On the Statistical Efficiency of Reward-Free Exploration in Non-Linear RL
This work addresses sample efficiency in RL for researchers, offering both positive algorithmic results and negative hardness insights, though it is incremental in building on prior settings like linear MDPs and low-rank MDPs.
The paper tackles reward-free reinforcement learning under non-linear function approximation, proposing the RFOLIVE algorithm for sample-efficient exploration under minimal structural assumptions and showing that previous explorability assumptions are statistically unnecessary, while also providing a statistical hardness result indicating an exponential separation between low-rank and linear completeness settings.
We study reward-free reinforcement learning (RL) under general non-linear function approximation, and establish sample efficiency and hardness results under various standard structural assumptions. On the positive side, we propose the RFOLIVE (Reward-Free OLIVE) algorithm for sample-efficient reward-free exploration under minimal structural assumptions, which covers the previously studied settings of linear MDPs (Jin et al., 2020b), linear completeness (Zanette et al., 2020b) and low-rank MDPs with unknown representation (Modi et al., 2021). Our analyses indicate that the explorability or reachability assumptions, previously made for the latter two settings, are not necessary statistically for reward-free exploration. On the negative side, we provide a statistical hardness result for both reward-free and reward-aware exploration under linear completeness assumptions when the underlying features are unknown, showing an exponential separation between low-rank and linear completeness settings.