Heuristic Search over Program Transformations

Abstract

Abstract. In prior work, we have developed a method for the automatic reconstruction of buggy Prolog programs from correct programs to model learners' incorrect reasoning in a tutoring context. The method combines an innovative variant of algorithmic debugging with program transformations. Algorithmic debugging is used to indicate a learner's error and its type; this informs a program transformation that "repairs" the expert program into a buggy variant that is closer at replicating a learner's behaviour. In this paper, we improve our method by using heuristic search. To search the space of program transformations, we estimate the distance between programs. Instead of only returning the first irreducible disagreement between program and Oracle, the algorithmic debugger now traverses the entire program. In the process, all irreducible agreements and disagreements are counted to compute the distance metrics, which also includes the cost of transformations. Overall, the heuristic approach offers a significant improvement to our existing blind method.