A comparison of learning and replicator dynamics using experimental data

We compare the explanatory power of two popular adjustment models, replicator dynamics and simple belief learning dynamics. The data, gathered over 18 laboratory sessions each with 8 to 16 profit-motivated players, comprise more than 100 runs (each with 10 or more stages or periods) of single and two population 2×2 bimatrix games with unique interior equilibria. To equalize the number of free parameters we suppress individual player and treatment effects known to be highly significant in the simple belief learning model. Nevertheless the belief learning model narrowly outperforms the replicator model on the single population data and greatly outperforms it on the two population data.