Document

advertisement
A System based on Swarm
Intelligence and Ant Foraging
Techniques
By Kristin Eicher-Elmore
What is Swarm Intelligence?
Swarm Intelligence is a system in
which more than one unsophisticated
agents work together to create a
solution to difficult tasks.
Some definitions relevant to
Swarm Intelligence
• Collective behavior: The process of a group
of agents working together to achieve a
common goal.
• Reactive behavior: The reaction of an agent
to an outside stimulus such as a light.
• Emergent Phenomena: The process where
new behaviors develop dynamically during
the process of solving a task.
Why is using Swarm Intelligence
Techniques Important for
Robotics Systems?
Cost Effectiveness of:
Hardware and
Software
Cost Effectiveness of Hardware
• Simple agents have inexpensive hardware
that can be easily replaced if an agent is
damaged or lost in a hazardous
environment.
• Inexpensive hardware leads to the ability to
create large groups of agents that will be
able to cover a large area.
Cost Effectiveness of Software
• Using simple agents means that the
Software must be kept relatively simple and
uncomplicated. These systems generally
will not have the memory space for
complex algorithms. Thus, the reaction
times will generally be quicker for fast
reaction times.
Purpose of the System
To create a model for a system that will use
features of the ant foraging techniques to find the
shortest path to a goal for Search and Rescue
applications.
• Military uses
• Fire and disaster rescue
• Police uses
Any situation where there is danger and the need
to get to a victim quickly.
Ant Foraging Techniques
Ant foraging techniques were chosen
because of the ant’s ability to find the
shortest path to a goal.
Ant Foraging Technique
Definitions
• Stigmergy: Indirect communication used for
communication between different insects such as
ants. It is opposed to direct cues such as visual or
auditory ones.
• Pheromones: The chemical scent used by ants to
communicate with one another in an indirect way.
• Mass recruitment: The process by which ants are
directed towards a food source through the use of
pheromone trails.
How does Mass Recruitment
work to find the shortest path?
• The first ant to find the food source and return to
the nest leaves a pheromone trail for the other ants
to follow.
• Another ant follows this trail since it has the
freshest and strongest scent and leaves a scent trail
reinforcing the path.
• The path is now established and it will be the
shortest one because of the fact that the first one to
return took the least time finding the food.
Problems with adhering strictly
to ant foraging techniques
Ants will meander around until they find a
food source. When they return this path is
usually the shortest but wandering will not
work with a robot without ensuring that it
has a good efficient search algorithm.
The algorithm: How this system
ensures a good solution
• The use of colored zones.
• Constant changing of search methods
• Constant search for food source through
each search iteration
• Adequate obstacle avoidance
• Quick and Responsive RF Communication
The use of colored zones
• Once the robot reaches this marker the
search method is changed to a forward
search and this ensures that the robot will
keep moving on and to keep the boe-bot
from doubling back if it is making a left or
right wall hug search.
• This feature serves to force a progression
towards the goal.
Making progress with colored
zones
Constantly changing search
methods
• Changing search methods from a forward to
a right wall hug, and a left wall hug search
make sure that the robot will not keep trying
the same route over and over and wander
aimlessly.
• These search methods are stored in memory
to be communicated to the follower ants as
a map.
Robot changing search methods
Constant search for the food
• The food is searched for prior to every step
forward the robot makes. This ensures that
the robot will not miss it.
• When the robot senses the food it will enter
a separate search loop that does not involve
the switching of search methods performed
when in travel mode. This further ensures
that the food will not be passed by.
Food Search
Quick Obstacle Avoidance
• If the robot becomes stuck in a corner it will
make a sweep of the surrounding area to
find the farthest path from the wall closest
to the robot that is clear for both sensors.
• The robot also moves quickly through
obstacles.
Robot becoming Unstuck in a
Corner
Robot moving through Obstacle
Course
Quick and Responsive RF
Communication
• Fast wireless communication means the
follower robots can make a quick trip to the
food goal.
The Scout Robot Communicating
to Follower Robots
The System Algorithm Attempts
to Find the Shortest Path by:
• Using Zones to mark progress so that scouts make
quicker progress by not becoming stuck in one
area.
• Using more than one search method so that the
robot does not end up hugging one wall or
traveling forward and going along every obstacle
until the goal is reached.
• Sensing for the food at a constant rate so it isn’t
passed
• Obstacle Avoidance techniques that make sure the
robots do not become stuck in a corner for too
long.
Emergent Behavior: Nature vs.
Boe-Bot
Similarities
Differences
Sensing around obstacles No pheromone decay
A follower ant will scout Pheromone information
its own way to a food
is used as a guide rather
source if it becomes lost than a strict trail
Platform
All code is written in pBasic
for a Board of Education BS2pe
chip using the Parallax Basic
Stamp Editor
Hardware
• Parallax 433 Mhz Transceiver
• Ultrasonic Ping Sensors
• Photo-Resistors
The Code contains two
Controller Subsets
• Scout Search Loop
• Follower Search Loop
Each robot contains the same code, but a
flag indicates whether the robot starts out as
a scout or a follower
Observations and Results
•
•
•
•
Obstacle Avoidance
Getting out of corners
Finding the Light
XOR Error Checking and RF
communication
• Maze size and Progression
Obstacle Avoidance
• The code is successful at keeping the robot away
from both walls and moving forward for forward
search, and hugging the right and left walls for
forward search. The robot is always successful at
this.
