Learning and Implementation on the Internet

We address the problem of learning and implementation on the Internet When agents play repeated games in distributed environments like the Internet they have very limited a priori information about the other players and the payo matrix and the play can be highly asynchronous Consequently standard solution concepts like Nash equilibria or even the serially undominated set do not apply in such a set ting To construct more appropriate solution concepts we rst describe the essential properties that constitute reasonable learning behavior in distributed environments We then study the convergence behavior of such algorithms these results lead us to propose rather non traditional solutions concepts for this context Finally we discuss implementation of social choice functions with these solution concepts We would like to thank Roger Klein and Herv e Moulin for useful discussions and seminar participants at Princeton Rutgers Stonybrook and UBC for helpful comments This research was supported by NSF Grant ANI Email friedman econ rutgers edu shenker parc xerox com