Timezone: »
We consider the problem of off-policy evaluation—estimating the value of a target policy using data collected by another policy—under the contextual bandit model. We establish a minimax lower bound on the mean squared error (MSE), and show that it is matched up to constant factors by the inverse propensity scoring (IPS) estimator. Since in the multi-armed bandit problem the IPS is suboptimal, our result highlights the difficulty of the contextual setting with non-degenerate context distributions. We further consider improvements on this minimax MSE bound, given access to a reward model. We show that the existing doubly robust approach, which utilizes such a reward model, may continue to suffer from high variance even when the reward model is perfect. We propose a new estimator called SWITCH which more effectively uses the reward model and achieves a superior bias-variance tradeoff compared with prior work. We prove an upper bound on its MSE and demonstrate its benefits empirically on a diverse collection of datasets, often seeing orders of magnitude improvements over a number of baselines.
Author Information
Yu-Xiang Wang (UC Santa Barbara)
More from the Same Authors
-
2021 Spotlight: Logarithmic Regret in Feature-based Dynamic Pricing »
Jianyu Xu · Yu-Xiang Wang -
2021 Poster: Privately Publishable Per-instance Privacy »
Rachel Redberg · Yu-Xiang Wang -
2021 Poster: Logarithmic Regret in Feature-based Dynamic Pricing »
Jianyu Xu · Yu-Xiang Wang -
2021 Poster: Optimal Uniform OPE and Model-based Offline Reinforcement Learning in Time-Homogeneous, Reward-Free and Task-Agnostic Settings »
Ming Yin · Yu-Xiang Wang -
2021 Poster: Towards Instance-Optimal Offline Reinforcement Learning with Pessimism »
Ming Yin · Yu-Xiang Wang -
2021 Poster: Near-Optimal Offline Reinforcement Learning via Double Variance Reduction »
Ming Yin · Yu Bai · Yu-Xiang Wang -
2017 Poster: Higher-Order Total Variation Classes on Grids: Minimax Theory and Trend Filtering Methods »
Veeranjaneyulu Sadhanala · Yu-Xiang Wang · James Sharpnack · Ryan Tibshirani -
2016 Poster: Total Variation Classes Beyond 1d: Minimax Rates, and the Limitations of Linear Smoothers »
Veeranjaneyulu Sadhanala · Yu-Xiang Wang · Ryan Tibshirani -
2015 : Yu-Xiang Wang: Learning with differential privacy: stability, learnability and the sufficiency and necessity of ERM principle »
Yu-Xiang Wang -
2015 Poster: Differentially private subspace clustering »
Yining Wang · Yu-Xiang Wang · Aarti Singh