centralized training, decentralized execution

Centralised training with decentralised execution is the dominant design pattern in cooperative MARL, and the idea is delightfully practical. During training, in simulation, you are allowed to peek at everything — every agent's observation, the true state, everyone's action. But when the trained agents are deployed for real, each one must act on only its own local view, because that is all a real robot or player ever has. CTDE lets you exploit the god's-eye view to learn, then throw it away to run.

Concretely, the policies that get deployed are decentralised: agent i's actions depend only on its own observation history. The learning machinery, however, is centralised — typically a critic or a mixing network that conditions on the global state and the joint action. That centralised critic gives much cleaner learning signals and stabilises the moving target created by other learners, while the decentralised policies it shapes remain executable in the field.

This pattern is the backbone of nearly every modern cooperative algorithm: MADDPG, COMA, VDN, QMIX, and MAPPO are all CTDE methods. Its limit is honest: it assumes you actually have access to global information at training time (usually true in a simulator, often false in the wild) and it does not remove the underlying coordination problem — it just gives the learner a better vantage point from which to solve it.