Sample complexity of variance-reduced policy gradient: weaker assumptions and lower bounds

Several variance-reduced versions of REINFORCE based on importance sampling achieve an improved \(O(\epsilon ^{-3})\) sample complexity to find an \(\epsilon\)-stationary point, under an unrealistic assumption on the variance of the importance weights. In this paper, we propose the Defensive Policy Gradient (DEF-PG) algorithm, based on defensive importance sampling, achieving the same result without any assumption on the variance of the importance weights. We also show that this is not improvable by establishing a matching \(\Omega (\epsilon ^{-3})\) lower bound, and that REINFORCE with its \(O(\epsilon ^{-4})\) sample complexity is actually optimal under weaker assumptions on the policy class. Numerical simulations show promising results for the proposed technique compared to similar algorithms based on vanilla importance sampling.