UW Computer Science Department UW Computer Science Department Optimizing Interaction Potentials for Multi- Agent Surveillance Dr. William M. Spears Dr. Diana F. Spears Wesley Kerr Suranga Hettiarachchi Dimitri Zarzhitsky University of Wyoming
UW Computer Science Department Optimizing Interaction Potentials for Multi-Agent Surveillance Dr. William M. Spears Dr. Diana F. Spears Wesley Kerr Suranga.
Dec 22, 2015
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
UW Computer Science DepartmentUW Computer Science Department
Optimizing Interaction Potentials for Multi-Agent Surveillance
Dr. William M. Spears
Dr. Diana F. Spears
Wesley Kerr
Suranga Hettiarachchi
Dimitri Zarzhitsky
University of Wyoming
UW Computer Science DepartmentUW Computer Science Department
Scenario
Terrain detector
Target detector
Separation radius Separation radius
Maximize area coverage and probabilityof detection of the targets by the ensemble.
UW Computer Science DepartmentUW Computer Science Department
Asset Control
• Control the motion of assets via interactions with neighboring assets.
• Interactions are determined via physics-based potentials (F = ma simulation). This is called “artificial physics” or AP.
• Optimize potentials to achieve the best performance using genetic algorithms.
UW Computer Science DepartmentUW Computer Science Department
Class of Potentials Examined
• We evolved asset-asset forces of two forms:
• Also, a viscous friction term is evolved, which ranges from no friction to full friction (1.0 to 0.0)
.attractive isit otherwise
close, asset too if repulsive is
/
:law force AP Standard
F
rGF p
.attractive isit otherwise
close, asset too if repulsive is
)()(224
:law force (LJ) Jones-Lennard
7
6
13
12
F
r
c
r
dF jiji
UW Computer Science DepartmentUW Computer Science Department
Target Control
• Stationary targets for baseline study.
• Gollum target controller:– Targets try to cross from left to right through
environment while sneaking through foliage.
• Super-Gollum target controller:– Also tries to avoid the UAV sensor footprint.
UW Computer Science DepartmentUW Computer Science Department
EnvironmentGenerator
Run F=ma simulatorwith particularinteraction potential.Measure performancew/n environments
Genetic Algorithmevolving population ofinteraction potentials
Particularinteractionpotential
Create environments
Return performance to GA
Output best interaction potential if desired performance met or time elapsed.
Architecture
UW Computer Science DepartmentUW Computer Science Department
Example Scenario
• 5-20 Micro-Air Vehicles (assets) at constant altitude.
• Environment size = 200x200 with some % foliage. • Targets of interest: 100 tanks.• Sensors have a fixed field of view.
– Target sensor coverage = 2
– Terrain sensor coverage = 2
• Surveillance over an area L2 > M 2 >> M 2
• GOAL: Maximize number of tanks detected that have been visible at some point in time.
UW Computer Science DepartmentUW Computer Science Department
Environment Generator
tank
forest
Separationradius
Foliage fieldof view
Target fieldof view
Note: separationradius can depend on foliage
Only 3 MAVsshown here
MAV
UW Computer Science DepartmentUW Computer Science Department
Experimental Comparison
• Methods:– We compare the evolved AP force laws with the
evolved LJ force laws (using 10 assets) against:• Stationary targets
• Gollum targets
• Super-Gollum targets
– Sensitivity analyses are also measured with respect to the number of assets and the fidelity of the target detector.
UW Computer Science DepartmentUW Computer Science Department
Stationary Targets
99.7
99.75
99.8
99.85
99.9
99.95
100
% Targets Found
0 10 20 30 40 50 60 70 80 90
% Foliage
AP Performance / Stationary Targets
ArtificialPhysics
Lennard-Jones
Both approaches work quite well when targets are stationary
UW Computer Science DepartmentUW Computer Science Department
Gollum Targets
0
20
40
60
80
100
% Targets Found
0 10 20 30 40 50 60 70 80 90
% Foliage
AP Performance / Gollum Targets
ArtificialPhysics
Lennard-Jones
LJ holds up better when targets are Gollum controlled
UW Computer Science DepartmentUW Computer Science Department
Super-Gollum Targets
LJ holds up better when targets are Super-Gollum controlled
0
20
40
60
80
100
% Targets Found
0 10 20 30 40 50 60 70 80 90
% Foliage
AP Performance / Super-Gollum Targets
ArtificialPhysics
Lennard-Jones
UW Computer Science DepartmentUW Computer Science Department
Super-Gollum Sensitivity
LJ is robust to increasing/decreasing number of assets
UW Computer Science DepartmentUW Computer Science Department
Super-Gollum Sensitivity
LJ holds up well when target detection probability lowered
UW Computer Science DepartmentUW Computer Science Department
Conclusions
• In general the LJ potential outperformed the AP potential.
• The evolved potential is robust with respect to the loss of one half of the assets or sensor degradation.
• The evolved potential is robust to changes in the percentage of foliage.
• This robustness emerges despite the fact that the evolved potential was not explicitly trained for these degradations.
UW Computer Science DepartmentUW Computer Science Department
Extensions
• Currently extending LJ to include a virtual mass term. If asset is over open area mass increases, and velocity decreases (obeying conservation of momentum).
• Also examining sweeping behavior controlled via kinetic theory.
UW Computer Science DepartmentUW Computer Science Department
Gollum Targets
LJM places assets over open areas more often, improving performance
0
20
40
60
80
100
% Targets Found
0 10 20 30 40 50 60 70 80 90
% Foliage
AP Performance / Gollum Targets
ArtificialPhysics
Lennard-Jones
Lennard-JonesMass
Related Documents
