Repository for experimental development of fully autonomous AI for the AI4REALNET project. The work focuses on the development of solutions for multi-agent systems, using the flatland simulation environment.
These packages were developed with python 3.10.11 For package versions see the requirements.txt. In the .vscode folder, a launch.json and settings.json are available to run the different models and perform unittesting.
Currently the repository contains the following baseline models upon which further development should be done:
The original code from the paper Learning When to Communicate at Scale in Multiagent Cooperative and Competitive Tasks by Singh et al., published at the ICLR in 2019, has been adapted to the flatland environment and integrated into the T3.4 codebase.
The JBR_HSE group developed the best RL solution to the Flatland challenge at NeurIPS 2020. The solution was already developed specifically for flatland and hence only needed to be slightly adapted for integration into the T3.4 codebase.
The interpretable communication protocols and model architecture are currently WIPs and not functional.
WIP :)