Minimizing Computing Costs of Policy Trees in a POMDP-based Intelligent Tutoring System

Uncertainties exist in intelligent tutoring. The partially observable Markov decision process (POMDP) model may provide useful tools for handling uncertainties. The model may enable an intelligent tutoring system (ITS) to choose optimal actions when uncertainties occur. A major difficulty in applying the POMDP model to intelligent tutoring is its computational complexity. Typically, when a technique of policy trees is used, in making a decision the number of policy trees to evaluate is exponential, and the cost of evaluating a tree is also exponential. To overcome the difficulty, we develop a new technique of policy trees, based on the features of tutoring processes. The technique can minimize the number of policy trees to evaluate in making a decision, and minimize the costs of evaluating individual trees.