siyamsajeebkhan / google_research_football_kaggle

Google Research Football Kaggle Competition

This repo contains part of the source code I used for Google Research Football Kaggle Competition. Please refer to this blog post for a detailed account of of my method.

The source code consists of 2 parts: [imitation learning] and [reinforcement learning]. These are the 2 major components I used for this challenge.

Run Instruction

Imitation Learning

First, enter the imitation_learning folder and spin up a docker container:

$ cd imitation_learning
$ ./build_docker.sh
$ ./run_docker.sh

Then, run the following make command inside the docker container which will automatically download the historical episodes via kaggle API, perform ETL on the raw episode data, and train the imitation learning model.

$ make etl_and_train

Please refer to the Makefile to see the components of the pipeline.

The configuration variables and hyperparameters can be found in yamls/config.yaml.

Reinforcement Learning

First, enter the rl_agent folder and spin up a docker container:

$ cd imitation_learning
$ ./build_docker.sh
$ ./run_docker.sh

Then, run the following make command inside the docker container which train a PPO agent against a hard coded rule-based bot:

$ make train_ppo

The configuration variables and PPO hyperparameters can be found in config.yaml. Feel free to edit them for your own use.

Notice that there is NO pretrained model provided in this repo. If you intend to incorporate a pretrained model, change the load_pretrained config variable to True in config.yaml and add your pretrained model to the pretrained folder.