Towards Semi-Automatic Learning-Based Model Transformation

Kiana Zeighami, Kevin Leo, Guido Tack, Maria Garcia de la Banda

Abstract: Recently, [16] showed that the nogoods inferred by learning solvers can be used to improve a problem model, by detecting constraints that can be strengthened and new redundant constraints. However, the detection process was manual and required in-depth knowledge of both the learning solver and the model transformations performed by the compiler. In this paper we provide the first steps towards a (largely) automatic detection process. In particular, we discuss how nogoods can be automatically simplified, connected back to the constraints in the model, and grouped into more general “patterns” for which common facts might be found. These patterns are easier to understand and provide stronger evidence of the importance of particular constraints. We also show how nogoods generated by different search strategies and problem instances can increase our confidence in the usefulness of these patterns. Finally, we identify significant challenges and avenues for future research.