This is a capstone project for the reinforcement learning specialization by the University of Alberta which provides some of the utility code. The environment is provided by OpenAI Gym.
The design of the reinforcement system is in RL_system.py, and training is done in RL_system_training.ipynb.
This project trains a reinforcement learning agent to successfully land a 'lunar lander' without crashing. It uses a neural network as a function approximator to the action-values and selects actions according to a softmax policy. Training is done using Expected SARSA with the Adam optimizer. Experience replay is used to improve sample efficiency.
The above GIF is a recording from RL_system_training.ipynb of the RL agent running in the Lunar Lander Environment. It's able to reach it's goal, but can probably be improved if trained for longer!