QUORUM SENSING FOR COLLECTIVE ACTION

AND DECISION-MAKING IN MOBILE AUTONOMOUS TEAMS

Donald A. Sofge

Naval Research Laboratory, 4555 Overlook Avenue SW, Washington, DC, U.S.A.

William F. Lawless

Paine College, 1235 15th Street, Augusta, GA, U.S.A.

Keywords: Autonomous, Multi-agent team, Ensemble, Quorum sensing, Collective, Bio-inspired control, Swarm.

Abstract: Design of controllers for teams of mobile autonomous systems presents many challenges that have been

addressed in biological systems, such as behavior-based control paradigms that are decentralized,

distributed, scalable, and robust. Quorum sensing is a distributed, decentralized decision-making process

used by bacteria and by social insects to coordinate group behaviors and perform complex tasks. It is used

by bacteria to control the colony behavior for a variety of functions, such as biofilm construction or

initiating pathogenicity inside a host. It is used by social insects including the ant Temnothorax albipennis

to collectively evaluate and select from amongst potentially many new nesting sites.Honeybees (Apis

mellifera) use quorum sensing to collectively choose a new nesting site when the swarm grows too large

and needs to split. It is shown that the quorum sensing paradigm may be used to provide robust

decentralized team coordination and collective decision-making in mobile autonomous teams performing

complex tasks. In this effort quorum sensing-inspired techniques are developed and applied to the design of

a decentralized controller for a team of mobile autonomous agents surveying a field containing buried

landmines.

1 INTRODUCTION

Today’s military is increasingly reliant on the use of

unmanned systems to perform a variety of missions

including surveillance, precision target designation,

mine detection, signals intelligence, and chemical-

biological-radiological-nuclear (CBRN)

reconnaissance, as described in the Office of the

Secretary of Defense FY2009–2034 Unmanned

Systems Integrated Roadmap (OSD, 2009). Many of

the programs, systems and technologies described in

OSD’s 25-year roadmap for unmanned systems

require the development of capabilities for

autonomous operations for teams of these systems

working together to execute missions. But

developing capabilities for controlling teams of

autonomous systems, and effectively utilizing these

teams to achieve mission objectives, presents many

technical challenges.

Fortunately, some of these challenges have been

addressed in biological systems such as colonies of

bacteria and social insects. In this paper we examine

quorum sensing in biology and propose its use as a

paradigm for implementing behavior-based control

that is decentralized, distributed, scalable, and robust

for teams of mobile autonomous systems.

Additionally, we propose that quorum sensing may

be used for ensemble decision-making tasks such as

collective classification in distributed autonomous

sensor platforms.

The increasing availability of autonomous and

unmanned vehicle platforms to military commanders

creates opportunities for the use of autonomous

vehicle teams to enhance situational awareness,

decrease response times, and gain tactical advantage

without increased risk to human life. Teams

composed of fully autonomous systems offer

potential to extend operational capabilities in critical

battlespace domains including the littoral zone,

undersea, space, on the battleground, and in other

challenging and hazardous environments. However,

coordinated command and control (C2) for teams of

195

Sofge D. and Lawless W..

QUORUM SENSING FOR COLLECTIVE ACTION AND DECISION-MAKING IN MOBILE AUTONOMOUS TEAMS .

DOI: 10.5220/0003122501950204

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 195-204

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

autonomous systems operating in complex, dynamic,

partially unknown and potentially hostile

environments offers both technological hurdles as

well as opportunities unique to each battlespace

domain.

In this technical effort quorum sensing is applied

to the development of behavior-based control for a

team of autonomous ground robots (agents) tasked

with searching a field for buried landmines. Each

robot is equipped with one or more sensors for

detecting mines, such as a metal detector,

magnetometer, or ground-penetrating radar. We

assume that the robots have different sensor suites

mounted on them to perform landmine detection,

and thereby may be considered heterogeneous from

a sensing and platform perspective. The goal is to

implement a decentralized control strategy such the

robots must collaborate in order to identify the

mines, while implicitly divvying up the labor

amongst available agents to provide efficient parallel

search.

2 BACKGROUND

AND RELATED WORK

Quorum sensing (QS) can be classified as a

decentralized decision-making process used to

coordinate behavior. The key characteristics of QS

are that each individual (1) senses either directly

(e.g. through molecular concentration) or indirectly

(e.g. by counting the number of interactions it has

with others of its kind) the number or density of its

own kind present, and (2) commences a

predetermined response, such as adopting a specific

behavior, once the quorum decision signal is

triggered. QS may also be part of a collective quality

assessment process, as practiced in ants and

honeybees in selection of new nesting sites. In such

cases the individuals make individual quality

assessments and then share these assessments with

the group.

2.1 QS in Bacteria

Bacteria achieve QS by detecting the density of

other bacteria in the area, and then using this signal

to regulate genes that in turn express behaviors (e.g.

swimming, biofilm construction, pathogenicity). QS

has been observed in many species of bacteria, but

has been studied extensively in only a few including

Vibrio fischeri, responsible for light production

(bioluminescence) in the Hawaiian bobtail squid; in

Escherichia coli (E. coli), which resides in the lower

intestinal tract and is often credited with causing

food poisoning in humans; and most extensively in

Pseudomonas aeruginosa, where QS has been found

to be employed in biofilm formation, swimming, and

cell aggregation.

Bacteria use signaling molecules called

autoinducers to regulate QS. These molecules are

continuously secreted and detected by the bacteria,

forming a kind of communication network within

the colony. Once a specific density threshold of

autoinducer molecules is crossed, behavior changes

are induced through changes in gene expression.

Bacteria also communicate between species,

using a different molecule to communicate than the

one used within their own species (Ng, 2009). It is

estimated that there are 10 times as many bacteria

present in the human body as there are cells within

the body. These bacteria, many species of which

have not yet been identified, play an integral role in

the proper functioning of the human body, while a

few can cause serious and even fatal diseases.

Identifying the chemical signaling mechanisms for

various species of bacteria, both for signaling within

species and between species, is an active and on-

going focus of bacteriology research (Mehta, 2009).

2.2 QS in Ants

While QS has been observed in a variety of social

insects, it has been studied most extensively in

Temnothorax albipennis and Leptothorax albipennis

ants (Pratt, 2002). Ant colonies nest in small

crevices between rocks, or inside small spaces inside

sticks. When the nest is broken, scout ants fan out in

search of a new nesting site. When a promising nest

site is found, the scout ant assesses the quality of the

new site, and returns to the old nest. She waits a

period of time inversely proportional to the quality

of the new site before recruiting nestmates to follow

her to the new nest site, a process called tandem

running.

Ants perform tandem running visiting many

candidate sites, recruiting other ants to visit the site

they have chosen to nominate. While these site

visits, recruiting, and tandem running are taking

place, the ants are sensing the number of encounters

they have with other ants. Once the number of

encounters reaches a threshold a quorum decision is

triggered and all of the ants return to the old nest and

begin carrying the brood, queen, and fellow ants to

the new nest site.

This process represents a more aggressive form

of QS than that employed by bacteria in that the

individuals compete to directly influence the

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

196

outcome of the collective decision, selecting a new

nesting site. The combination of individual site

quality assessment, recruitment, and voting (by their

presence) comprises an ensemble decision process

that provides a means for the colony to quickly and

efficiently find and relocate to a new nesting site.

2.3 QS in Bees

Honeybees (Apis mellifera) are social insects that,

like TA ants, utilize QS to collectively evaluate and

select new nest sites (Seeley, 2004). When a colony

of bees becomes too large, the queen will leave the

hive with a group of workers in order to start a new

hive elsewhere. Once outside the nest, the workers

form a swarm that may attach itself to a log, tree

branch, or other convenient location where it can

rest for a few days. A few of the scouts set off in

search of a nest site to house the new hive.

When a potential site has been found, the scout

assesses the quality in terms of whether or not it is

infested with ants, how protected it is from the

weather, how much sunlight it receives, etc., then

returns to the swarm and performs a waggle dance to

recruit other bees to the site. The number of

repetitions of the dance is proportional to the quality

of the site. Other scouts will fly to the potential new

nesting site and perform their own quality

assessments and recruiting. Once a quorum number

of bees has been reached at the new site they all

return to the swarm and begin a new behavior called

piping, causing the swarm to take off and relocate to

the new nest site.

Bees utilize a similar process to search for new

food sources, including random search, assessment,

waggle dance and recruitment. A search algorithm

called the Bees Algorithm inspired by this process

has been applied to a variety of combinatorial

optimization problems including server allocation

and job shop scheduling (Pham, 2005).

2.4 QS in Computational Intelligence

Quorum sensing clearly has potential for use in

applications of computational intelligence, but it has

surprisingly received little attention per se from the

artificial intelligence community. This may be due

in part to the fact that although QS was discovered

and studied in Vibrio fischeri in the late 1960s, for

many years it was thought to be limited to marine

bacteria such as Vibrio fischeri and Vibrio harveyi.

