Bandit problems with Levy payoff processes

Abstract

We study two-armed Levy bandits in continuous-time, which have one safe arm that yields a constant payoff s, and one risky arm that can be either of type High or Low; both types yield stochastic payoffs generated by a Levy process. The expectation of the Levy process when the arm is High is greater than s, and lower than s if the arm is Low. The decision maker (DM) has to choose, at any given time t, the fraction of resource to be allocated to each arm over the time interval [t,t+dt). We show that under proper conditions on the Levy processes, there is a unique optimal strategy, which is a cut-off strategy, and we provide an explicit formula for the cut-off and the optimal payoff, as a function of the data of the problem. We also examine the case where the DM has incorrect prior over the type of the risky arm, and we calculate the expected payoff gained by a DM who plays the optimal strategy that corresponds to the incorrect prior. In addition, we study two applications of the results: (a) we show how to price information in two-armed Levy bandit problem, and (b) we investigate who fares better in two-armed bandit problems: an optimist who assigns to High a probability higher than the true probability, or a pessimist who assigns to High a probability lower than the true probability.