We introduce a novel stochastic contextual bandit model, where at each step the adversary chooses a distribution over a context set. The learner observes only the context distribution while the exact context realization remains hidden. This allows for a broader range of applications, for instance when the context itself is based on predictions. By leveraging the UCB algorithm to this setting, we propose an algorithm that achieves a $\tilde{\mathcal{O}}(d\sqrt{T})$ high-probability regret bound for linearly parametrized reward functions. Our results strictly generalize previous work in the sense that both our model and the algorithm reduce to the standard setting when the environment chooses only Dirac delta distributions and therefore provides the exact context to the learner. We further obtain similar results for a variant where the learner observes the realized context after choosing the action, and we extend the results to the kernelized setting. Finally, we demonstrate the proposed method on synthetic and real-world datasets.

@inproceedings{kirschner2019context}(d\sqrt{T})$ high-probability regret bound for linearly parametrized reward functions. Our results strictly generalize previous work in the sense that both our model and the algorithm reduce to the standard setting when the environment chooses only Dirac delta distributions and therefore provides the exact context to the learner. We further obtain similar results for a variant where the learner observes the realized context after choosing the action, and we extend the results to the kernelized setting. Finally, we demonstrate the proposed method on synthetic and real-world datasets.},
	author = {Kirschner, Johannes and Krause, Andreas},
	booktitle = {Proc. Neural Information Processing Systems (NeurIPS)},
	month = {December},
	title = {Stochastic Bandits with Context Distributions},
	year = {2019}}