Asynchronous Advantage Actor-Critic in Tensorflow 2.0 w/ Keras
A clean implementation of A3C focused on readability. The goal was to make everything as close as possible to pseudocode while also illustrating important aspects of implementation that are glossed over in the theory. Some of these aspects include:
- How to share gradient updates across asynchronous agents (using the Hogwild! approach)
- How to elegantly add entropy term to loss in DL frameworks (methology carries over to PyTorch)
git clone https://github.com/henry-prior/keras-a3c.git
cd keras-a3c
pip install -r requirements.txt
python runner.py