Trains an agent with (stochastic) Policy Gradients on Pong

Based on this awesome blog post from the great Andrej Kaparthy: https://karpathy.github.io/2016/05/31/rl/

Install

Create a virtual env and activate it.
python -m venv venv
source venv/bin/activate
Install requirements.
pip install -r requirements.txt
Then install the gym.
pip install "gym[atari]"
Accept licences.
pip install "gym[accept-rom-license, atari]"

Run

python src/pong.py"

Architecture

Weights
W1: (200 x 6400)
W2: (1 x 200)

Math

References

Andrej Kaparthy blog
David Silver - Reinforcement Learning
Richard S. Sutton - Reinforcement Learning: An Introduction