New Algorithms Optimize Contextual Slate Bandits with Limited Adaptivity

This paper explores the contextual slate bandit problem, a scenario where a learner must select a slate of items from multiple sets to maximize rewards, with rewards following a Generalized Linear Model (GLM). The focus is on settings with limited adaptivity, meaning the system cannot update its policy continuously. The researchers propose two novel algorithms: B-SlateGLinCB for batched adaptivity and RS-SlateGLinCB for rarely-switching adaptivity.B-SlateGLinCB divides the learning horizon into a logarithmic number of batches, updating its policy only between batches. RS-SlateGLinCB, on the other hand, performs a limited number of parameter updates throughout the process. Both algorithms demonstrate impressive regret bounds that are independent of the GLM's non-linearity parameter, a common scaling factor in other GLM bandit algorithms.Crucially, these algorithms are computationally efficient, requiring only polynomial time per round despite the exponential number of possible slates. Simulations show they outperform existing limited-adaptivity baselines and remain competitive with fully adaptive state-of-the-art methods. Their strong performance was also validated in a practical in-context example selection task for language models, highlighting their real-world applicability.