Jumping to Conclusions

Résumé

This talk presents a continuous-time game of strategic experimentation with two-armed bandits in which payoffs on the risky arm are generated by a compound Poisson process. The distribution of payoff increments depends on an unkown state of the world, whereas the arrival rate of such increments does not. As a consequence, players learn exclusively from the size of the lump-sum payoffs they obtain. Under perfect monitoring, the piecewise constancy of beliefs allows for an analysis of stationary Markov equilibria with essentially discrete-time techniques. In turn, these Markov equilibria constitute perfect Bayesian equilibria when players observe all payoffs, but not each other’s actions. Strongly symmetric equilibria cannot improve on the symmetric Markovian one. Whether asymmetric equilibria can achieve higher payoffs than Markovian equilibria is the focus of ongoing work. (with Johannes Hörner (Yale University), Nicolas Klein (University Paris-Panthéon-Assas))