DID YOU KNOW? DTIC has over 3.5 million final reports on DoD funded research, development, test, and evaluation activities available to our registered users. Click
HERE to register or log in.
Accession Number:
ADA516462
Title:
Agents Technology Research
Descriptive Note:
Final rept. Jan 2005-Dec 2009
Corporate Author:
AIR FORCE RESEARCH LAB ROME NY INFORMATION DIRECTORATE
Report Date:
2010-02-01
Pagination or Media Count:
57.0
Abstract:
This report provides a comprehensive description of three separate efforts pursued by the agents technology research group. The efforts were focused on state abstraction methods for reinforcement learning, the multi-agent credit assignment problem, and distributed multi-agent reputation management. State abstraction is a technique used to allow machine learning technologies to cope with problems that have large state spaces. This report details the development and analysis of a new algorithm, Reinforcement Learning using State Abstraction via NeuroEvolution RL-SANE, that utilizes a new technology called neuroevolution to automate the process of state abstraction. The multi-agent credit assignment problem is a situation that arises when multiple learning actors within a domain are only provided with a single global reward signal. Learning is difficult in these scenarios because it is difficult for each agent to determine the value of its contribution to obtaining the global reward. In this report we describe the problem in detail and one specific approach we investigated that uses a Kalman filter to derive local rewards from global rewards. Multi-agent reputation management is important in open domains where the goals or the interests of the agents are diverse and potentially in conflict with one another. Reputation and trust can be used by the agents to determine which other agents in the system it should cooperate with and which it should not. This report details the development of the Affinity Management System AMS, an approach for managing and learning trust in a distributed fashion that utilizes self-modeling.
Distribution Statement:
APPROVED FOR PUBLIC RELEASE
