Adaptive Decentralized Routing in Small-World Networks

We study routing in a small-world network modeled by an augmented grid, where Kleinberg's expected polylog bound offers a theoretical explanation for the empirically observed efficiency of greedy routing. We improve upon greedy through decentralized machine learning, using (local) ensembles of experts to choose routing edges and learn from feedback. In experiments with both synthetic and real networks, we observe improvements over greedy routing even when nodes are blind (do not know their neighbors' locations) or a fraction of nodes are uncooperative.