Adapting to Misspecification in Contextual Bandits
share
A major research direction in contextual bandits is to develop algorithms that are computationally efficient, yet support flexible, general-purpose function approximation. Algorithms based on modeling rewards have shown strong empirical performance, but typically require a well-specified model, and can fail when this assumption does not hold. Can we design algorithms that are efficient and flexible, yet degrade gracefully in the face of model misspecification? We introduce a new family of oracle-efficient algorithms for ε-misspecified contextual bandits that adapt to unknown model misspecification – both for finite and infinite action settings. Given access to an online oracle for square loss regression, our algorithm attains optimal regret and – in particular – optimal dependence on the misspecification level, with no prior knowledge. Specializing to linear contextual bandits with infinite actions in d dimensions, we obtain the first algorithm that achieves the optimal O(d√(T) + ε√(d)T) regret bound for unknown misspecification level ε. On a conceptual level, our results are enabled by a new optimization-based perspective on the regression oracle reduction framework of Foster and Rakhlin, which we anticipate will find broader use.
READ FULL TEXT