Kagan Tumer's Publications

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


Decreasing Communication Requirements for Agent Specific Rewards in Multiagent Learning. A. Iscen, C. Holmes Parker, and K. Tumer. In AAMAS-2011 Workshop on Adaptive and Learning Agents, Taipei, Taiwan, May 2011.

Abstract

In many different multiagent domains that require cooperation, success of the agents is heavily dependent on the communication between the agents. For better team performance, shaping individual rewards is essential. As a reward shaping method, difference rewards have shown previous success on many different domains, but the communication requirements are high. This paper defines the set of environment variables on which the agent's reward on the system depends. The definition is used to separate the information needed for the difference reward from the rest of the information about the environment. This concept of the effective area of an agent is explained with an example from the stateless gridworld domain. The experiments show that the performance of the agents with difference reward depends on the amount of information on their effective area. Moreover, if the communication method is designed carefully, the agents can have same quality of reward with 1\% to 10\% communication.

Download

[PDF]399.0kB  

BibTeX Entry

@incollection{tumer-iscen_ala11,
        author = {A. Iscen and C. Holmes Parker and K. Tumer},
        title = {Decreasing Communication Requirements for Agent Specific Rewards in Multiagent Learning},
        booktitle = {AAMAS-2011 Workshop on Adaptive and Learning Agents},
	month = {May},
	address = {Taipei, Taiwan},
	editors = {P. Vrancx and M. Knudson and M. Grzes},
	abstract={In many different multiagent domains that require cooperation, success of the agents is heavily dependent on the communication between the agents. For better team performance, shaping individual rewards is essential. As a reward shaping method, difference rewards have shown previous success on many different domains, but the communication requirements are high. This paper defines the set of environment variables on which the agent's reward on the system depends. The definition is used to separate the information needed for the difference reward from the rest of the information about the environment. This concept of the effective area of an agent is explained with an example from the stateless gridworld domain. The experiments show that the performance of the agents with difference reward depends on the amount of information on their effective area. Moreover, if the communication method is designed carefully, the agents can have same quality of reward with 1\% to 10\% communication.},
	bib2html_pubtype = {Workshop/Symposium Papers},
	bib2html_rescat = {Multiagent Systems},
        year = {2011}
}

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