The official course title is "Advanced Topics in Artificial Intelligence." It is cross-listed as Computer Science 3710 (CRN 34548) and as ISSP 3565 (CRN 36602). These course numbers are used for all AI graduate seminars. It is okay to take this course more than once, because the topics are different in different years. For instance, last year Prof. Pollack taught a graduate seminar on Planning using these course numbers.
Auditors are welcome, but must keep up with the course readings.
The class meets on Mondays and Wednesdays from 10:00 am to 11:20 am. The first meeting, on August 30, will be in room 234 of Eberly Hall (formerly Alumni Hall; just behind LRDC and to the East). The subsequent meetings will be in a smaller room that is more conducive to discussions. Contact me if you miss the first meeting in order to find out where the other class meetings will be.
The objectives of the course are to understand the major work in the field of intelligent tutoring systems and to learn how build simple a simple intelligent tutoring system. A brief historical digression is the simplest way to explain what an intelligent tutoring system is.
The first tutoring systems were originally called CAI (Computer Aided Instruction), CBI (Computer-Based Instruction) and other names. Nowadays, they would be called instructional multimedia. They present the student with a window of expository material conveyed as text, graphics, animations or video, then ask the student a question. Depending on which response the student gives, the system brings up different windows of information next. Such tutoring systems are have been used in schools, industry and the military for 30 years. The Web is an ideal media for such systems. However, such technology is not covered here, because a graduate program at Pitt (Instructional Design and Technology) offers courses and degrees for those interested in it.
Tutoring systems based on Artificial Intelligent (AI) are called Intelligent Computer Aided Instruction (ICAI) or Intelligent Tutoring Systems (ITS). There are many types. One common type is the coached practice environment. The tutor coaches the student as the student solves a multi-step problem, such as solving a complex algebra word problem or discovering the laws governing a simulated economy. The student works with software tools, such as spreadsheets, graphs and calculators. The tools are often designed especially for the task, such as a kind of scratch paper that facilitates entering the types of equations or other notations that the task demands. To solve the problem, the student must make many user interface actions before entering a final solution. After each action, the coach remains silent or it may make a comment. Its comments may take the form of text or other indicators (e.g., highlighting entries, beeping). Most tutors remain silent as long as the student is making good progress toward a solution. They make comments only when the student has made a mistake, has wandered down an unproductive path or has asked for help. These comments are one main form of instruction. When the problem is finished, the coach may review certain key steps in the solution, a process called reflective follow up. This is another important form of instruction.
Another type of ITS engages the student in a Socratic dialog. The system begins by asking the student a question, such as "Why don't they grow rice in British Columbia?" The student can enter an answer in natural language, such as "It's too mountainous." The tutor responds with further questions, such as "If British Columbia were flat, could they grow rice there?" The tutor designs questions that will uncover and remedy the students' misconceptions, often by getting them to recognize contradictions in their beliefs.
Another type of ITS is critiques work that the student has produced. The student submits a piece of work, such as the design for a heat engine, a computer program, or an essay on a psychological experiment. The tutor analyzes the student's work and prints out a critique. For instance, it might point out that the heat engine could be improved by adding another stage, or that the computer program would be more elegant if it used recursion for a certain function, or that the essay missed 2 of the 5 main points that should have been made.
All these tutors are intelligent in that they can do the task that was assigned to the student, or at least, they can analyze the student's solution to the task and determine its quality. Thus, most tutors include an expert system. For instance, a tutor for physics would include an expert system that can solve physics problems. AI is needed for such expert systems. However, AI is often needed for other components of the system. If the tutor communicates in natural language, then AI is used to process the language. If the tutor uses qualitative simulation in order to test a student's design, then AI is needed in order to perform the simulation. If the tutor tries to select just the right comment in order to provoke learning without discouraging the student, then AI is needed to model the student's cognitive and motivational state. So many different kinds of AI are used in ITS that they are sometimes called "AI complete."
Nonetheless, many practical systems have been built and evaluated. Whereas CAI tutors generally raise students' grades by 0.4 standard deviation units, ITS generally raise their grades by 1.0 standard deviation units. This is not bad, given that the best instruction on the planet (professional human tutors) raise grades by only 2.0 standard deviation units. A main focus in this couse will be analyzing such empirical result, and in particular those involving human tutors, in order to find out how good ITS are and what do they need to do in order to become as good or better than human tutors.
For most classes, students will read a selection of papers and email one or more questions to the instructor per paper. These questions are the kind of questions that the student would ask in a public talk or when reviewing the paper for a journal. Student should try to ask the best questions possible and will be taught this archane art, in-so-far as it is possible to teach it.
Registered students (but not auditors) will submit a final project. Typically, this will involve implementing a simple tutoring system. A prolog program that implements a very simple tutor is here. You may wish to use it as a starting point for your system development.
Students should have a graduate-level introduction to AI, such as those provided by the AI core sequence at Pitt (CS 2710 & CS 2750; a.k.a. ISSP 2160 & ISSP2170). Student should also be proficient in either Lisp or Prolog. Students will be responsible for finding their own Lisp and Prolog programming environments.
There is no textbook. Students will photocopy readings handed out in advance. The list of readings is here.
I am usually available immediately after class for office hours. Otherwise, just call me first to see if I am in (624-7458), then stop by. My office is room 823 in the LRDC building. I read email (vanlehn@cs.pitt.edu) once a day unless I am unusually busy.
Grades will be based on class participation and the final project.
Last update: July 12, 1999