Toolbox Multi-Agent Systems

Multi-agents systems (MAS) are a core area of research in artificial intelligence. Such systems consist of several decision-making agents which interact in a shared environment. Using MAS a wide range of applications can be addressed including autonomous driving, the semantic web, smart grids, multi-robot factories, automated trading, etc.

This toolbox is intended to be a visualization and simulation instrument in which we document state-of-the art methods and provide guidelines on when to use which approach. 

The multi-agent systems toolbox currently contains two prototypical environments and several of their subproblems. 

  1. Multi-agent pathfinding in which a number of agents need to move from initial locations to given goal locations.
  2. Multi-agent pick up and delivery in which a number of agents have to pick up and deliver packages. These tasks are generated randomly or taken from a predefined list. 

In this project, visualization and simulation is done in Unity and environments can be set up using a configuration file. Moreover we use the ML-agents toolkit which allows for training models and adding your own algorithms. Several state-of-the art planning algorithms have been implemented and added to the framework. We have also added some toy examples that clearly show how to use the ML-agents toolkit and how to add your own algorithms. Besides visualizing your algorithms you can also analyze statistics that are saved while running the algorithms. You can also add your own metrics. 

This project is being developed in the context of the AI Flanders Research Program. Check out the existing environments below.

Environments

Multi-agent path finding

Multi-agent pick up and delivery