Applying the DQN-Reinforcement Learning Algorithm to Gym's Lunar Lander environment. Train an agent to safely land a Lunar Lander on the moon without crashing.
The original paper explaining the Algorithm can be found here: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf
What things you need to install the software and how to install them
OpenAIs Gym toolbox
pip install gym
PyTorch
pip install torchvision
Run the Notebook "Deep-Q-Network.ipynb" to train your agent. The notebook provides an comparison of the agents ability to land the lunar lander before and after training. You should see strong improvements. Optinonally you can skip the training procedure and jump ahead to point 4. in the notebook to load a pretrained model and watch the trained agents performance.
Tweak several Hyperparameters in the model.py (Hidden Layer size etc.) to see if you can achieve performance gains. Also you can play around with the exploration parameter epsilon and ReplayBuffer size in the dqn_agent.py
The project is part of Udacities Nanoprogram: Deep Reinforcement Learning
https://eu.udacity.com/course/deep-reinforcement-learning-nanodegree--nd893