Robust Initial Conditions for Learning Dynamics A preliminary version

Although a complete dynamic model includes the equations of motion as well as the initial conditions, many of the learning dynamics in the literature do not entail a serious theory of initial conditions. This paper explores and evaluates five approaches to complete learning models with a theory of initial conditions. We find that (1) the initial period’s fit is sensitive to the model of initial conditions chosen, with some initial condition models consistently superior to others over games. (2) Looking at periods 2 onward, leading dynamic models are robust to initial condition models according to one-period-ahead goodness-of-fit criteria. (3) Longer horizon measures of fit provide the insight that models of initials conditions with initially high diversity sometimes outperform better specified models of initial conditions .