21
89

Model selection for contextual bandits

Abstract

We introduce the problem of model selection for contextual bandits, where a learner must adapt to the complexity of the optimal policy while balancing exploration and exploitation. Our main result is a new model selection guarantee for linear contextual bandits. We work in the stochastic realizable setting with a sequence of nested linear policy classes of dimension d1<d2<d_1 < d_2 < \ldots, where the mm^\star-th class contains the optimal policy, and we design an algorithm that achieves O~(T2/3dm1/3)\tilde{O}(T^{2/3}d^{1/3}_{m^\star}) regret with no prior knowledge of the optimal dimension dmd_{m^\star}. The algorithm also achieves regret O~(T3/4+Tdm)\tilde{O}(T^{3/4} + \sqrt{Td_{m^\star}}), which is optimal for dmTd_{m^{\star}}\geq{}\sqrt{T}. This is the first model selection result for contextual bandits with non-vacuous regret for all values of dmd_{m^\star}, and to the best of our knowledge is the first positive result of this type for any online learning setting with partial information. The core of the algorithm is a new estimator for the gap in the best loss achievable by two linear policy classes, which we show admits a convergence rate faster than the rate required to learn the parameters for either class.

View on arXiv
Comments on this paper