Artificial Intelligence in the Military

Download Report

Transcript Artificial Intelligence in the Military

Artificial Intelligence in the
Military
Presented by
Carson English, Jason Lukis,
Nathan Morse and Nathan Swanson
Overview
•
•
•
•
•
History
Neural Networks
Automated Target Discrimination
Tomahawk Missile Navigation
Ethical issues
History
• 1918 – first tests on guided missiles
• 1945 – Germany makes first ballistic
missile
• 1950 – AIM-7 Sparrow
– “fire-and-forget
History
• 1973 – remotely piloted vehicles (RPVs)
– Used to confuse enemy air defenses
• 1983 – tomahawk missile first used by navy
– Uses terrain contour matching system
• 1983 – Reagan make his famous star wars speech
• 1988 – U.S.S. Vincennes mistakenly destroys
Iranian airbus due to autonomous friend/foe radar
system
History
• 1991 – Smart bombs used in Gulf War to
selectively destroy enemy targets
– Praised for its precision and effectiveness
Neural Networks
•
•
•
•
Inspired by studies of the brain
Massively parallel
Highly connected
Many simple units
Structure of a neuron in a neural net
Neural net with three neuron layers
Three Main Neural Net Types
• Perceptron
• Multi-Layer-Perceptron
• Backpropagation Net
Perceptron
Multi-Layer-Perceptron
Backpropagation Net
Areas where neural nets are useful
·
·
·
·
·
·
·
·
·
·
pattern association
pattern classification
regularity detection
image processing
speech analysis
optimization problems
robot steering
processing of inaccurate or incomplete inputs
quality assurance
simulation
Limits to Neural Networks
• the operational problem encountered when
attempting to simulate the parallelism of neural
networks
• inability to explain any results that they obtain
Automated Target Discrimination
As researched by the Computational
NeuroEngineering Laboratory in Gainsville, FL
•
•
•
•
•
•
SAR (Synthetic Aperture Radar)
CFAR (Constant False Alarm Rate)
QGD (Quadratic Gamma discriminator)
NL-QGD (multi-layer perceptron)
Example
Results
Synthetic Aperture Radar
•
•
•
•
Data collection for ATD
Self-illuminating imaging radar
Creates a height map of a surface
Maintains spatial resolution regardless of
distance from target
• Can be used day and night regardless of
cloud cover
Picture of SAR rendering
Two Constant False Alarm
method for determining targets
Quadratic Gamma discrimination
Non Linear QGD
Example
Results
• After training, all three discriminators were
run on a data set representing 7km2 of
terrain. Target detection threshold was set
to 100%.
• CAFR resulted in 4,455 false alarms.
• QGD resulted in 385 false alrams.
• NL-QGD resulted in 232 false alarms.
Tomahawk Missile Navigation
• Missile contains a map of terrain
• Figures out its current position from
percepts (radar & altimeter)
• Uses a modified Gaussian least square
differential correction algorithm, a step size
limitation filter, and a radial basis function
Weight matrix
Radial Basis Function
Gaussian Least Square Correction
Necessary Condition
Sufficient Condition
Step size limitation filter
Tolerence error = 10^-8
Ethics
• Accountability
– Legal
– Political
– Example: Aegis defense system shoots down an Iranian
Airbus jetliner in 1988
• Use of AI in warfare
• Ethics of Research and Development
– Potential uses
– Military Funding of AI
– Passing of the blame “just doing my job”
Sources
• “Target Discrimination in Synthetic Aperture Radar (SAR) using
Artificial Neural Networks” Jose C. Principe, Munchurl Kim, John W.
Fisher III. Computational NeuroEngineering Laboratory. EB-486
Electrical and Computer Engineering Department. University of
Florida.
• Sandia National Laboratories. http://www.sandia.gov/radar/sar.html
• Jet Propulsion Laboratory: California Institute of Technology.
http://southport.jpl.nasa.gov/desc/imagingradarv3.html
• Wageningen University, The Netherlands.
http://www.gis.wau.nl/sar/sig/sar_intr.htm