Modeling Students' Learning Behaviors in Open Ended Learning Environments

Abstract

Open-Ended Learning Environments (OELEs) describe a class of environments that provide students with learning goals that are usually in the form of complex problem-solving and model-building tasks. OELEs also scaffold students’ tasks with a set of tools, but provide them the freedom on how they combine and use these tools to progress towards their goals. Novice learners may have difficulties in accomplishing their tasks in OELEs, and the system can be designed to provide adaptive scaffolding to assist them in acquiring useful information, constructing problem solutions, and assessing their solution outcomes. Providing appropriate and adequate adaptive scaffolds requires a comprehensive understanding of students’ learning behaviors and accurate assessment of their learning performance. This information can be accumulated in the form of learner models as students work in the system.

This dissertation presents an approach for designing and developing accurate and refined learner modeling schemes using student data collected from different OELEs. The approach aims to address the data impoverishment problem by applying a Reinforcement Learning (RL) technique combined with Monte Carlo Tree Search (MCTS) to augment initial data set of students’ action sequences collected when students work with OELEs in classroom environments. The goal of the RL+MCTS approach is to learn more accurate models of students' learning behaviors with Hidden Markov Models (HMMs). By setting different reward functions in RL, two sets of reinforced models were generated to categorize and capture evolutions of students’ learning behaviors. These can then be used as the basis to predict students’ performance and provide adaptive scaffolds to help them develop better learning behaviors and improve their learning performance. Statistic evaluations and empirical analysis were applied to assess the reinforced models. Experiments with two OELEs (i.e., Betty’s Brain and CTSiM) showed promising results, which demonstrates that this approach to be a good starting point for apply