Model-Based RL Insights
Sergey Levine discusses the effectiveness of sequence models in handling long sequences and the small change in beam search to favor high rewards. He highlights the surprising success of the model as a predictive model and its ability to represent the whole joint distribution over states and actions. Additionally, he mentions the difficulty of performing a 100-step rollout with conventional models.In this clip
From this podcast
The Gradient
Sergey Levine on Robot Learning & Offline RL
Related Questions