MADDPG

MADDPG (Multi-Agent Deep Deterministic Policy Gradient) is the official code release from OpenAI’s paper Multi-Agent Actor-Critic for Mixed Cooperative-Competitive Environments. The repository implements a multi-agent reinforcement learning algorithm that extends DDPG to scenarios where multiple agents interact in shared environments. Each agent has its own policy, but training uses centralized critics conditioned on the observations and actions of all agents, enabling learning in cooperative, competitive, and mixed settings. The code is built on top of TensorFlow and integrates with the Multiagent Particle Environments (MPE) for benchmarking. Researchers can use it to reproduce the experiments presented in the paper, which demonstrate how agents learn behaviors such as coordination, competition, and communication. Although archived, MADDPG remains a widely cited baseline in multi-agent reinforcement learning research and has inspired further algorithmic developments.