Empirical estimation of hybrid strategic learning models on
			experimental game data

Colin Camerer
Caltech

This talk will describe hybrid models used to characterize how people
change behavior in response to experience ("learn") in finite games.
The hybrid model assumes that people are reinforced by payoffs they
receive but also respond to foregone payoffs.  (Many models which are
popular in economics are special cases of the hybrid.)  The model fits
and predicts better than simpler models in a wide variety of games
(mixed-equilibrium games, coordination, dominance-solvable,
signaling).  I will also describe a "low-fat" version of the model in
which parameter values in the learning model are themselves functions
of the data.  This permits the model to react to "surprises" in an
intersting way and also economizes on the number of parameters that
need to be estimated or fixed a priori.  I will also describe some
results on "strategic teaching", in which agents understand they are
playing repeated with a learning agent and choose period t strategies
which maximize long-run payoffs (e.g., behaving in a trustworthy way,
or an aggressive way).