Off-Policy Evaluation
A method in reinforcement learning used to estimate the performance of a new target policy using historical data collected by a different, older policy.
Think of It Like This
Like trying to predict how well a new basketball strategy will work by only watching tapes of last year's games played differently.
This is crucial in high-stakes environments like healthcare or robotics where deploying an untested policy to gather data is dangerous or expensive. It often utilizes techniques like importance sampling to re-weight the historical data. However, variance can explode if the target policy behaves radically differently from the logging policy.