TY - GEN
T1 - Networked reinforcement learning
AU - Oku, Makito
AU - Aihara, Kazuyuki
PY - 2008
Y1 - 2008
N2 - Recently, many models of reinforcement learning with hierarchical or modular structures have been proposed. They decompose a task into simpler sub-tasks and solve them with multiple agents. In these models, however, topological relations of agents are severely restricted. By relaxing the restrictions, we propose networked reinforcement learning where each agent in a network acts in parallel as if the other agents are parts of the environment. Although convergence to an optimal policy is no longer assured, we show by numerical simulations that our model performs well at least in some simple situations.
AB - Recently, many models of reinforcement learning with hierarchical or modular structures have been proposed. They decompose a task into simpler sub-tasks and solve them with multiple agents. In these models, however, topological relations of agents are severely restricted. By relaxing the restrictions, we propose networked reinforcement learning where each agent in a network acts in parallel as if the other agents are parts of the environment. Although convergence to an optimal policy is no longer assured, we show by numerical simulations that our model performs well at least in some simple situations.
KW - Hierarchical reinforcement learning
KW - Modular reinforcement learning
KW - Partially observable markov decision process
UR - http://www.scopus.com/inward/record.url?scp=78449232010&partnerID=8YFLogxK
M3 - 会議への寄与
AN - SCOPUS:78449232010
SN - 9784990288020
T3 - Proceedings of the 13th International Symposium on Artificial Life and Robotics, AROB 13th'08
SP - 469
EP - 472
BT - Proceedings of the 13th International Symposium on Artificial Life and Robotics, AROB 13th'08
T2 - 13th International Symposium on Artificial Life and Robotics, AROB 13th'08
Y2 - 31 January 2008 through 2 February 2008
ER -