Learning Feedback in Intelligent Tutoring Systems

Acronym: 
FIT
Term: 
March 2012 - March 2015
Research Areas: 
D
Abstract: 

Since about 1990, intelligent tutoring systems (ITS) have been getting more and more popular. Designing an ITS usually requires precise models of the underlying domain as well as of how a human tutor would respond to student mistakes. As such, the applicability of ITSs is typically restricted to well-defined domains where such a formalization is possible, and large scale applications where development costs do not play a signicant role. For ill-defined domains, human tutors still by far outperform the performance of ITSs, or the latter are not applicable at all. The goal of the FIT project is to develop novel ITS methods which extend the applicability of ITS systems to ill-defined domains by means of machine learning techniques which can autonomously infer structures and feedback options from given data (e.g., student solutions). For this purpose, prototypebased methods and recent developments for general non-vectorial data structures will be extended such that they allow to simultaneously structure solution spaces, learn metrics for structures, align student solutions with clusters of other solutions, and infer appropriate feedback based thereon. The adaptation mechanisms will be developed in fully unsupervised scenarios or settings with only partial feedback to take into account the requirements for ITSs in ill-defined domains.

Methods and Research Questions: 

Details are provided at the SPP (Schwerpunkt-Programm) web-page about Autonomous Learning.

Outcomes: 

During the first two years of the project's duration, several milestones have been reached: First, user studies to investigate different feedback mechanisms provided the justification to develop a general concept for feedback provision in ill-defined domains. Next, the definition of a universal, domain-independent structural representation for solutions,along with the design and implementation of an appropriate software infrastructure, paved the way to process student solutions from various domains and react to students' feedback requests in real time. To ensure meaningful feedback, different structural similarity measures have been investigated and compared in terms of their suitability to reproduce syntactic and semantic relationships between solutions. Another crucial cornerstone was to establish a general mathematical foundation to transfer well-known prototype-based machine learning schemes to arbitrary structural similarity metrics. This opens the way for prototype-based clustering and classification on sets of solutions which are represented by the aforementioned similarity measures. In the most recent contributions, the first successful steps towards an autonomous learning of metric parameters have been taken, and first promising results have been accepted for publication in the near future.

Publications: