[KongchangAI]
Concept集中训练-分散执行 / Centralized Training with Decentralized Execution

CTDE

A mainstream training paradigm in multi-agent reinforcement learning. During training, global information is shared to reduce policy-update interference; during execution, each agent makes decisions independently. Representative algorithms include QMIX and MAPPO.

Timeline (last 90 days)

Oct 13

MAPPO adopts the CTDE paradigm: each agent has its own actor, but during training they share a centralized critic that can observe the global state to reduce gradient variance, while the actor at deployment time still relies only on local observations.

Unverified50%

All Facts (1)

Unverified

MAPPO adopts the CTDE paradigm: each agent has its own actor, but during training they share a centralized critic that can observe the global state to reduce gradient variance, while the actor at deployment time still relies only on local observations.

50%

Source Articles

Training a 4-Player Squad in a 2D Battle Royale with MAPPO: 40 Seconds of Gameplay and Lessons LearnedOct 13The Nonstationarity Problem in Multi-Agent Reinforcement Learning: A Deep Dive into PACTOct 12