pymarl2

Why we included this project

This is a research codebase built around a simple observation: in cooperative multi-agent reinforcement learning, implementation details can matter as much as the algorithm itself. It reimplements QMIX and a range of value-based and actor-critic MARL methods, applying code-level tricks like value normalization, reward scaling, orthogonal initialization, n-step returns, larger batch sizes, and per-scenario hyperparameter tuning. The payoff is near-perfect win rates on the StarCraft Multi-agent Challenge. Researchers and graduate students will find it a useful reference for reproducing strong baselines and for figuring out which tweaks actually matter. It also includes configs for Google Football and an attention-based QMIX variant for communication tasks, so it's a solid starting point for benchmarking cooperative multi-agent methods, though it's a research codebase rather than a drop-in production library.