Where to start learning RL

Original article was published on Artificial Intelligence on Medium

I would say I am competent in robot learning (robotics + reinforcement learning). I had the privilege of being pushed to this in my Ph.D., but you can too. The themes are repeatable and effective.

Learn by doing

We learn new skills in computer science to make things. Taking ideas to code is where value is made (except for a few theoretical professors). The crux of this post is that you need to find your problem space.

There is a long list of resources to learn about RL theory at the end of this, but with how broadly applicable AI methods are — you have to choose where. This comes down to an overlay of three motives:

  1. Problems you enjoy working on.
  2. Problems that have global impact.
  3. Problems that will get you a job and stability.

Decide on a problem space for RL where you like what you’re doing, it’ll do something to help the world, and hopefully, other people will catch on and give you a bigger platform to make change.

An example simulated robot arm task — called Reacher3d. Uses Mujoco and Gym.

What have I built? I work with robots. I want robots to do many simple tasks, all over the place. They can move our furniture, drive our cars, deliver our boxes, and more. All this should come within a decade. A decade out, this looks like learning low-level locomotion controllers. The core repository for learning robot dynamics and control is found here. (Most of the research is still on private before publication.)

Build fundamentals and depth with writing or reflection

I’ve written about 20 posts on Medium, and it’s an amazing compliment to any education program. It’s time to reflect on what you built and how it fits into a bigger picture. It’s a time to make sure others can comprehend your results. A common weakness of the best graduate students I meet — an inability to clearly break down their ideas. As a senior graduate student, I am focused on making my work last and be reused after I finish my degree.

Research papers, blog posts, etc are all writing forms that act as permanent recreations of your mind and self. There’s little that lets individuals continue to be of use and interacted with after their career, but high quality writing may be the most accessible tool we have for now.

Posts I have written on RL to date. It’s a wonderful subject and there’s always more to explore.

  1. 3 skills to master before RL.
  2. What is a Markov Decision Process anyways?
  3. The hidden linear algebra of reinforcement learning.
  4. Fundamentals iterative methods of reinforcement learning.
  5. Convergence of reinforcement learning algorithms

Learn PyTorch

PyTorch is becoming dominant in the are of machine learning research, and because reinforcement learning is young, it’s mostly research. You can find the statistics here. PyTorch is very fluid and pythonic, so don’t worry about getting too bogged down in learning it, it can happen along the way.