Deep Q-Learning agent learns how to navigate a world full of bananas.
The agent in this project uses the following RL techniques:
- Prioritized Experience Replay
- Double DQN (DDQN)
- Dueling DQN
To be successful in this environment, the RL agent must learn to collect yellow bananas (reward +1
) while avoiding blue bananas (reward -1
).
The state has a size of 37
and contains information such as the agent's velocity and ray-based perception of objects in front of the agent.
Action space:
0
: Forward1
: Backward2
: Turn left3
: Turn right
The environment is solved if the agent achieves an average score of +13
over 100 consecutive episodes.
Follow the instructions here to:
- Create a
conda
environment. - Clone the Udacity Deep RL repository.
- Install Python packages into the environment.
- Create an IPython kernel using the environment.
The OpenAI Gym instructions can be skipped.
In order to watch the agent play the game, you also need to download the environment by following the instructions here.
Once you've completed the setup, you can:
- Open
Navigation.ipynb
. - Select the kernel created during setup.
- Run all the cells in the notebook to train the agent.
Follow the instructions here, load the saved neural network weights (checkpoint.pth
), and watch the trained agent play the game!