Optimal Mixture Weights for Off-Policy Evaluation with Multiple Behavior Policies

Abstract

Off-policy evaluation is a key component of reinforcement learning which evaluates a target policy with offline data collected from behavior policies. It is a crucial step towards safe reinforcement learning and has been used in advertisement, recommender systems and many other applications. In these applications, sometimes the offline data is collected from multiple behavior policies. Previous works regard data from different behavior policies equally. Nevertheless, some behavior policies are better at producing good estimators while others are not. This paper starts with discussing how to correctly mix estimators produced by different behavior policies. We propose three ways to reduce the variance of the mixture estimator when all sub-estimators are unbiased or asymptotically unbiased. Furthermore, experiments on simulated recommender systems show that our methods are effective in reducing the Mean-Square Error of estimation.

arXiv commentOffline Reinforcement Learning Workshop at Neural Information Processing Systems, 2020

References in corpus (9)

• Scikit-learn: Machine Learning in Python
• Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation Algorithms
• Breaking the Curse of Horizon: Infinite-Horizon Off-Policy Estimation
• RecSim: A Configurable Simulation Platform for Recommender Systems
• More Efficient Off-Policy Evaluation through Regularized Targeted Learning
• Doubly robust off-policy evaluation with shrinkage
• Black-box Off-policy Estimation for Infinite-Horizon Reinforcement Learning
• Adaptive Estimator Selection for Off-Policy Evaluation
• Triply Robust Off-Policy Evaluation