Solve Problems using Reinforcement Learning 强化学习

Reinforcement Learning is used in ChatGPT, inventory management software, warehouse robots, and translation tools.

This exciting field in machine learning solves sequential decision-making problems, i.e. take a sequence of decisions to maximize performance.

This week, we'll train a simple lunar lander designed ourselves. Before the meeting:

https://colab.research.google.com/drive/1J9N6bbtvs8ExuTe6P48_MhCzdSvJcTbi#forceEdit=true&sandboxMode=true

Click to open the notebook. Then click 'File' and select 'Save a copy in Drive' from the dropdown menu.

Please run the copied notebook cell by cell as well as read Chapter 9.

You'll benefit from the discussion if you run it in advance.

Resources

Although their chapters are arranged differently, BOOK 1 and BOOK 2 (free online) cover the same content. Feel free to read either one.

BOOK 1:

https://learning.oreilly.com/library/view/the-reinforcement-learning/9781800200456/

Github:

https://github.com/PacktWorkshops/The-Reinforcement-Learning-Workshop

BOOK 2 (free online):

http://www.incompleteideas.net/book/RLbook2020.pdf

Reinforcement Learning: An Introduction - Sutton and Barto (2018) is the most cited RL book.

It has three parts:

The first part (Chapters 2–8) elaborates the tabular case which has exact solutions.

The second part (Chapters 9–13) talks about function approximation.

The third part (Chapters 14–16) focuses on RL's relationships to psychology, neuroscience, and case-studies (Atari game playing, Watson’s wagering strategy, and the Go playing programs) respectively.