How should computer-based educational tools represent Machine Learning (ML) systems for didactic purposes? We address this question using constructionist learning theory and the intelligent agent paradigm of AI. ML in this context is understood as generating and improving ”goal-directed” system behaviors by iteratively maximizing a ”goal function”.

We give a theoretical outline of the problem domain along the questions: How independent can ML concepts be from concepts of classical computer science (CS)? What are central concepts and processes that ML possesses? What are important properties of structural models of this kind of systems conducive to comprehension? Finally, we propose some design features of educational informatics tools for teaching ML and outline further research needs.