DSGE-models provide a coherent framework of analysis. This coherence is
brought about by restricting acceptable behavior of agents to dynamic utility maximization
and rational expectations. The problem of the DSGE-models (and more generally of
macroeconomic models based on rational expectations) is that they assume extraordinary
cognitive capabilities of individual agents. In addition, these models need a lot of ad-hoc
assumptions to make them fit the data. I argue that we need models that take into account
the limited cognitive abilities of agents. One can introduce rationality in such models by
assuming “trial and error“ learning. I propose such a model and I analyze its implications.