PettingZoo

Why we included this project

Anyone training agents that learn alongside or against other agents needs a standard way to define and run those environments. PettingZoo fills that role for Python, much as Gymnasium does for single-agent reinforcement learning. It bundles a wide set of ready-made reference environments, including multi-player Atari games, coordination-heavy cooperative games, classic board and card games, and cooperative simulators, so you can benchmark a new algorithm without building your own worlds first. Two interaction styles are supported, a sequential turn-based cycle and a parallel mode for environments where agents act simultaneously, and researchers comparing MARL methods tend to value having both. Every environment carries a strict version number, which keeps experiments reproducible and lets teams report results others can actually verify.