Strategic Experimentation with Unobservable Outcomes and Reversible Actions (Seminar-At-A-Distance)

Joint work with Jerome Renault and Nicolas Vieille.

 

Abstract:

We study a model of strategic experimentation. Each of two players faces a bandit machine with one safe arm that always produces 0 and one risky arm, whose use is costly, that either produces always 0 (Bad arm) or produces a prize m with probability lambda whenever used (Good arm). The types of the risky machines of the two players are the same. Players choose at each stage which arm to use. In this project we study properties of the equilibrium if this game. In particular we show that if lambda > 1/2, in all equilibria, there is at most one experimentation that is done when the belief of the player who does it is below the one-agent cut-off.