MAgent

MAgent is a research platform for many-agent reinforcement learning, presented as an AAAI 2018 demo paper. Unlike platforms built for one or a few agents, it was designed to scale reinforcement-learning experiments from hundreds to millions of agents in grid-world scenarios such as large-scale battles, to study collective behaviour. Important: this repository is no longer maintained. Its README points to MAgent2, a maintained fork from the Farama Foundation that can be installed with pip.

It remains of historical interest to researchers working on multi-agent and swarm reinforcement learning. MAgent makes no assumptions about agent design, so you can use rule-based policies or deep-learning frameworks. However, the original code targets Python 2.7 or 3, requires building native dependencies (CMake, Boost, JsonCpp, websocketpp), and its last update was in 2022, so it is unlikely to work smoothly on modern systems.

What to use instead: for new work, use MAgent2, the maintained fork from the Farama Foundation that can be installed with pip.