This is an example of Deep Q Network reinformcement learning in a simple game environment. The goal is to pick up as many of the yellow bananas as possible. Yellow bananas are rewarded positively, while Blue bananas are penalized.
Some pre-trained model weights are provided. A GPU is preferred for training.
For more information, please see the Introduction to the original github project [here] ( https://github.com/udacity/deep-reinforcement-learning/tree/master/p1_navigation ).
This is an example of Deep Q Network reinformcement learning in a simple game environment. The goal is to pick up as many of the yellow bananas as possible. Yellow bananas are rewarded positively, while Blue bananas are penalized.
The environment is provided as a custom Unity project from Udacity, found here in the Banana_Linux folder.
The original Linux environment can be downloaded [here] (https://s3-us-west-1.amazonaws.com/udacity-drlnd/P1/Banana/Banana_Linux.zip).
The environment has 37 states. Four actions are available: forward, backward, left and right.
Two improvements to the basic Q-Learning algorithm are implemented: experience replay, and fixed Q-targets.
| File | Description |
|---|---|
| navigation_project.ipynb | Main ipython file |
| Banana_Linux | Unity code repository |
| dqn_agent.py | Learning agent |
| model.py | PyTorch network model |
| checkpoint.pth | Model weights |
| install.sh | Installation |
| requirements.txt | Necessary modules |
| report.md | Report on the project |
| training.png | Image |
| README.md | This file |
- Clone the assocated repository
- Run the install.sh file
source install.sh jupyter-notebook navigation_project.ipynb- Follow the instructions in the notebook