Learning to cooperate, compete, and communicate

Multiagent RL is harder than standard RL because agents must learn while adapting to one another in competitive and communication-heavy settings.

The article argues that decentralized methods often struggle in these environments, especially when agents need to coordinate or model other agents’ behavior.

MADDPG addresses this with a centralized critic during training, giving access to other agents’ observations and actions to stabilize learning.

This centralized training approach helps agents learn stronger cooperative, competitive, and speaker-listener communication behaviors.