Expectations play an important role in modern economic theories. 2011 Nobel price winners Sargent and Sims do research centered around the rational expectations framework.
The econometric learning approach, as I understand it, tries to model economic agents as forming expectations, on the basis of some learning methodology extracting and updating forecasting models. This means independent whether the expectations are rational.
As in machine learning in general such models might be extracted purely from data or combined with some mathematical forms that's parameters might be calibrated and re-calibrated towards concrete market behavior in an intelligent manner (in principle like outlined in the Blank Swan Of Metal Treatment, if I have this right)
But it might be that it is preferable that the agents start with little knowledge.
However, the methodological ingredients are in Mathematica 8 (comprehensive Statistics and Probability Superfunctions, fitness for the petabyte age) and mlf (creating understandable computational models from data in an multi-strategy and multi-method environment atop Mathematica).
This is the first time that I announce that we work on integrating both into a system that speaks the domain-specific language of econometrics and emphasize on the above mentioned learning aspects and hope to make systems more adaptive.
I find the challenges very well expressed in Learning in macroeconomics .. especially do we need to see market participants as players in the econometrics game, as we saw options trades as players of the game of Black Scholes options theory?