logic problem. Then they will “teach” their computer how to beat the opponent by inducing and applying the underlying rule that allows them to win. Students have multiple opportunities to check their rules and receive feedback on whether or not they are correct. When they think they know the rule, they teach “good Moby,” then watch the two Moby characters battle it out to see if their teaching was successful.

In this case, the logic problem involves identifying the factors (such as sun or water) that are required for flowers to appear on a colored grid. The software is designed to be flexible so that teachers can easily swap the flower logic problem with science content by using a simple menu system.

The teachable agent is software developed by the AAA lab, headed by Associate Professor of Education Dan Schwartz, under the auspices of the Stanford Center for Innovations in Learning. The software allows students to teach a computer everything from cell biology to trigonometry and logic. And it is in the teaching, Schwartz and his research team believe, that the deepest learning takes place.

“People learn when they teach,” says Schwartz, “and nobody gets hurt if the kid is a bad teacher, so the technique is better than peer teaching. We've been able to demonstrate that the student adopts the thinking of the agent. It opens up a lot of possibilities because it is quite a different tool. Usually the computer teaches the student. Here, the student is teaching the computer.”

Another recent Teachable Agent project involved Stanford students learning about cell metabolism. The students created a concept map which contained nodes of inter-related pieces of an equation, such as oxygen, algae and sunlight. After mapping the causal relationships of each, for example oxygen and sunlight lead to algae formation, they were able to “ask” the computer what would happen in various scenarios. If the student had correctly established the causal relationships, the computer would be able to solve the problem because it had been “taught” correctly.

The goal of Teachable Agents is to develop and understand motivating and effective instruction in which students teach the computer to explore the premise that people learn best when they teach.

Schwartz is also investigating the use of teachable agents for children with fundamental literacy and arithmetic difficulties. Supported by a grant from NSF, he is collaborating with neuroscientist Bruce McCandliss from Cornell University to test new teachable agents in New York City schools.

“The AAA lab is dedicated to creating innovative software that sets a model that other people can follow,” says Schwartz, who believes teachable agents will catch on and become widespread as a teaching tool. “But at the same time we are doing research that has practical value in how people learn.”

Back to News