Explainable Product Configuration for Telecom via Constraint Reasoning

Abstract: Product Configuration is one of the most successful applications of Artificial Intelligence and is widely adopted in industry. Nowadays,intelligent systems not only have to be able to generate outputs but these outputs are desired to be understandable and interpretable to humans. This thesis explores how well the increasingly popular SMT solvers are performing reasoning tasks for product configuration systems from the telecom domain. The main focus is to determine how well-suited the method is for both the general validation of a configuration and the computation of explanations for configurations that contain defects. Since computing explanations that are easily understandable is by nomeans trivial, we try to enumerate all possible explanations for a problem. The results show that for the used telecom use-case there is no single best method among the tested methods for both of these tasks, and different methods are suited for each one. While solvers that supportincremental solving, such as SAT and SMT, are especially suited for explanation computation, their performance on checking the initial satisfiability of the model is often worse. General modeling techniquessuch as MiniZinc prove to be very competitive for validation checks but suffer from the repeated compiling when trying to compute explanations. Further, using Knowledge Compilation, one can reduce the problem to amore specific problem class where the reasoning can be executed more efficiently in theory. In practice, this method suffers from the explosion of variables in the compiled model and cannot out perform the other methods in our experiments.