The stochastic critical node problem over trees

We tackle a stochastic version of the critical node problem (CNP) where the goal is to minimize the pairwise connectivity of a graph by attacking a subset of its nodes. In the stochastic setting considered, the outcome of attacks on nodes is uncertain. In our work, we focus on trees and demonstrate that over trees the stochastic CNP actually generalizes to the stochastic critical element detection problem where the outcome of attacks on edges is also uncertain. We prove the NP-completeness of the decision version of the problem when connection costs are one, while its deterministic counterpart was proved to be polynomial. We then derive a nonlinear model for the considered CNP version over trees and provide a corresponding linearization based on the concept of probability chains. Moreover, given the features of the derived linear model, we devise an exact Benders decomposition (BD) approach where we solve the slave subproblems analytically. A strength of our approach is that it does not rely on any statistical approximation such as the sample average approximation, which is commonly employed in stochastic optimization. We also introduce an approximation algorithm for the problem variant with unit connection costs and unit attack costs, and a specific integer linear model for the case where all the survival probabilities of the nodes in case of an attack are equal. Our methods are capable of solving relevant instances of the problem with hundreds of nodes within 1 hour of computational time. With this work, we aim to foster research on stochastic versions of the CNP, a problem tackled mainly in deterministic contexts so far. Interestingly, we also show a successful application of the concept of probability chains for problem linearizations significantly improved by decomposition methods such as the BD.