Reinforcement Learning for Modeling Large-Scale Cognitive Reasoning (Phase II)
Loading...
Authors
Zhao, Ying
Bellonio, Jennie
Subjects
combat identification
reasoning
reinforcement learning models
rule-based AI
reasoning
reinforcement learning models
rule-based AI
Advisors
Date of Issue
2018-04
Date
Presented April 10-12, 2018
Period of Performance: 10/01/2017-09/30/2018
Period of Performance: 10/01/2017-09/30/2018
Publisher
Monterey, California: Naval Postgraduate School
Language
en_US
Abstract
Project Summary: In related research we demonstrated that Big Data (BD) techniques and analytics can provide potentially useful information in assisting the combat identification (CID) problem and improving the CTAP (Common Tactical Air Picture). CID is notoriously a very difficult function, often more art than science process is still very manual, and decision makers can experience cognitive overload so analytics is just one aspect of CID. CID would then be a good case study to investigate machine learning and artificial intelligence (AI). The focus of our research has been the question: Are machine learning (ML) and artificial intelligence (AI) systems able to learn and use the existing knowledge models for better and timely decision making for CID? Soar is an open source tool as a cognitive architecture, developed by the University of Michigan. Expert systems include reinforcement learning (Soar-RL). We first used Soar with a thesis student who was a Tactical Action Officer (TAO) that represented the "teacher", and we empirically showed that learning did take place in Soar. We then began using a Soar version which is integrated into Naval Simulation Software (NSS) as an agent to further aid in the CID decision making process, and eventually offered the opportunity to integrate real world data into the process. We also used LLA (Lexical Link Analysis) that provided initial learning rules to Soar. LLA is an unsupervised deep learning tool that can discover the correlations among the attributes, and therefore is used to discover the initial rules for NSS and Soar-RL. Our research empirically answered our core question that the ML/AI method, a.k.a. the Soar-RL integrated with LLA, can learn and use the existing knowledge models for better and timely decision making for CID. We saw error rates go from 3.7% to 0.4% with initial rules provided by LLA.
Type
Report
Description
NPS NRP Executive Summary
Series/Report No
Department
Information Sciences (IS)
Organization
Naval Research Program
Identifiers
NPS Report Number
NPS-18-N013-A
Sponsors
N810/ Information Warfare Branch
NPS Naval Research
NPS Naval Research
Funder
NPS-18-N013-A
Format
4 p.
Citation
Distribution Statement
Approved for public release; distribution is unlimited.
Rights
This publication is a work of the U.S. Government as defined in Title 17, United States Code, Section 101. Copyright protection is not available for this work in the United States.