Joseph E. Beck

High-Level Student Modeling with Machine Learning

We have constructed a learning agent that models student behavior at a high level of granularity for a mathematics tutor. Rather than focusing on whether the student knows a particular piece of knowledge, the learning agent determines how likely the student is to answer a problem correctly and how long he will take to generate this response. To construct this model, we used traces from previous users of the tutor to train the machine learning agent. This agent used information about the student, the current topic, the problem, and the students efforts to solve this problem to make its predictions. This model was very accurate at predicting the time students required to generate a response, and was somewhat accurate at predicting the likelihood the students response was correct. We present two methods for integrating such an agent into an intelligent tutor.

Macroadapting Animalwatch to Gender and Cognitive Differnces with Respect to Hint Interactivity and Symbolism

We have built empirical models of elementary-school students behavior from analyzing student interaction with a mathematics tutor with the objective of building teaching policies for individually different students. This model incorporates external information about the student, namely cognitive development and gender. It also incorporates hint features, namely the degree of interactivity and symbolism of each hint given. We found that boys benefit better from non-interactive and low-intrusive hints, while girls benefit better from highly interactive hints. We found that low symbolic hints are more effective for low cognitive ability students than highly symbolic ones, and the opposite happens for high cognitive ability students.

Using the Student Model to Control Problem Difficulty

We have created a student model which dynamically collects information about a student’s problem solving ability, acquisition of new topics and retention of earlier topics. This information is provided to the tutor and used to generate new problems at the appropriate level of difficulty and to provide customized hints and help. Formative evaluation of the tutor with 20 students provides evidence that the student model constructs problems at the correct level of difficulty. The problem generation technique is extensible for use in other problem-based domains. This paper describes the design and implementation of the student model and illustrates how the tutor adjusts the difficulty of a problem based on the student model.

Using a Learning Agent with a Student Model

In this paper we describe the application of machine learning to the problem of constructing a student model for an intelligent tutoring system. The proposed system learns on a per student basis how long an individual student requires to solve the problem presented by the tutor. This model of relative problem difficulty is learned within a two-phase learning algorithm. First, data from the entire student population are used to train a neural network. Second, the system learns how to modify the neural networks output to better fit each individual students performance. Both components of the model proved useful in improving its accuracy. This model of time to solve a problem is used by the tutor to control the complexity of problems presented to the student.

Adaptation of Problem Presentation and Feedback in an Intelligent Mathematics Tutor

We have developed an intelligent tutor for teaching remedial mathematics to community college students. This domain is fairly narrow in scope and is an important component of the college curriculum. The target audience often retains fragments of knowledge from previous courses which can aid them in learning; alternately, misconceptions can present conceptual stumbling blocks if students have misremembered what they learned previously. Thus, a system built with a strong student model can greatly benefit the teaching process. The tutor described in this paper tracks student skills along with a general acquisition factor, and uses this information for topic selection, problem generation, problem presentation, and dynamic feedback.

Using Rollouts to Induce a Policy from a User Model

This research describes the application of an executable user model to generate policies to adapt software to best fit the user. Our approach first gathers data describing how users behave, and uses these data to induce a computational model that predicts how users will perform in a particular situation. Since system designers have differing goals, our architecture takes an arbitrary goal that the designer would like to see users achieve. Our architecture than using rollout techniques to determine how software should act in a particular situation with the user in order to achieve the desired goal.

W3 - Applying Machine Learning to ITS Design/Construction

Machine learning is applicable to many aspects of ITS construction including student modeling, learning tutoring strategies, and providing education partners for a student. With respect to student modeling, learning techniques can be used to induce the student’s current state of knowledge. Learning teaching strategies promises to allow the development of more flexible systems that can adapt to the unique requirements of different populations of students, and to differences in individuals within those populations.