The extent to which bacteria utilize signaling to

achieve decentralized coordinated action was not

appreciated until recently (Ng, 2009).

A vibrant and expanding area of computational

intelligence research is based on modeling the

behavioral paradigms of social insects and applying

them to groups or teams of autonomous man-made

systems. Ant Colony Optimization (ACO) was

inspired by the movement of ants in locating food

sources, and the optimal paths they establish to

move the food back to the colony. ACO was

proposed by Dorigo (1992) as a search heuristic for

finding an optimal path in a graph, and has spawned

a class of heuristic algorithms for performing

optimization tasks. ACO algorithms may be

considered a subclass of stigmurgic methods

(Bonabeau, 1998) in which agents utilize

communication through the environment. Key

features of ACO algorithms include the use of

pheromones to create paths along which the ants (or

agents) move, and the processes for strengthening

and weakening such paths. ACO algorithms have

been applied to a variety of challenging optimization

tasks including the traveling salesman problem

(Dorigo, 1996), job shop scheduling (Merkle, 2002),

and distributed clustering (Bonabeau, 1998).

QS may be considered a stigmurgic method, but

it is not an ACO algorithm since it doesn’t use

pheromones, it doesn’t adapt agent paths to

gradually improve its solution(s), and the focus of

QS is emergent collaboration to achieve collective

decisions. As such, it would be more appropriate to

categorize QS as a method for distributed multiagent

collaboration rather than as an optimization

technique.

Sahin and Franks (2002) researched the

measurement of spaces by animals, including

Leptothorax albipennis ants, for potential use in

developing behaviors for autonomous mobile robots.

While the ants they studied were utilizing QS for

nest assessment they focused instead on the

mechanism for nest quality assessment, specifically

how they measured the size of the potential new

nesting sites. In their “Future Lines of Research”

section they discuss collective decision-making and

quorum sensing, and suggest exploring the use of

social behavior in complex measurements and

decision-making.

Wokoma (2003) proposed the use of a QS-based

protocol to provide self-organized clustering to

optimize communications routing in distributed

sensor networks. They conclude that the QS-based

protocol is more scalable than a centralized

approach, and can adjust to changes in the

environmental signal and network topology because

there is no dependence on any particular node.

Peysakhov and Regli (2005) proposed a server

QUORUM SENSING FOR COLLECTIVE ACTION AND DECISION-MAKING IN MOBILE AUTONOMOUS TEAMS

197

population management scheme for wireless mobile

ad hoc networks based upon QS, specifically

Leptothorax albipennis ants. They implement a QS

based protocol that automatically rebalances service

availability on server hardware configured as a

wireless server network. They conclude that the

solution exhibits properties of emergent stability,

decentralized control, and resilience to disturbances.

QS can be viewed as a simple form of voting, but

it differs from voting as practiced in ensemble

decision systems and human organizations in that in

QS no overall tally (counting) of votes is required,

the agents involved are necessarily mobile, and the

quorum decision is triggered based upon a density

threshold being exceeded. Related work in voting in

ensemble decision systems is discussed in the next

section.

2.4.1 Ensemble Decision-making

Recent advances in computational intelligence have

produced techniques and algorithms for combining

predictions, estimates, and decisions from multiple

sources, such as expert systems or neural network

models, such that the ensemble decision is at least as

good (and often significantly better) than that of any

one expert or model (Polikar, 2006). In human social

affairs we routinely practice ensemble decision-

making in numerous fora including elections, jury-

trials, product reviews and rankings, medical

treatment decisions (e.g., asking for a second or third

opinion before surgery), talent contests, and

scientific peer review. In the application of both

computational intelligence and human intelligence,

the use of ensemble decision-making allows the

individual to benefit from the knowledge and

experience of the group, and to thereby reduce the

risk of making poor decisions.

Ensemble decision systems (Polikar, 2006) have

been developed that use a population of decision

models to perform collective decision-making. The

strength of this approach is that if the model errors

are uncorrelated, then the overall ensemble decision

will be more accurate. Such systems often apply

voting schemes in which each classifier in the

ensemble is given a set of inputs and “votes” on the

classification. The votes are tallied, and a

combination rule is applied, such as majority or

consensus.

Biological systems such as ant colonies and other

social insect groups routinely demonstrate the ability

to coordinate information and collaborate in large

numbers to solve extremely challenging problems

collectively, such as building a new nest with

hundreds of complex interconnected chambers and

passages, or carrying objects many times the size

and weight of a single individual, despite the lack of

any form of centralized or coordinated planning or

control.

