Esquilax: A Multi-Agent Simulation and RL Library

[zombie-einstein]

I've released Esquilax a library for simulation of large scale multi-agent systems (think swarms, social networks etc.) and their use for training and testing multi-agent reinforcement learning and other ML techniques.

It provides transformations that allow users to map interaction functions over groups of agents (e.g. agents observing each other on a 2d space) such that developers can concentrate on model design rather than low-level algorithms. It also includes various modelling utilities, and functionality for multi-agent RL and neuro-evolution training.

The full documentation can be found here and examples can be found on the Github repo here.

You can also see this project for a larger project example. This animation was produced using an RL agent trained in an Esquilax environment. In the simulation each agent sees only its local view of the environment and is rewarded for staying close to other agents (but not colliding), the flocking behaviour emerges from the RL policy applied to each agent individually.