Evaluating resources for teaching myself reinforcement learning

Question

While trying to teach myself the topic of reinforcement learning, I find many sites, books, people and other resources with information about that subject. (I am trying to learn the subject from those resources).

However, some resources contradict each other or conflict with one another. While sometimes it's easy to see what isn't correct (when there is a contradiction, there's often one version that's more accurate, for example), sometimes it's not so easy. Then there are times when there are simply different resources for the same topic in the field, and I cannot decide from which it might be better to learn reinforcement learning.

I'd like to be able to judge or evaluate which resources about reinforcement learning could be better to learn from.

I am aware that it's near impossible to say "this is the best one", but what I'm looking for is a method for choosing the resource from which I can learn the most. This means that in an introduction, resources which go into very fine details are not as good as ones that gives an overview of the subject.

In a way, I'm asking how can I build a curriculum for myself based entirely on resources that deal with reinforcement learning (books, sites, people etc.) to guide my learning of a subject of which I know very little? Judging one resource over another is one of the only things which are making the learning difficult.

This is relevant to reinforcement learning specifically because it's a quickly changing field. New practices/techniques and resources are being generated very rapidly.

---

I can see that the question itself is very long, but that's meant to keep it focused on what my problem really is: some resources are better than others when it comes to learning a new subject in a quickly evolving field...

Answer 1

If self-learning is a goal, you don't need to necessarily choose between resources at all. I had to self-teach a tremendous amount of material this year, and to do so, I roughly borrowed the methodology from Barry Farber's incredible How To Learn Any Language.

The basic idea is to grab multiple (quality) resources of different kinds, and begin to work through all of them. Do a little bit of each one every day, and make all of these activities as active as possible. (For instance, create flashcards of concepts you want to review. While reading physical books, don't be afraid to underline things you don't understand and revisit older underlines.)

Every content creator attempts to get at the material in a way that they feel emphasizes what is most important. This means that you end up learning multiple central conceits, and you develop a deep and capable understanding relatively quickly. If you find yourself getting stuck on one source, do your underlining, and move on to the next one. The next day, you can revisit the first source with the new understandings that you have gained from the other resources, and thereby continue to progress.

Answer 2

As a self-learner you have the extra burdens of resource selection and course design. You can lighten the load by sharing in the work of others. Find universities doing courses and research in the subject and review their course materials and source selections. For the cutting edge things, find papers that have been subjected to peer review, and are sufficiently current to maintain their relevance to current methods.

While I support open data and open science, this may be a case where you need to access the material behind a pay-wall, and the expense might be justified.

If, after all methods have failed to help you select the best resources you might have simply attempt to implement the different suggestions and evaluate them for correctness, completeness and relevance to your situation.

Answer 3

If you want to start with a single source, I would suggest the book, "Reinforcement Learning: An Introduction" by Sutton and Barto. The authors have an established history and a track record for publishing peer-reviewed articles in the field. While this book is dated, published in 1998, I have not found a better introduction to the field yet. Before you can understand the latest developments in the filed of RL, you need to understand the basics and be able to distinguish between various approaches. Jumping in right away to the latest publications in the field would likely be very difficult to follow. I acknowledge that I likely have bias, having used this book as a Ph.D. student, then using it later as a professor.