The study of such biological systems has

recently given rise to the field of Swarm Intelligence

(Garnier, 2007), which focuses specifically on the

emergence of intelligence through the interactions of

a large number of individuals, with each acting

according to its own behavioral plan. Swarm-

inspired behavior-based approaches to control of

teams of autonomous systems offer several

advantages over more traditional approaches (e.g.,

linearized optimal control (Robinett, 2010),

including robustness in dynamic environments,

decentralized and fully distributed controls, low

computational complexity (each individual is

executing a simple set of rules or behaviors), and

scalability since only local interactions are

considered (hence there is also no single point-of-

failure for the entire system).

In related efforts for controlling teams of

autonomous ground vehicles and teams of

unmanned air vehicles we utilize a physics-inspired

approach called physicomimetics (artificial physics)

(Spears, 2004; Wiegand, 2006) where vehicles are

modeled as particles, interactions between them are

governed by force laws, and observation goals are

represented by attractors. These techniques will

allow human operators to employ teams of

autonomous vehicles to perform missions and

provide information to enhance situational

awareness without increasing manning requirements

for planning and control of the platforms.

2.4.2 Behavior-based Autonomous Team

Control

In a related effort (Sofge, 2009) we are investigating

planning and navigation for teams of underwater

gliders to improve the accuracy of assimilative

ocean prediction models for undersea warfare. The

ocean environment presents numerous challenges for

unmanned systems such as difficulty communicating

with teammates underwater (increasing the need for

autonomy), difficulty localizing the vehicle

underwater and maintaining accurate positioning

(e.g., inertial navigation systems are highly sensitive

to drift due to currents and other ocean dynamics

and the resultant accumulation of error), and

difficulty controlling highly underactuated systems

such as undersea gliders. The undersea environment

also offers significant advantages for unmanned and

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

198

autonomous systems such as the low likelihood of

colliding with other objects (other than the bottom or

one another); stealth provided by ocean cover;

extended stay times for passive monitoring due to

the low energy need (since gliders use buoyancy

control to move forward, no propellers are required);

emerging technologies for harvesting energy from

the ocean; and long-range propagation of acoustic

signatures for identification.

In other efforts we are developing an information

theoretic approach to optimizing underwater

distributed sensor networks (DSNs), and algorithms

for merging bathymetric datasets (e.g., ocean floor

profiles) that have been collected at various times by

a variety of means including sonar arrays towed by

survey ships, side-scan sonar collected by undersea

vehicles, surveillance aircraft, and even space-based

observing platforms (satellites).

Autonomous sensor networks are under

development or have already been put into operation

for many purposes including weather forecasting

and prediction (Bell, 2010); volcanic gas emissions

monitoring (Galle, 2010); tsunami early-warning

systems (PTWC, 2010); monitoring bridges, tunnels,

pipelines, and other critical structures (Chebrolu,

2008); and for monitoring networking and

communications channels in order to detect possible

activity by terrorists (Lawless, 2010). A key aspect

of all of these networked systems, as well as the

teams of autonomous systems described previously,

is the need to coordinate information flows amongst

the individual members of the team (or within the

network); and to reconcile, fuse or merge, and

integrate the various bits of information streaming in

from disparate sources into a coherent picture for

use by human operators.

3 METHODOLOGY

The goal is to demonstrate that the quorum sensing

paradigm may provide robust decentralized

coordination and collective decision-making for

mobile autonomous teams performing complex

tasks. Quorum sensing is applied to the design of a

decentralized planner for a team of mobile

autonomous agents surveying a field containing

buried landmines. The key features of this approach

are (1) each agent only interacts with its local

environment, thus minimizing communication

requirements and avoiding complexity (and

bandwidth) scaling problems as the number of

agents increases, (2) collaboration between agents is

necessary to accomplish the task, both for collective

decision-making and division of labor, but it is an

emergent property (not explicit), (3) the approach is

robust to variations in the size (and topology) of the

field, number of targets (mines), number of agents,

sensor performance, and quorum size.

While QS is inspired by the behavior of social

insects such as ants, it relies on different

mechanisms than those employed by other artificial

ant algorithms such as ACO (described previously),

and QS is presented as a method to achieve

distributed collaboration and decision-making for a

team of agents, not as an optimization technique.

Therefore no direct comparison of ACO and QS

methods in performing this task was performed. An

ACO-based solution utilizing pheromone trails may

exist, but that is beyond the scope of this study.

