Documentation about centralized Learning for Multi Agent Reinforcement Learning

I know that it is now possibile in Mathworks to train multiple agents within the same environment for a collaborative task, using the so called "centralized" learning for agents of the same group. I understand the benefits of this approach and from the documentation it is clear how to code it. However, I was not able to find anywhere in the documentation any reference about the theory and the specific computations that this approach implies. I don't doubt that it works, but I would like to know more about the technical details if possible. To be more specific: I'm looking for references & informations related to the "LearningStrategy" property of object "rlMultiAgentTrainingOptions"

[Lin]

Hello Federico Toso，Do you have the answer to this question now? Can you share?

Answer 1

[the cyclist]

References for MATLAB functions are typically in two locations:

• at the bottom of the page for specific functions (or sometime one click away, in a document page about the underlying methods and algorithms)
• in the code itself for the function (which can be seen using "type functionName.m")

Without knowing the functions you are trying to understand, it's not possible to be more specific.

[Federico Toso]

Thank you for the answer. Unfortunately I could not find any reference even after your advices. I edited my question adding specific informations about the object I'm interested in. Please let me know if you have further tips

[the cyclist]

It seems to me that this documentation page describes the algorithm, and this page also has a lot of detail.

The second page also lists this reference:

[1] Sutton, Richard S., and Andrew G. Barto. Reinforcement Learning: An Introduction. Second edition. Adaptive Computation and Machine Learning. Cambridge, Mass: The MIT Press, 2018.

[Federico Toso]

Thank you for the additional infos. This covers a lot of details about Reinforcement Learning training, but there isn't anything specific about the "centralized" LearningStrategy property of MultiAgent training