Oracle-Efficient Pessimism: Offline Policy Optimization in Contextual Bandits
This work addresses computational inefficiency in offline policy optimization for contextual bandits, which is an incremental improvement for researchers and practitioners in reinforcement learning.
The paper tackles the problem of offline policy optimization in contextual bandits by developing a general oracle-efficient algorithm that reduces to supervised learning, achieving statistical guarantees comparable to prior pessimistic approaches and showing experimental advantages over unregularized methods across various configurations.
We consider offline policy optimization (OPO) in contextual bandits, where one is given a fixed dataset of logged interactions. While pessimistic regularizers are typically used to mitigate distribution shift, prior implementations thereof are either specialized or computationally inefficient. We present the first general oracle-efficient algorithm for pessimistic OPO: it reduces to supervised learning, leading to broad applicability. We obtain statistical guarantees analogous to those for prior pessimistic approaches. We instantiate our approach for both discrete and continuous actions and perform experiments in both settings, showing advantage over unregularized OPO across a wide range of configurations.