Companies like Amazon and iTunes have nifty programs and algorithms tracking consumer activity so that if, for instance, you are a fan of soul music you'll likely find tunes by Stevie Wonder and Angie Stone in your suggestions box.
A team of scientists at The University of Arizona is extending that concept and, with a newly funded $300,000 grant, applying it to research on developing artificial intelligence and educational technology that would benefit the military, business and, someday, K-12 schools and higher education institutions.
The team – headed up by the UA computer science department head Paul R. Cohen – is attempting to maximize a tutoring system model by using enormous amounts of data about learners to improve the feedback provided by an intelligent tutoring system.
"Teaching people means making a sequence of dependent decisions," said Cohen, the principal investigator on the grant, awarded by the Defense Advanced Research Projects Agency, or DARPA, which is part of the U.S. Department of Defense.
"We're trying to optimize the value of each decision by reasoning algorithmically about how it sets up the student for future learning opportunities," Cohen added.
Cohen is working with Carole Beal, a UA cognitive science professor, UA computer science doctoral degree candidate Derek Green and Yu-Han Chang, a research scientist with the Information Sciences Institute at the University of Southern California on the one-year project to create the system.
The program would, in effect, already know what a student knows and could be able to match that knowledge with comparable students before suggesting particular texts, exams, videos, educational games, demonstrations and other Web-based tools to help improve that individual's knowledge.
"More concretely, it should help the student learn efficiently, with good recall, and, ideally, understanding as evidenced by transfer," the team noted in the grant proposal.
Cohen says to think of it this way: Say you're curious about learning about percussion instruments, so you type the term into a general search engine. Say a main page returns with some good details and, toward the end of the page, a list of hyperlinks for additional information. "The question is, ‘What do you look at next?'"
The program would know exactly.
"It's so far from what we do in the classroom where you have 30 students in a room in front of one person who is doing an absolutely heroic job at trying to customize the curriculum – but no one person can do that job," Beal said.
Consider a traditional classroom setting.
The curriculum is often linear, moving from one module to the next. And given the current demands on educators, chances for individualized and tailor-made instruction have sometimes become increasingly difficult. In the context of what to teach next, the answer is sometimes obscure.
"And sometimes the traditional curriculum is designed by people who know the material so well," Beal said. "But we're hoping to use some statistical information to effectively promote learning even if that's the way the expert would not do it."
Customized learning is "the motivating principle surrounding intelligent tutoring," Cohen said. "The idea is that the educational experience should be different for different students, and there are a number of reasons: different students have very, very different learning styles and different background, and you would like to adjust your lesson plan accordingly," he said.
"You often don't know the right decision until you see how well they do on the test in the end," Cohen added. "That's when you decide what to show them next."
Yet, the technology Cohen and his team are working to develop would essentially do that. One important point to make is that the technology would also be more apt to accurately funnel students in a particular direction with the more information about others in the system. That is to say, the more the system knows about learners in general, the more specific it can be about individual students.
The project is based on data mining work Cohen and Beal completed in the past. Earlier in the year, the team completed a pilot. Preliminary findings suggested students who used the model learned more quickly and were able to retain information and that the program improved as more students participated.
Using AnimalWatch, an online tutoring system developed by Beal and her colleagues, the team will test college students on finite-field arithmetic problems and trace how well they do on the math problems, which will be offered at random to one group and in a sequence to another. The team will also test student engagement and mastery or skill level.
"The first generation of intelligent tutoring systems incorporated master teachers' expertise," Cohen said. But this is the second generation.
"We're using what students know to help other students. It's not about teacher expertise, it's about statistics, and it's thoroughly 21st century."
But the role of the educator does not by any means become inconsequential.
"We never get the technology so good that we don't need the teacher," Cohen said. "But it means the teacher's role would have to change."