Multi-Agent Reinforcement Learning From the Ground Up: PPO, GRPO, CTDE, and MAPPO
Multi-agent reinforcement learning (MARL) studies how several agents learn to act in a shared environment. This post is a background survey of the cooperative MARL literature, written so that a reader with no reinforcement-learning background can follow it from start to finish. It is organized in four parts. Part 1 builds the single-agent foundations — the vocabulary, then the formulas, with the purpose of every symbol. Part 2 covers the two policy-update algorithms everything else rests on, PPO and GRPO, each with a worked example. Part 3 moves to many agents: the CTDE paradigm, the methods that established it (DDPG, MADDPG, COMA, VDN, QMIX), and MAPPO. Part 4 describes the benchmark environments. A summary and a glossary close the post. ...