Online Optimization in X-Armed Bandits

We consider a generalization of stochastic bandit problems where the set of arms, Χ, is allowed to be a generic topological space. We constraint the mean-payoff function with a dissimilarity function over Χ in a way that is more general than Lipschitz. We construct an arm selection policy whose regret improves upon previous result for a large class of problems. In particular, our results imply that if Χ is the unit hypercube in a Euclidean space and the mean-payoff function has a finite number of global maxima around which the behavior of the function is locally Holder with a known exponent, then the expected regret is bounded up to a logarithmic factor by √n, i.e., the rate of the growth of the regret is independent of the dimension of the space. Moreover, we prove the minimax optimality of our algorithm for the class of mean-payoff functions we consider.