How to Train Multiple Reinforcement Learning Agents In Basic Grid World? (Multiple Agents)

My problem is very similar to "Train Reinforcement Learning Agent in Basic Grid World", however I wish to expand this problem to a multi-robot (multiple agent) scenario with unique start and goal coordinates per agent, rather than just one agent. I have created a basic gridWorld environment with obstacles, in which I wish to train 2-10 robots to get to their goal points without collision. I have read a lot of the multi-agent examples provided with the tool box, however, they seem to differ from my specific problem and I can't quite put my finger on how to accomplish my problem utilizing those examples. Thank you for your time and consideration.

Answer 1

[Emmanouil Tzorakoleftherakis]

Training multiple agents simultaneously is currently only supported in Simulink. The predefined Grid World environments in Reinforcement Learning Toolbox are implemented in MATLAB so if you want to do multi-agent training, you would need to wrap the MATLAB environment in a MATLAB Function block in Simulink, and then follow the template/steps in the multi-agent examples you have been looking at.

[Donald Alexander]

Thank you for your response Emmanouil. I have attempted to do this through the reinforcementLearningDesigner tool available. However, it tells me that the rlQAgent type (the type used in the basic grid world example) is not supported.

Error: "Reinforcement Learning Designer not support this agent type".

Is it possible to use the rlQAgent with a multi-robot situation in simulink? Are there ways to avoid this error?

Thank you

[Emmanouil Tzorakoleftherakis]

The Q agent is unfortunately not supported yet for use within Reinforcement Learning Designer. You would have to set it up programmatically

[Arman Ahmadian]

It is very unfortunate that Matlab only supports MARL in simulink only. I've been waiting for multiple releases of matlab to be able to train multiple agents without the hassle of turning matlab environment classes into simulink and yet we're into 2022b and still no support.