Multi-agent Learning and the Reinforcement Gradient
M. Kaisers, and K. Tuyls. Proc. of 9th European Workshop on Multi-agent Systems (EUMAS 2011), Maastricht University, (2011)
Abstract
The number of proposed reinforcement learning algorithms appears to be ever-growing. This article tackles the diversification by showing a persistent principle in several independent reinforcement learning algorithms that have been applied to multi-agent settings. While their learning structure may look very diverse, algorithms such as Gradient Ascent, Cross learning, variations of Q-learning and Regret minimization all follow the same basic pattern. Variations of Gradient Ascent can be described by the projection dynamics and the other algorithms follow the replicator dynamics. In combination with some modulations of the learning rate and deviations for the sake of exploration, they are primarily different implementations of learning in the direction of the reinforcement gradient.
%0 Conference Paper
%1 Kaisers2011
%A Kaisers, Michael
%A Tuyls, Karl
%B Proc. of 9th European Workshop on Multi-agent Systems (EUMAS 2011)
%D 2011
%I Maastricht University
%K dy-,evolutionary game learning learning,namical systems,reinforcement theory,gradient
%T Multi-agent Learning and the Reinforcement Gradient
%U http://michaelkaisers.com/publications/2011_EUMAS_MKaisers.pdf
%X The number of proposed reinforcement learning algorithms appears to be ever-growing. This article tackles the diversification by showing a persistent principle in several independent reinforcement learning algorithms that have been applied to multi-agent settings. While their learning structure may look very diverse, algorithms such as Gradient Ascent, Cross learning, variations of Q-learning and Regret minimization all follow the same basic pattern. Variations of Gradient Ascent can be described by the projection dynamics and the other algorithms follow the replicator dynamics. In combination with some modulations of the learning rate and deviations for the sake of exploration, they are primarily different implementations of learning in the direction of the reinforcement gradient.
@inproceedings{Kaisers2011,
abstract = {The number of proposed reinforcement learning algorithms appears to be ever-growing. This article tackles the diversification by showing a persistent principle in several independent reinforcement learning algorithms that have been applied to multi-agent settings. While their learning structure may look very diverse, algorithms such as Gradient Ascent, Cross learning, variations of Q-learning and Regret minimization all follow the same basic pattern. Variations of Gradient Ascent can be described by the projection dynamics and the other algorithms follow the replicator dynamics. In combination with some modulations of the learning rate and deviations for the sake of exploration, they are primarily different implementations of learning in the direction of the reinforcement gradient.},
added-at = {2016-12-19T12:09:05.000+0100},
author = {Kaisers, Michael and Tuyls, Karl},
biburl = {https://www.bibsonomy.org/bibtex/226d0f40788734a24a15d595bff326d4a/swarmlab},
booktitle = {Proc. of 9th European Workshop on Multi-agent Systems (EUMAS 2011)},
interhash = {48b9d4b35512511f37214a28cb17e04a},
intrahash = {26d0f40788734a24a15d595bff326d4a},
keywords = {dy-,evolutionary game learning learning,namical systems,reinforcement theory,gradient},
publisher = {Maastricht University},
timestamp = {2016-12-19T12:18:59.000+0100},
title = {{Multi-agent Learning and the Reinforcement Gradient}},
url = {http://michaelkaisers.com/publications/2011_EUMAS_MKaisers.pdf},
year = 2011
}