2006 Fiscal Year Final Research Report Summary
Modelling and Control of Human-Agent Society
| Project/Area Number |
16560354
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
System engineering
|
|---|
| Research Institution | Kyushu University |
|---|
Principal Investigator |
MURATA Junichi Kyushu University, Faculty of Information Science and Electrical Engineering, Associate Professor, システム情報科学研究院, 助教授 (60190914)
|
|---|
| Project Period (FY) |
2004 – 2006
|
| Keywords | Agents / Virtual society of agents / Society models / Multi-agent systems / Reinforcement learning |
|---|
| Research Abstract |
Human societies have several issues that come from low morals, and their extent and degree have been magnified by use of advanced communication networks. Computers are embedded in appliances and other devices, and they work as agents for humans supporting human activities. These agents are connected to each other through the networks, and now human-agent double-layered societies are emerging. This study aimed at developing agents that can contribute to suppressing the above issues in human societies, and the following results have been derived.
Agents have been developed which suppress the issues by intervening human activities indirectly. Several models of human-agent societies were built, and through their simulations, it has turned out that evaluations by other people on a person's activity have a strong effect in suppressing the issues. Based on this finding, agents have been developed that send observations and evaluations to their human masters to guide the human activities. Their
… More
effectiveness has been confirmed by simulations.
Agents have been developed which suppress the issues by intervening human activities directly. Considering the above development, and also intending to avoid a possible situation that the human master stops using the agent disliking its direct intervention, agents have been developed that take actions for human masters based on observations, other people's evaluations, and the evaluations from their masters as well to suppress the issues. Its validity has been verified by simulations.
In order to give the agents adaptation abilities, use of learning algorithms and their improvement have been studied. Considering that the system comprises many humans and agents, reinforcement learning algorithm was adopted. Coordinated learning algorithm has been devised where the agents send necessary information only to each other. Also, a faster algorithm has been developed by partitioning the problem considering sub-goals. Their advantages have been confirmed by simulations. Less
|
