Optimal Probe Scheduling for Network Endpoint Localization

The objective of the network endpoint localization problem is to estimate the endpoints (source and destination) of some message passed through a communication network based upon the set of partially synchronized sensors that intercept the message. The topology of the network is unknown, however some prior information about node degrees in the network be available to the monitor. A preliminary training phase of the problem involves scheduling message transmissions in the network and observing the activated sensor set. This paper describes strategies for the training phase. The strategies are designed so as to reduce ambiguity in the estimation phase. Using the graph edit distance between local subnetworks around each probing site as a measure of ambiguity, the problem of optimal training design is formulated as a Markov decision process (MDP). Since the state space of the process is continuous, standard linear programming solutions are not applicable. Scheduling algorithms based on greedy approaches and relaxations to multi-armed bandits are discussed.