Graph-SND

Graph-SND evaluation on a precomputed graph reduces pairwise computations from all unordered pairs to only graph edges. Graph-SND exactly recovers SND when the aggregation graph is the complete graph with unit weights. Graph-SND is defined as a weighted mean of pairwise behavioral distances over graph edges. Graph-SND modifies only the aggregation layer of SND while leaving policies, training, rollout estimation, and pairwise distances unchanged. Graph-SND can express complete, local, and sparse random aggregation regimes.