Heterogeneous Noise and Stable Miscoordination

Coordination games admit two types of equilibria: coordinated pure equilibria in which everyone plays the same action, and inefficient mixed equilibria with miscoordination. The existing literature shows that populations will converge to one of the pure coordinated equilibria from almost any initial state. By contrast, we show that plausible learning dynamics, in which agents sample the aggregate behavior of the opponent’s population and best reply to their samples, can induce stable miscoordination if there is heterogeneity in the sample sizes: some agents base their choices on noisy small samples (anecdotal evidence), while others rely on large samples.

Joint work with Srinivas Arigapudi and Amnon Schreiber