Learning with Bounded Memory in Stochastic Models.

Learning with bounded memory in stochastic frameworks is incomplete in the sense that the learning dynamics cannot converge to an REE. The properties of the dunamics arising from such rules are studied for models with steady states. If in standard linear models the REE is in a certain sense expectationally stable (E-stable), then the dynamics are asymptotically stationary and forecasts are unbiased. We also provide similar local results for a class of nonlinear models with small noise and their approximations.