Very good!

Thank you.

very good

Very good

Very nice

Fantastic

Perfect!

loved it

not bad

Great!

great!

Great!

great

Good!

Good!

good

nice

good

good

Nice

I enjoyed this introduction course in Reinforcement Learning (RL). It explained in detail the fundamentals of RL such as k-armed bandits, Contextual Bandits and - of course - Markov Decision Processes (MDP). The lectures explained the conceps with nice examples and as well as the math behind (Bellman equations). The coursebook was the great "RL bible" ("Reinforcement Learning - An Introduction", 2nd edition by Sutton & Bartto); the lectures followed the first 4 chapters of the book quite closely.

I liked the programming assignments. It took some time to understand the structure of the tools used (e.g. the little known RLGlue) but after that it was quite straight forward, especially since the Notebook had great support for testing the solutions before submitting the assignment.

It was also interesting to see the guest lectures talk about the world outside the simple example MDPs used as examples, such as RL in the real world (using Contextual Bandits as a foundation), and about solving huge Fleet Management problems with RL.

One thing I missed in this course was more details about MDP and linear programming, which was mentioned in passim by the lecturers, and was an essential tool for solving the Fleet Management Problem (using approximate linear programming). Perhaps some of the next courses will discuss linear programming more...

I thought that the course content was extremely interesting, and the tests and programming were informative.

I did think though that the lectures were a little terse and could have given more information and worked through more examples. I think the presenters of this course and the people who constructed it could learn a lot from how, say, Andrew Ing's Coursera courses and Geoffrey Hinton's Coursera courses are put together and presented.

Specifically, the actual video time was very short and huge dependence was placed on the text book (which is very good textbook). I found Jupyter note book buggy and had to reset it a few times, but that might be me: I am not familar with it. I think as well, in a preliminary section, there could have been more on the Jupyter notebook and programming - even if this was just a document. As a user inexperienced with the Jupyter notebook, I found debugging and running test code in the lecturer's notebook in order to find my errors really hard. I often had to reset the notebook. Some assistance would have been appreciated here. In other courses that I have done, the prgramming environment has been more flexible which has made debugging easier, but I accept that my concerns here may be due to my inexperience.

Hi,

I don't know whether this feedback will reach the correct ears or not.

I have already completed the course before and now I am doing it again. One thing that I found is the coding assignments are using library and is not letting the student do the thing from scratch. Things will be very clear to the student if the build everything from scratch using the basic libraries. for eg. not using rl_glue, but coding up the environment, coding up the agent. Using abstraction is good, but for those who already know the things. Since this course is more about the fundamentals of RL, it should teach the basics of building environment, agent from scratch. Maybe we can use library once we have done it from scratch, like starting from week 3 or course 2. I persnally was not able to get the full understanding of the things untill I implemented the things from scratch.

Thanks:)

overall course very nice. A great effort !

The course is useful only if you read the book simultaneously. The book describes things in a much better way than the teachers of the course. If you want to get a good understanding of the concepts, I suggest you first read the topics from the book and then listen to the lectures.

Pros:

1. Good quizzes to understand the concepts.

2. Programming exercises are a must to understand the ideas correctly.

Cons:

1. One complaint I have for the course is, it lacks mathematical rigor in its explanations. For example, the course and the book just say that value updates eventually converge to the value corresponding to the policy, but they never reason about why that must be the case. Even putting a heuristic argument or atleast pointing to a "proof of convergence" would have satisfied the mathematician in me.