Reinforcement Learning, Part 5: Overcoming the Practical Challenges of Reinforcement Learning Video - MATLAB

The first four videos in this series covered how great reinforcement learning is and how you can use it to solve some really hard control problems. So you may have this idea that you can essentially set up an environment, place an RL agent in it, and then let the computer solve your problem while you go off and drink a coffee or something. Unfortunately, even if you set up a perfect agent and a perfect environment and then the learning algorithm converges on a solution, there are still drawbacks to this method that we need to talk about. So in this video, I'm going to address a few possibly non-obvious problems with RL and try to provide some ways to mitigate them. Even if there aren't straightforward ways to address some of the challenges that you'll face, at the very least it'll get you thinking about them. The problems that we'll address in this video come down to two main questions: The first is once you have a learned policy, is there a way to manually adjust it if it's not quite perfect?