Model-Based-Diagnosis for Assistance in Programming Exercises

Implementing AI methods can be an effective way to understand their inner workings, in particular when learners have to locate bugs. However, programming tasks are time consuming and can be extremely challenging for students. In order to provide assistance, code evaluation platforms have been developed that give immediate feedback in the form of discrepancies between expected and actual output for test cases. While such tests clearly indicate wether or not an implementation is faulty, they do not assist learners in locating the fault in their implementation. We propose to diagnose solution attempts, explaining potential faults with respect to abstract behavior. By framing programming tasks as functional models, we can diagnose the underlying concepts of a task and provide feedback. In this paper we focus on abstract data types as a basis for AI algorithms. The diagnosis method described in this paper produces an explanation of a fault in the form of a description based on a reconstruction of hidden program states. Applying the method to student submissions in a programming class shows that the proposed method can effectively identify and locate faults.