Efficient multi-agent experimentation and multi-choice bandits

The first best in the multi-agent experimentation problem of Klein and Rady (2011) can be reinterpreted as a multi-choice bandit problem where a researcher experiments on up to two projects simultaneously. We exploit this analogy to characterize the optimal strategy: If the cost or the discount rate are low enough, after working on a single project unsuccessfully, the researcher takes on both projects at once. When the “bad project†can also yield successes, as in Keller and Rady (2010), we provide an example where the researcher delays taking on the second project, and we show that she can give up earlier in particular if false positives are arbitrarily unlikely.