Behavior Policy Gradient Supplemental Material

摘要: We first derive an analytic expression for the gradient of the variance of an arbitrary, unbiased off-policy policy evaluation estimator, OPE (H, θ). Importance-sampling is one …