AbstractMulti-agent systems need to communicate to coordinate a shared task. We show that a recurrent neural network (RNN) can learn a communication protocol for coordination, even if the actions to coordinate are performed steps after the communication phase. We show that a separation of tasks with different temporal scale is necessary for successful learning. We contribute a hierarchical deep reinforcement learning model for multi-agent systems that separates the communication and coordination task from the action picking through a hierarchical policy. We further on show that a separation of concerns in communication is beneficial but not necessary. As a testbed, we propose the Dungeon Lever Game and we extend the Differentiable Inter-Agent Learning (DIAL) framework. We present and compare results from different model variations on the Dungeon Lever Game.
%0 Journal Article
%1 doi:10.1080/01969722.2019.1677335
%A Ossenkopf, Marie
%A Jorgensen, Mackenzie
%A Geihs, Kurt
%D 2019
%I Taylor & Francis
%J Cybernetics and Systems
%K itegpub vs
%N 8
%P 672-692
%R 10.1080/01969722.2019.1677335
%T When Does Communication Learning Need Hierarchical Multi-Agent Deep Reinforcement Learning
%U https://doi.org/10.1080/01969722.2019.1677335
%V 50
%X AbstractMulti-agent systems need to communicate to coordinate a shared task. We show that a recurrent neural network (RNN) can learn a communication protocol for coordination, even if the actions to coordinate are performed steps after the communication phase. We show that a separation of tasks with different temporal scale is necessary for successful learning. We contribute a hierarchical deep reinforcement learning model for multi-agent systems that separates the communication and coordination task from the action picking through a hierarchical policy. We further on show that a separation of concerns in communication is beneficial but not necessary. As a testbed, we propose the Dungeon Lever Game and we extend the Differentiable Inter-Agent Learning (DIAL) framework. We present and compare results from different model variations on the Dungeon Lever Game.
@article{doi:10.1080/01969722.2019.1677335,
abstract = {AbstractMulti-agent systems need to communicate to coordinate a shared task. We show that a recurrent neural network (RNN) can learn a communication protocol for coordination, even if the actions to coordinate are performed steps after the communication phase. We show that a separation of tasks with different temporal scale is necessary for successful learning. We contribute a hierarchical deep reinforcement learning model for multi-agent systems that separates the communication and coordination task from the action picking through a hierarchical policy. We further on show that a separation of concerns in communication is beneficial but not necessary. As a testbed, we propose the Dungeon Lever Game and we extend the Differentiable Inter-Agent Learning (DIAL) framework. We present and compare results from different model variations on the Dungeon Lever Game.},
added-at = {2019-11-13T10:46:05.000+0100},
author = {Ossenkopf, Marie and Jorgensen, Mackenzie and Geihs, Kurt},
biburl = {https://www.bibsonomy.org/bibtex/25a2aa183e3e4f6e5ed9c8dfc62d599fd/vskassel},
doi = {10.1080/01969722.2019.1677335},
eprint = {https://doi.org/10.1080/01969722.2019.1677335},
interhash = {afcaf7548c0b312b06cafb5ad220ff34},
intrahash = {5a2aa183e3e4f6e5ed9c8dfc62d599fd},
journal = {Cybernetics and Systems},
keywords = {itegpub vs},
number = 8,
pages = {672-692},
publisher = {Taylor & Francis},
timestamp = {2019-11-13T10:46:05.000+0100},
title = {When Does Communication Learning Need Hierarchical Multi-Agent Deep Reinforcement Learning},
url = {https://doi.org/10.1080/01969722.2019.1677335 },
volume = 50,
year = 2019
}