Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning

Presented by Haoxin Lin • Model-based Reinforcement Learning (MBRL) can reduce the reliance of policy optimization on realworld data -learning a dynamics model 𝑇(𝑠 ! , 𝑟|𝑠, 𝑎) from real-world data -facilitating policy exploration within the learned dynamics model