Transcript A Survey of Artificial Intelligence Applications in Water

A Survey of Artificial Intelligence
Applications in Water-based
Autonomous Vehicles
Daniel D. Smith
CSC 7444
December 8, 2008
Autonomous Vehicles
• Vehicle which can perform all the functions
required of it without outside intervention
while operating in an uncontrolled
environment.
• Types:
– Land-based
– Water-based (surface and underwater)
– Air-based
Past and Current Research in
Biological Engineering
• Program uses Autonomous Waterbased vehicles for a variety of purposes
– Water quality monitoring
– Bird predation reduction
– Pollution tracking
• Research is moving into areas involving
multiple agents which need to interact
with each other and the environment in
intelligent ways.
Past and Current Vehicles
Problems with traditional
control methods
• Complex - especially for underwater
vehicles
• Non-adaptive
• Can be slow
Neural Networks
and
Self-Organizing Maps
Neural Networks
• Some systems use the neural network
along side a more traditional controller
to provide on-line adjustments to the
controller itself.
• Other systems utilize the neural network
as one stage of a multi-stage process.
A Neural Network Controller
for Diving of a Variable Mass
Autonomous Underwater
Vehicle
Mazda Moattari and Alireza Khayatian
Variable Mass Submarine
• System developed to compensate for
changing dynamics of vehicle
• As vehicle burns fuel, the mass of the vehicle
changes
• Neural network provides correction to
traditional PID control system to keep dive
angle correct.
• Correction is done by using a second neural
network to estimate the Jacobian of the
output of the control system.
Self-tuning PID Controller
Control of Underwater
Autonomous Vehicles Using
Neural Networks
Michael Santora, Joel Alberts,
and Dean Edwards
Submarine Guidance
• Simulation for control of a submarine’s
heading and depth
• Assumptions:
– No obstacles
– Constant speed
– Waypoint reached if location was within a
1m radius circle of the actual waypoint.
Submarine
Controller and Neural Network
Autonomous Underwater
Vehicle Guidance by
Integrating Neural Networks
and Geometric Reasoning
Gian Luca Foresti, Stefani Gentili,
and Massimo Zampato
Vision-based Guidance
• Neural network used as the
first stage of a two stage
artificial vision system
• Neural network is trained on
test images to help locate the
edges of underwater
pipelines.
• After training, correctly
classified 93% of 100 test
images.
Training Image
Classified Image
A Self-Organizing Map Based
Navigation System for an
Underwater Robot
Kazuo Ishii, Shuhei Nishada, and Tamaki Ura
SOM with Learning
• 20 x 20 node map
• 5000 training data sets
• On-line, map adapts to the
environment.
Genetic Algorithms
A Hierarchical Global Path
Planning Approach for AUV
Based on Genetic Algorithm
QiaoRong Zhang
GA Description
• Use octree to decompose 3D space into
uniform regions.
• Label cells as Full, Empty, or Mixed
• GA constructs path from Source to Goal
through Empty and Mixed Cells
– Uses 3 genetic operations:
• Reproduction: Fit individuals (paths) progress to the next
generation
• Crossover: Create new individuals from the fittest of the
previous population
• Mutation (Insert, Delete, Replace)
– Fitness is a combination of shortest distance and
most empty cells in path.
Line of Sight Guidance
with Intelligent Obstacle
Avoidance for Autonomous
Underwater Vehicles
Xiaoping Wu, Zhengping Feng, Jimao Zhu,
and Robert Allen
Tuning Fuzzy Logic with GA
• AUV has fuzzy logic planner
– 2 inputs: Distance and angle to obstacle
– 1 output: Heading correction to avoid
• GA used to minimize cross-track error by
tuning the fuzzy logic planner
• Fitness is determined by smallest cross-track
error over a safe distance
• Percentage of fit individuals of each
population kept for next generation
Results of Simulation
Before Tuning
After Tuning
Evolutionary Path Planning for
Autonomous Underwater
Vehicles in a Variable Ocean
Alberto Alvarez, Andrea Caiti, and Reiner Onken
Optimizing energy cost
• Population is N randomly generated potential
paths from source to goal
• Fitness is determined by computing the
energy cost of moving the vehicle along the
path taking into account ocean currents.
• N/2 individuals with lowest cost (fittest)
chosen
• Parents and offspring kept
• Mutation is limited to the less fit individuals of
the population and involves randomly moving
one waypoint of the path.
Evolutionary Path Planning
and Navigation of
Autonomous Underwater
Vehicles
V. Kanakakis and N. Tsourveloudis
B-Spline Genetic Algorithm
• Off-line path planning
• B-Spline path defined by:
– Start, End, and Second Point
– Six free-to-move points
• Population size of 30
• Single point crossover with mutation
• Fitness function defined by:
6
f   ai  f i
i1
Questions?
Thank you