Kagan Tumer's Publications

Display Publications by [Year] [Type] [Topic]


Efficient Reward Functions for Adaptive Multi-Rover Systems. K. Tumer and A. Agogino. In AAMAS-05 Workshop on Learning and Adaptation in Multiagent Systems, Utrecht, Netherlands, July 2005.

Abstract

This paper addresses how efficient reward methods can be applied to multiple agents co-evolving in noisy and changing environments, under communication limitations. This problem is approached by ``factoring'' a global reward over all agents into agent-specific rewards that have two key properties: 1) agents maximizing their agent-specific rewards will tend to maximize the global reward, 2) an agent's action has a large influence over its agent-specific reward allowing it to evolve quickly. Agents using these agent-specific rewards are tested in episodic and non-episodic, continuous-space multi-rover environment where rovers evolve to maximize a global reward function over all rovers. The environments are dynamic (i.e. changes over time) and can be noisy and can restrict communication between agents . We show that a control policy evolved using these agent-specific rewards outperforms global reward methods by up to 400\%.More notably, in the presence of a larger number of rovers or roverswith noisy and communication limited sensors, the proposed methodoutperforms global reward by a higher percentage than innoise-free conditions with a small number of rovers.

Download

[PDF]304.8kB  

BibTeX Entry

@incollection{tumer-agogino_lamas05,
        author = {K. Tumer and A. Agogino},
        title = {Efficient Reward Functions for Adaptive Multi-Rover Systems},
        booktitle = {AAMAS-05 Workshop on Learning and Adaptation in Multiagent Systems},
        address = {Utrecht, Netherlands},
	editors = {K. Tuyls and P. Jan't Hoen and S. Sen and K. Verbeeck},
        month = {July},
	abstract = {This paper addresses how efficient reward methods can be applied to multiple agents co-evolving in noisy and changing environments, under communication limitations. This problem is approached by ``factoring'' a global reward over all agents into agent-specific rewards that have two key properties: 1) agents maximizing their agent-specific rewards will tend to maximize the global reward, 2) an agent's action has a large influence over its agent-specific reward allowing it to evolve quickly. Agents using these agent-specific rewards are tested in episodic and non-episodic, continuous-space multi-rover environment where rovers evolve to maximize a global reward function over all rovers. The environments are dynamic (i.e. changes over time) and can be noisy and can restrict communication between agents . We show that a control policy evolved using these agent-specific rewards outperforms global reward methods by up to 400\%.
More notably, in the presence of a larger number of rovers or rovers
with noisy and communication limited sensors, the proposed method
outperforms global reward by a higher percentage than in
noise-free conditions with a small number of rovers.},
	bib2html_pubtype = {Workshop/Symposium Papers},
	bib2html_rescat = {Multiagent Systems, Robotics},
        year = {2005}
}

Generated by bib2html.pl (written by Patrick Riley ) on Tue Jun 26, 2018 19:10:42