Leave it to the folks at Google to devise AI capable of predicting which machine learning models will produce the best results. In a newly-published paper ("Off-Policy Evaluation via Off-Policy Classification") and blog post, a team of Google AI researchers propose what they call "off-policy classification," or OPC, which evaluates the performance of AI-driven agents by treating evaluation as a classification problem. The team notes that their approach -- a variant of reinforcement learning, which employs rewards to drive software policies toward goals -- works with image inputs and scales to tasks including vision-based robotic grasping. "Fully off-policy reinforcement learning is a variant in which an agent learns entirely from older data, which is appealing because it enables model iteration without requiring a physical robot," writes Robotics at Google software engineer Alexa Irpan. "With fully off-policy RL, one can train several models on the same fixed dataset collected by previous agents, then select the best one."

In this work, we consider the problem of model selection for deep reinforcement learning (RL) in real-world environments. Typically, the performance of deep RL algorithms is evaluated via on-policy interactions with the target environment. However, comparing models in a real-world environment for the purposes of early stopping or hyperparameter tuning is costly and often practically infeasible. This leads us to examine off-policy policy evaluation (OPE) in such settings. We focus on OPE for value-based methods, which are of particular interest in deep RL, with applications like robotics, where off-policy algorithms based on Q-function estimation can often attain better sample complexity than direct policy optimization. Existing OPE metrics either rely on a model of the environment, or the use of importance sampling (IS) to correct for the data being off-policy. However, for high-dimensional observations, such as images, models of the environment can be difficult to fit and value-based methods can make IS hard to use or even ill-conditioned, especially when dealing with continuous action spaces. In this paper, we focus on the specific case of MDPs with continuous action spaces and sparse binary rewards, which is representative of many important real-world applications. We propose an alternative metric that relies on neither models nor IS, by framing OPE as a positive-unlabeled (PU) classification problem with the Q-function as the decision function. We experimentally show that this metric outperforms baselines on a number of tasks. Most importantly, it can reliably predict the relative performance of different policies in a number of generalization scenarios, including the transfer to the real-world of policies trained in simulation for an image-based robotic manipulation task.