Copyright © 1999 Massachusetts Institute of Technology
Perhaps computers could educate our children as well as the best human tutors. This dream has inspired decades of work in cognitive science. The first generation of computer tutoring systems (called Computer Aided Instruction or Computer Based Instruction) were essentially hypertext. They mostly just presented material, asked multiple-choice questions, and branched to further presentations depending on the student's answer (Dick and Carey 1990).
The next generation of tutoring systems (called Intelligent CAI or Intelligent Tutoring Systems) were based on building knowledge of the subject matter into the computer. There were two types. One coached students as they worked complex, multiminute problems, such as troubleshooting an electronic circuit or writing a computer program. The other type attempted to carry on a Socratic dialog with students. The latter proved to be very difficult, in part due to the problem of understanding unconstrained natural language (see NATURAL LANGUAGE PROCESSING). Few Socratic tutors have been built. Coached practice systems, however, have enjoyed a long and productive history.
A coached practice system usually contains four basic components:
Any of these components may utilize AI technology. For instance, the environment might contain a sophisticated simulation or an intelligent agent (see INTELLIGENT AGENT ARCHITECTURE), such as a simulated student (called co-learners) or a wily opponent. The student modeling module's job includes such classic AI problems as plan recognition and uncertain reasoning (see UNCERTAINTY). The pedagogical module's job includes monitoring an instructional plan and adapting it as new information about the student's competence is observed. Despite the immense potential complexity, many intelligent tutoring systems have been built, and some are in regular use in schools, industry, and the military.
Although intelligent tutoring systems are perhaps the most popular use of AI in education, there are other applications as well. A common practice is to build an environment without the surrounding expert system, student modeling module, or pedagogical module. The environment enables student activities that stimulate learning and may be impossible to conduct in the real world. For instance, an environment might allow students to conduct simulated physics experiments on worlds where gravity is reduced, absent, or even negative. Such environments are called interactive learning environments or microworlds. A new trend is to use networking to allow several students to work together in the same environment. Like intelligent tutoring systems, many intelligent environments have been built and used for real educational and training needs.
Other applications of AI in education include (1) using AI planning technology to design instruction; (2) using student modeling techniques to assess students' knowledge on the basis of their performance on complex tasks, a welcome alternative to the ubiquitous multiple-choice test; and (3) using AI techniques to construct interesting simulated worlds (often called "microworlds") that allow students to discover important domain principles.
Cognitive studies are particularly important in developing AI applications to education. Developing the expert module of a tutoring system requires studying experts as they solve problems in order to understand and formalize their knowledge (see KNOWLEDGE ACQUISITION). Developing an effective pedagogical module requires understanding how students learn so that the tutor's comments will prompt students to construct their own understanding of the subject matter. An overly critical or didactic tutor may do more harm than good. A good first step in developing an application is to study the behavior of expert human tutors in order to see how they increase the motivation and learning of students.
However, AI applications often repay their debt to empirical cognitive science by contributing results of their own. It is becoming common to conduct rigorous evaluations of the educational effectiveness of AI-based applications. The evaluations sometimes contrast two or more versions of the same system. Such controlled experiments often shed light on important cognitive issues.
At this writing, there are no current textbooks on AI and education. Wenger (1987) and Polson and Richardson (1988) cover the fundamental concepts and the early systems. Recent work generally appears first in the proceedings of the AI and Education conference (e.g., Greer 1995) or the Intelligent Tutoring Systems conference (e.g., Frasson, Gauthier, and Lesgold 1996). Popular journals for this work include The International Journal of AI and Education (http://cbl.leeds.ac.uk/ijaied/), The Journal of the Learning Sciences (Erlbaum) and Interactive Learning Environments (Ablex).
Dick, W., and S. Carey. (1990). The Systematic Design of Instruction. 3rd ed. New York: Scott-Foresman.
Frasson, C., G. Gauthier, and A. Lesgold, Eds. (1996). Intelligent Tutoring Systems: Third International Conference, ITS96. New York: Springer.
Greer, J., Ed. (1995). Proceedings of AI-Ed 95. Charlottesville, NC: Association for the Advancement of Computing in Education.
Polson, M. C., and J. J. Richardson. (1988). Foundations of Intelligent Tutoring Systems. Hillsdale, NJ: Erlbaum.
Wenger, E. (1987). Artificial Intelligence and Tutoring Systems. San Mateo, CA: Morgan Kaufmann.
Copyright © 1999 Massachusetts Institute of Technology