Using neural networks to predict student behavior in intelligent tutoring systems

One of the important components in an intelligent tutoring system is the student model. This model is used to predict what the student may do next as well as to serve as a repository of past student solutions. The student model is important in that it can help to direct the student to unknown material when enough concepts have been mastered and to material that needs to be reviewed when the student is unsure. Some student models have tried to predict student solution steps by restricting the interface to the point where the student cannot make an unknown move. Others do not concentrate on prediction, but instead concentrate on remedying errors in problem solutions.Since the problem of prediction is difficult, the research reported in this dissertation concentrates on solving this problem by using a tool that previously has not been used to implement a predictive student model. This tool is the neural network which has the ability to generalize over a set of student answers. This ability gives the network the capacity to answer as the student would on problems that the network has never seen before. Given this exciting possibility, this dissertation chronicles a series of experiments performed with backpropagation and the development of a methodology to place neural networks into the student model of an intelligent tutoring system.Specifically, the experiments were performed in the domain of subtraction. For each experiment, neural networks were used to simulate the cognitive processes of a student solving subtraction problems. The subtraction data presented to a network could be correct, erroneous, or noisy. Further, the data could be incomplete to test the network's ability to reflect or predict the student's actions on unseen data. On average, the neural network in the final experiment was able to recognize over half of the missing data and 99% of the present data. Further, the network performed adequately in the presence of noisy subtraction data enforcing the overall conclusion that neural networks have the ability to function well in a predictive student model.