We study the off-policy evaluation problem---estimating the value of a target policy using data collected by another policy---under the contextual bandit model. We consider the general (agnostic) setting without access to a consistent model of rewards and establish a minimax lower bound on the mean squared error (MSE). The bound is matched up to constants by the inverse propensity scoring (IPS) and doubly robust (DR) estimators. This highlights the difficulty of the agnostic contextual setting, in contrast with multi-armed bandits and contextual bandits with access to a consistent reward model, where IPS is suboptimal. We then propose the SWITCH estimator, which can use an existing reward model (not necessarily consistent) to achieve a better bias-variance tradeoff than IPS and DR. We prove an upper bound on its MSE and demonstrate its benefits empirically on a diverse collection of datasets, often outperforming prior work by orders of magnitude.
Summary/Notes »Author Information
Yu-Xiang Wang (Carnegie Mellon University / Amazon AWS)
Alekh Agarwal (Microsoft Research)
Miroslav Dudik (Microsoft Research)
Related Events (a corresponding poster, oral, or spotlight)
-
2017 Talk: Optimal and Adaptive Off-policy Evaluation in Contextual Bandits »
Tue Aug 8th 11:06 -- 11:24 AM Room C4.5
More from the Same Authors
-
2019 Poster: Warm-starting Contextual Bandits: Robustly Combining Supervised and Bandit Feedback »
Chicheng Zhang · Alekh Agarwal · Hal Daume · John Langford · Sahand Negahban -
2019 Poster: Fair Regression: Quantitative Definitions and Reduction-Based Algorithms »
Alekh Agarwal · Miroslav Dudik · Steven Wu -
2019 Oral: Warm-starting Contextual Bandits: Robustly Combining Supervised and Bandit Feedback »
Chicheng Zhang · Alekh Agarwal · Hal Daume · John Langford · Sahand Negahban -
2019 Oral: Fair Regression: Quantitative Definitions and Reduction-Based Algorithms »
Alekh Agarwal · Miroslav Dudik · Steven Wu -
2019 Poster: Provably efficient RL with Rich Observations via Latent State Decoding »
Simon Du · Akshay Krishnamurthy · Nan Jiang · Alekh Agarwal · Miroslav Dudik · John Langford -
2019 Oral: Provably efficient RL with Rich Observations via Latent State Decoding »
Simon Du · Akshay Krishnamurthy · Nan Jiang · Alekh Agarwal · Miroslav Dudik · John Langford -
2018 Poster: Hierarchical Imitation and Reinforcement Learning »
Hoang Le · Nan Jiang · Alekh Agarwal · Miroslav Dudik · Yisong Yue · Hal Daume -
2018 Poster: A Reductions Approach to Fair Classification »
Alekh Agarwal · Alina Beygelzimer · Miroslav Dudik · John Langford · Hanna Wallach -
2018 Oral: Hierarchical Imitation and Reinforcement Learning »
Hoang Le · Nan Jiang · Alekh Agarwal · Miroslav Dudik · Yisong Yue · Hal Daume -
2018 Oral: A Reductions Approach to Fair Classification »
Alekh Agarwal · Alina Beygelzimer · Miroslav Dudik · John Langford · Hanna Wallach -
2018 Poster: Practical Contextual Bandits with Regression Oracles »
Dylan Foster · Alekh Agarwal · Miroslav Dudik · Haipeng Luo · Robert Schapire -
2018 Oral: Practical Contextual Bandits with Regression Oracles »
Dylan Foster · Alekh Agarwal · Miroslav Dudik · Haipeng Luo · Robert Schapire -
2017 Poster: Contextual Decision Processes with low Bellman rank are PAC-Learnable »
Nan Jiang · Akshay Krishnamurthy · Alekh Agarwal · John Langford · Robert Schapire -
2017 Talk: Contextual Decision Processes with low Bellman rank are PAC-Learnable »
Nan Jiang · Akshay Krishnamurthy · Alekh Agarwal · John Langford · Robert Schapire -
2017 Poster: Active Learning for Cost-Sensitive Classification »
Akshay Krishnamurthy · Alekh Agarwal · Tzu-Kuo Huang · Hal Daumé III · John Langford -
2017 Talk: Active Learning for Cost-Sensitive Classification »
Akshay Krishnamurthy · Alekh Agarwal · Tzu-Kuo Huang · Hal Daumé III · John Langford -
2017 Tutorial: Real World Interactive Learning »
Alekh Agarwal · John Langford