• If the robot somehow gets very close to a wall on
one side, the ultrasonic becomes blinded. During
debugging it was found to record a large distance
when it is in fact right up close to it. All of the
sensors do this. So sometimes they get stuck
running straight into a wall at a slight angle.
Getting out of Corners
• Involves doing a sweep of the area and
chooses the first direction that is away from
obstructions on both sides of the robot in a
direction away from the obstruction.
• On average only two tries are required to
get out of a corner. At most three.
• The robot always chose the right direction.
Finding the Light
• Very successful since sensors are checked a every
step
• Once the robot senses it a separate sweep and
search is made until the light source has been
approached.
• Each robot has always found the light if close
enough and situations were rare of a robot going
by it when close unless another robot was
blocking the light.
• Average distance when light found was five-seven
inches away.
XOR Error Checking and RF
Communication
• XOR checksums are calculated at both ends and
compared before a message is accepted as correct.
• The scout will send out a message three times with
two seconds in between to ensure the correct
message is received.
• However during debugging and testing
communication never failed after the first attempt.
Maze size and Progression
• If the maze walls are too far apart then when the
robots go over a colored zone or marker, they
don’t realize they are making progress. They
might double back and think it was new ground
they were seeing when in fact it was the same
marker it has already seen.
• There did not seem to be any way to solve this in
code. The only solution seems to be keeping the
walls from being too far apart.
Live Demonstration
Recorded Demo
Challenges and Changes
• Communication and Error Checking
• Hardware Changes
• Mapping Technique Changes
Communication and Error
Checking
• At first there was more communication going on.
Each robot, scout and follower transmitted and
received. This was changed because of an eventual
lack of memory space.
• Because both scouts and followers transmitted and
received the XOR error checking scheme was
more exact and involved the receiver sending error
messages to the transmitter asking for another
transmission. Again this was simplified due to
little memory space.
Hardware Challenges and
Changes
• All code was simplified because major hardware
changes were needed.
• The main challenge was a lack of memory space
due to the needs of the transceiver.
• An extra chip a BS2 and a bread board were
added.
• The extra chip made it necessary to consider
building a battery pack that would hold five
batteries since more voltage was needed. A power
supply temporarily solved this problem.
Hardware Challenges and
Changes cont.
• One chip the BS2pe ran logic and movement, the
BS2 ran the transceiver.
• The biggest problem that could not be resolved:
chip to chip communication. The BS2pe would
not stop its program execution to notice the
interrupt from the BS2 with the transceiver.
• The BS2pe ran at 6000 instructions per second and
the BS2 ran at 4000 instructions per second. The
BS2 ran at a speed too slow to interrupt the
program execution of the BS2pe, so the code was
simplified.
Mapping Techniques
• At first actual directions were used instead
of search techniques. This used too much
memory space and because each robot
moves differently due to differences in
servo motors, search technique mapping
was more efficient.
Future Improvements
• Obstacle Avoidance and Colored Zones
• Finding the Goal (victim)
• Greater number of Agents and Scouts
Obstacle Avoidance and Zones
• Obstacle Avoidance code could remain the same
yet with more durable robots with better traction
and the ability to deal with potholes, etc..
• Instead of contrasting markers used to keep track
of progress, gps devices could be used that would
keep track of where the robot is in relation to its
starting point and the robot could actually see
forward progression from the starting point.
Finding the Victim
• Instead of using light sensors, a thermal
infrared camera could be used to identify
victims.
Greater Number of Agents and
Scouts
• A very large number of Scouts would be used to
create better coverage of an area.
• Once the victim was found by the quickest agent,
RF communication with more sophisticated error
checking could be used to bring followers
equipped with special equipment bringing
temporary relief like oxygen and water until
rescuers could reach the injured.
Conclusion
A successful swarm has these components:
• Collective behavior
• Emergent behavior
• Reactive Behavior
Collective Behavior in this
System
• Each agent shares the goal of finding the
food.
• When one Scout finds this food, a guide is
sent to the rest of the ants so that they can
find the food as well.
• All ants are cooperating together to find the
food.
Emergent Behavior
• Dynamic behaviors emerge during each run
of the program.
• A follower might find an optimal solution
better than the guide it received from the
Scout because it does not follow the
directions blindly but as a hint of the right
moves to make to the goal sensing for the
light as it goes.
Emergent Behavior cont.
• A separate search for the light source with the
proper obstacle avoidance and sweep methods for
the light if it has been sensed can create differing
behaviors in each ant even if they took the same
route enabling that no mistakes of missing the
light can be made.
• Changing search methods over time create
possible changes in behavior that keep an ant from
being stuck in a rut following one method.
Reactive Behavior
Apparent intelligence in insects comes from
there reactions to their environment.
Robots do can be made to react in similar
ways with very simple sensors and
hardware. Thus, swarm intelligence is an
ideal way to create large and simple systems
that can solve difficult problems with ease.
Download