Abstract:

In a world in which information is collected increasingly quickly and comprehensively, it is virtually impossible to produce didactic material of new topics for teaching in an adequate time that meets the expectations of students and the requirements of lecturers.

The question is whether methods commonly used to analyse Big Data could also be helpful in the field of Machine Supported Teaching (MST). A long-term vision of this approach could be to automatically create material from all kinds of available data sources that facilitates instructors to teach the corresponding subject matters.

In this paper, we focus on classical Machine Learning (ML) algorithms for exploring education material. ML methodologies can be divided into three major classes: unsupervised learning, supervised learning, and reinforcement learning. For each of the classes there exists symbolic and sub-symbolic approaches.
We show that only adapted methods are promising. Even supervised symbolic methods such as Decision Tree Learning or Bayesian Learning are designed to ensure that the trained models represent the learning object as adequately as possible, which is in general unsuitable for human learning (and teaching). What is needed is an approach that does not overburden students but considers their level of knowledge.

Unsupervised methods could serve to prefilter a given set of information. Helpful is also a closer look at the theoretical basis of reinforcement learning (RL), where (artificial) intelligent agents explore and exploit the problem space automatically. By skilfully intervening in the (symbolic) RL process, it is possible to gradually make the acquired knowledge accessible to the human observer as well.

To verify the principal considerations, we highlight some fundamental challenges exemplarily in the field of elementary pawn endgames in chess. Even simple questions raise a lot of technical difficulties to overcome. By not focusing the attention on the mere solution of a position, but rather keeping the solver in mind, general rules for the correct treatment become more significant. Automatic classification of difficulty levels is decisive - for both, problems and solutions strategies. This categorization takes the student's level of learning appropriately into account in the presentation of the subject matter.

In the outlook it is discussed that the described method may - in principle - be able to generate textbooks automatically. The goal would not be to replace human teachers, but in contrast to support them in representing a learning object appropriately.

Keywords:

Machine Supported Teaching, Machine Learning.