This repository contains a basic reimplementation of mean embeddings as published in "Deep Reinforcement Learning for Swarm Systems" (https://jmlr.org/papers/v20/18-476.html) in pytorch as well as pytorch geometric. Additionally, it contains a policy implementation based on Message Passing Networks as shown in Chapter 2 of my dissertation.
The tasks can be found in an extra repository and learning is based on a fork of stable-baselines which has been extended to deal with graph observation spaces. Both will be installed as a dependency.
This repo uses uv as a package manager. To install, first clone the repository and optionally install python 3.12
uv python install 3.12
Create a virtual environment
uv venv --python 3.12
and run
uv sync
to install the dependencies.
The folder scripts contains files for running training and evaluation. Adjust log_path folders to your needs.