Advanced Model Learning
This lecture of Berkeley’s deep rl course covers more advanced techniques in advanced model learning (especially when it comes to applying this to more realistic cases).
In Q-Learning, as Deepmind’s atari research in 2013 shows, it’s possible to take in large images as states and produce a reliable reinforcement learning method. However, for something like images, it often becomes very difficult to create an accurate model because
1) the images are too large and 2) the model is a simple observation.
This lecture covers state of the art research (including a lot of work done by Chelsea Finn, who gave this guest lecture) in representing our model. They include
1) using an encoder 2) models in the image space 3) inverse modelling and 4) alternative quantitiy modelling
Overall, this was just a great lecture, and really highlights how active an area of research reinforcement learning is.
My notes can be found below: