An experimental investigation of stochastic adjustment dynamics

This paper describes an experiment designed to test which, if any, stochastic adjustment dynamic most accurately captures the behaviour of a large population. The setting is a large population coordination game, the Language Game of Neary (2012), in which actions are strategic complements and two homogeneous groups have differing preferences over equilibria. We find that subject behaviour is highly consistent with the myopic best-response learning rule with deviations from this rule that are (i) dependent on the myopic best-response payoff but not on the deviation payoff, and (ii) directed in the sense of being group-dependent. We also find a time trend to deviations, with the magnitude tapering off as time progresses. This is in contrast to much of the theoretical literature that supposes a variety of other specifications of learning rules and both time-independent and payoff-dependent explanations for deviations.