Data-Driven Feasibility for the Resource Constrained Shortest Path Problem

Resource Constrained Shortest Path Problems (RCSPP) have wide applicability, representing a flexible model for network applications. Furthermore, they frequently arise as subproblems in decomposition-based methods, as occurs in column generation for Vehicle Routing Problems. In all these settings, being able to perform early detection of infeasibility helps to strongly reduce computing times. For instance, dynamic programming is often used to design RCSPP algorithms: labels representing partial solutions are iteratively created and extended, and these can be dropped if they are found to have no feasible (and profitable) completion. Many fathoming heuristics have been proposed in the literature. We experiment a data-driven approach in this context, using supervised learning models to deal with the problem of detecting infeasibility. We design features which are not dependent on instance size, having different computing cost. We compare the tradeoff between computational effort and performance which can be achieved, when a binary classifier is employed. Our results indicate such an attempt to be effective.