Landing Space Shuttle using Deep Q-Learning

This was my first attempt to learn reinforcement learning. The agent was trained to land a space shuttle in a randomly generated environment.

I started by learning Q-Learning through a project on solving Black Jack. However, because of the finite dimension limitation, I decided to use Deep Q-Learning, which combined the original method with a neural network. This approach showed promising results.

As shown above, after about 200 episodes, the agent learned how to land the shuttle reliably. Additionally, since I added the punishment for using too much fuel, the agent also learned to land the shuttle more confidently (200 frames vs. 900 frames).

For this project, I had to learn how to use PyTorch and implement different reinforcement learning algorithms in two weeks.