[4Xin2] Poster session 2

[4Xin2-23] A Study of Off-Policy Evaluation in Counterfactual Machine Learning

〇Mariko Sugimura¹, Ichiro Kobayashi¹ (1.Ochanomizu University)

Keywords:Counterfactual Machine Learning, Off-Policy Evaluation

Off-policy evaluation (OPE) is one of the research areas of Counterfactual Machine Learning (CFML), which aims to evaluate virtual policies using log data collected by another policy in the past. This makes it possible to evaluate new policies or obtain better policies without risky and costly online experiments. Although various OPE methods have been proposed, the accuracy with which each OPE method evaluates policies varies depending on the experimental environment, so it is not possible to measure the performance of a method in only one experimental environment. Therefore, in this study, we conducted experiments on various OPE methods in multiple environments using COBS, a benchmark suite for OPE, and confirmed the difference in performance of OPE in each method and environment. The results also led to a reconsideration of the compatibility between the methods and the environments.