Fig. 3

Cumulative reward, c_r(t), averaged over experiments on multiple independent samples of both synthetic (BA, BTER) and real world networks. In the BA network a, where probing the highest degree node is optimal, NOL “learns the heuristic”. NOL outperforms the baseline methods in BTER networks b, where the combination of heavy tailed degree distribution and relatively high clustering and modularity allows for discrimination. In some real networks (c, d, e), NOL either outperforms or closely tracks the best baseline, while in a real network with properties similar to a BA model f NOL is outperformed by the high degree baseline