A behavioral model for mechanism design: Individual evolutionary learning

We are interested in how Groves-Ledyard mechanisms perform when used repeatedly in a sequence of one-shot games where agents know only their own preferences. In particular, how fast do the mechanisms converge to the stage game Nash equilibrium and how does that speed of convergence depend on the mechanism parameter [gamma]. Prior theoretical and experimental work provide little guidance. Neither do existing behavioral models designed for small games with a small finite number of strategies. For example, even though experience weighted attraction learning is very successful in modeling behavior in one-shot games with very small, finite strategy spaces, it is not successful in modeling behavior in repeated games with a continuum strategy space unless one wants to be involved in fine tuning. We provide a behavioral model, individual evolutionary learning. The time to first convergence is predicted to be smooth and U-shaped in [gamma]. These predictions are robust to a wide range of parameter values. To test the IEL predictions, we ran our own experiments at the California Institute of Technology. Qualitatively, the data from those experiments are consistent with the IEL predictions about convergence and the U-shaped curve. Quantitatively, the human subjects are a little faster, a little less stable, and slightly less efficient than IEL. But for [gamma]Â =Â 50 and 100, the differences between humans and IEL are very small.