Multi-Agent RL

Train multiple agents to coordinate, cooperate, or compete effectively through multi-agent reinforcement learning, enabling complex multi-agent systems for software, robotics, and games. Multi-agent RL extends single-agent RL to settings with multiple simultaneously learning agents. Each agent observes the environment (and optionally other agents), and selects actions. The key challenge is non-stationarity: each agent perceives a changing world because the other agents are also learning. Solutions include centralized training with decentralized execution (CTDE), where agents share gradients during training but act independently at inference; and opponent modeling, where agents learn to predict other agents behavior.