Hacker News new | past | comments | ask | show | jobs | submit login
Reinforcement Learning and Optimal Control (mit.edu)
213 points by iron0013 61 days ago | hide | past | web | favorite | 12 comments

If you're looking for a great lecture course, I highly recommend CS 294. It's taught by Sergey Levine, who is in my opinion one of the foundational researchers at the frontier of deep reinforcement learning, and provides a comprehensive overview on the current state of the art: http://rail.eecs.berkeley.edu/deeprlcourse/

I took the time to look through two of the lectures from the slides available. This is so inspirational. Thanks a lot!

A link to all the lecture videos for Fall 2018 is available here if you'd like to listen to the actual recordings:


Any idea how it compares to the UCL course taught by researchers from DeepMind?


The Berkeley course is completely open source with all the lectures, homework problem sets, and exams publicly available on the site. Besides getting a grade, everything is available to you.

Unfortunately I'm not sure about how the UCL course compares as it seems like the materials are closed to non-UCL students. If you have access though, I'm sure there's lots to learn by taking the course!

Not meant as a critique to this book, but the fact that RL-based approaches rarely work for optimal control problems (in, for example, robotics) came as a surprise to me, given the hype and focus on RL [1]. It turns out that model-based methods for optimal control (e.g. linear quadratic control) invented quite a long time ago dramatically outperform RL-based approaches in most tasks and require multiple orders of magnitude less computational resources. Maybe there's some hope for RL method if they "course correct" for simpler control methods. At the moment, it seems like RL for robotics and control lies to the side of "research" and not "engineering."

[1]: https://www.alexirpan.com/2018/02/14/rl-hard.html

for linear systems modern control theory is already optimal (?) so given a linear system I'd assume an RL method would just approximate the optimal control method. I think the potential for RL methods is in capturing nonlinearities in actuators like artificial muscle.

I meant applying a linear approximation.

Bertsekas' earlier books (Dynamic Programming and Optimal Control + Neurodynamic Programming w/ Tsitsiklis) are great references and collect many insights & results that you'd otherwise have to trawl the literature for. I'm very interested to see what a book focused more narrowly on RL will be like-- Sutton's Introduction to Reinforcement Learning[0] is fantastic, but if you're going to do research on RL, another text such as this one is necessary.

0. http://www.incompleteideas.net/book/the-book-2nd.html

Seems like an interesting complement to the (free) book on planning algorithms by Steven Lavalle: http://planning.cs.uiuc.edu/

I liked the open courseware lectures with John Tsitsiklis and ended up with a few books by Bertsekas: neuro dynamic programming and intro probability. Both Bertsekas and Tsitsiklis recommended the Sutton and Barto intro book for an intuitive overview. I liked it.

This inspired me so much. Thank you for sharing!

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact