Official Review · NEURIPS 2023

This paper proposes a new reinforcement learning algorithm based on the replicability crisis and gives proof for the proposed method, providing a new perspective in this field. 1. The idea of replicable reinforcement learning is brand new and may provide a new perspective for reinforcement learning. 2. The proof of the related lemma is sufficient and strict. 1. The experiments are insufficient. Though the proposed replicable reinforcement learning requires much fewer samples, the performance of the replicable RL lacks. 2. More comparable experiments (i.e. DDPG, TD3, A3C, SAC, etc.) should be carried out to show the effectiveness of the proposed method. 3. It is better to provide some insightful analysis or conclusion from the replicable RL. 4. The organization may lack cohesion and coherence. Though several definitions and lemma are introduced, the relationship and function of each lemma lack a detailed claim. 5. The proposed replicable RL should be evaluated in the public benchmark such as gym， MuJoCo, etc. Please refer to the Weaknesses.