Machine learning may be applied to different aspects of the design of ITSs and MOOCs, such as student performance estimation and prediction and automatic design of knowledge components. In this collaboration,
we plan to focus on the problem of curriculum design, where the learning activities designed by an instructor are organized in a sequence. The main objective of the collaboration is to study the problem of curriculum design as a sequential decision process and develop novel algorithms that could be integrated in education systems that
automatically adapt the sequence of activities depending on the performance of the learner and make the learning process faster and more effective.
Research directions
The objective of the collaboration defines a number of research problems and during this collaboration we intend to follow different research directions:
- Task 1: A MAB approach to curriculum design. Recent works (Clement et al., 2014; Liu, Mandel, Brunskill and Popovic, 2014) framed curriculum design as a MAB problem. For instance, in prior work by Brunskill and colleagues (Liu et al., 2014), the choice is how to set various parameters on a fractions numberline problem (such as having tick marks or not) in order to enhance performance on the subsequent numberline (the reward signal of the arm). However, there remain many interesting and important open issues. For example, an important objective is how to help the student advance in his/her knowledge (or maximize progression), but in this case, unlike the standard stochastic assumption in MAB, the performance of the learner is non-stationary. This will require the development of novel algorithms taking into consideration the specific structure of the learner-teacher interaction.
- Task 2: Multi-task and transfer learning. While the objective of curriculum design is to provide a specific sequence of activities for each student, many students have similar learning skills and behave similarly. Thus, we expect to be able to use the interaction logs from different students and reuse them in designing a curriculum for other similar students. This approach, usually referred to as multi-task and transfer learning, allows to leverage on the large amount of data collected from all the students enrolled in the past to create a prior on the curriculum that could be easily and quickly adapted to the specific needs of each future student. Online data collected in large educational systems like MOOCs have the potential to enable this vision.
