Deep RL Bootcamp Lecture 9 Model-based Reinforcement

 

 

 

 

 

 

 

 

 

 

So, the process is similar to one-to-many RNN? 

 

 learn much more efficiently than model-free method

 

 

 

 

 

iteratively get better

 

 

 

less than 300 trials ~ 25min robot time  per task

 

 

 

 

 

 

 

 

 

 

 visual prediction from the observation

 

 

 

 during train of model, there is no reward. Some random motions are programmed. at the task time, there is a reward function, basically trying to move a pixel to  the goal position.