Supervised Off-Policy Ranking

Off-policy evaluation (OPE) leverages data generated by other policies to evaluate a target policy. Previous OPE methods mainly focus on precisely estimating the true performance of a policy. We observe that in many applications, (1) the end goal of OPE is to compare two or multiple candidate policies and choose a good one, which is actually a much simpler task than evaluating their true performance; and (2) there are usually multiple policies that have been deployed in real-world systems and thus whose true performance is known through serving real users. Inspired by the two observations, in this work, we define a new problem, supervised off-policy ranking (SOPR), which aims to rank a set of new/target policies based on supervised learning by leveraging off-policy data and policies with known performance. We further propose a method for supervised off-policy ranking that learns a policy scoring model by correctly ranking training policies with known performance rather than estimating their precise performance. Our method leverages logged states and policies to learn a Transformer based model that maps offline interaction data including logged states and the actions taken by a target policy on these states to a score. Experiments on different games, datasets, training policy sets, and test policy sets show that our method outperforms strong baseline OPE methods in terms of both rank correlation and performance gap between the truly best and the best of the ranked top three policies. Furthermore, our method is more stable than baseline methods.