Questions about the training and execution paradigm?

[BigBearBlacken]

As shown in the Architecture of MAT, it seems that the encoder and decoder needs all the observations and actions of all agents in the training phase. But in the execution phase, agent can't obtain other agents' observations and actions. So how does each agent choose actions according to the actions of other agents recursively in the execution phase? Or the algorithm adopts centralized training and centralized execution paradigm?

[morning9393]

Hiya,

Thanks for your attention, the vanilla MAT with recursion adopts centralized training and a centralized execution paradigm in deed, which concentrated more on modeling the interrelationship between agents. For decentralized scenarios, you may choose the MAT-dec version, which captures the interrelationship with an encoder and makes decisions with decentralized MLPs.

Hoping it is helpful, Muning

[BigBearBlacken]

Got it! Thanks for your reply!