Prerequisites

1. Git
2. VS Code
   1. Remote Development Extension Pack (if using Docker)
3. Docker + Docker Compose (if using Docker)
4. Python >= 3.6 (if not using Docker)

Quick start

Using Docker

1. git clone https://github.com/Minibrams/google-io-2022-rl.git
2. code google-io-2022-rl
3. Shift + CMD/CTRL + P -> Remote-Containers: Open Folder in Container

Not using Docker

1. git clone https://github.com/Minibrams/google-io-2022-rl.git
2. code google-io-2022-rl
3. python3 -m venv .venv (in VS Code terminal)
4. source .venv/bin/activate (in VS Code terminal)
5. pip install -r requirements.txt

Train your model

To train your model, run:

python train.py

Modify train.py and dqn/model.py as you see fit.

When running the train.py script, your model is run (and trained) against a random agent. Your model is saved to a file (dqn_model.h5) every 10 games.

Play your model

To play the connect-four game against your model, run:

python game.py

If a dqn_model.h5 file exists, the model is loaded before game start. Otherwise, you will be playing against a newly instantiated (random) agent.

Test your model

To test your trained model against a random agent, run:

python test.py

The model will be loaded from dqn_model.h5. The test will run 100 games against a random agent. The win rate will be reported at the end.