Experiments instead focused on validating the QS

approach.

3.1 Operationalizing QS

Our approach to applying QS to a target domain,

such as a search or optimization problem involving

multiple agents, is to first decompose the problem

into two or more distinct phases. Each phase is

characterized by parallel execution of agent

behaviors, with no centralized control of the team.

The first phase is fundamentally a parallel search

by the agents. Each agent must be capable of

performing a quality assessment or recognition of

whatever is being sought.

Next, each agent must have a mechanism for

communicating or expressing its assessment or

recognition. This could be communicated through

the environment with autoinducers in bacteria, or

communicated directly from one agent to another by

ants and bees. Recognition functions as a voting

mechanism in bacteria, while assessment is part of

the recruitment process in social insects. This

communication of assessment is the key to collective

decision-making.

Each agent must also have the ability to trigger

the quorum decision (since we require a distributed,

completely decentralized approach). The quorum

decision must be accepted by each agent. Once the

quorum decision state is accepted, the agent may go

back into another state such as random-walk, or

search, depending upon the task.

3.2 Area Coverage

Using a team of autonomous vehicles to search a

field at first glance appears to be a classic area

coverage problem. Such problems may often be

QUORUM SENSING FOR COLLECTIVE ACTION AND DECISION-MAKING IN MOBILE AUTONOMOUS TEAMS

199

solved efficiently by dividing the space to be

searched amongst the available agents, and having

each agent assume responsibility for covering a

specific area.

In this application, however, this approach will

not work since each agent requires the assistance of

other agents carrying different sensor packages to

confirm the identification of buried mines. A

Brownian-motion type random walk (with a single

random step taken at each time step) would not

make sense either, since the agents would spend an

inordinate amount of time retracing their steps and

revisiting the same places they had just visited.

While many strategies may be devised to address

this problem, we chose to start with a modified

random walk, with “walk-length” determined by the

size of the field and the number of agents. The walk-

length for the agents is determined by dividing the

field length by the number of agents and then

multiplying by three. For example, if the field is

50x50, and there are 10 agents, then the walk-length

will be 15. This gives each agent good field

coverage, but no attempt was made to optimize

walk-length with respect to overall team

performance. Each agent chooses a random direction

to move from its starting position on the grid by

selecting a direction toward one of its 8 neighbors

(standing still is not allowed). For each time step

that passes it will continue in that direction until the

full walk-length has been covered (e.g., 15 steps).

The agents then select another direction at random.

3.3 The Field, Agents and Mines

The field is implemented as a square cellular

toroidal grid (for simplicity) such that each agent

and each mine is located at a specific Cartesian grid

coordinate at each point in time. Both the mines and

the agents are randomly placed on the grid at the

beginning of each run. The agents will move while

the mines will remain in fixed locations throughout

the run.

Agents appear on the grid as small blue circles.

The mines initially appear as small red stars. Each

time an agent “recognizes” a mine, its star grows a

bit larger on the field until the quorum decision

threshold is crossed, at which point it is changed into

a green square (Figure 1).

3.4 Mine Detection

and Quorum Decisions

The QS paradigm requires that each agent must (1)

sense the number (or density) of its own kind, and

(2) commence a predetermined response once a

quorum decision threshold has been crossed. Since it

would be extremely inefficient to have all (or many)

of the agents congregate at each mine, we decided

that only a minimum requisite number (a quorum, by

definition) must visit each mine and mark it as

“recognized”. In addition, a mine can only be

recognized by an agent once. Recognition is a

stochastic process based upon the maximum sensing

range (a length of 3 cells was used for the

experiments), the accuracy of each agent (also a

controlled parameter), and a normally distributed

random number generated for each possible

recognition.

When an agent successfully recognizes a mine,

the mine’s hit-count is incremented. If the hit-count

exceeds the quorum threshold, the mine is

announced as recognized and its icon is converted

from a red star to a green square. The threshold is

the same for all mines and agents, and the requisite

action upon recognition (that is, a mine exceeding

the threshold) is to announce the mine (in an actual

real-world situation the presence and location of the

mine would be broadcast for further investigation

and/or remediation of the mine), and continue

searching for other mines. Once all of the mines are

located the simulation is stopped and the time taken

to find all of the mines is recorded.

3.5 Team Performance and Robustness

The performance of each agent team is measured

based upon the number of simulation time steps

from initialization of landmine and agent positions

until all landmines are recognized by a quorum of

agents. Since the starting positions, random-walk

process, and recognition process are stochastic, each

experiment is repeated 100 times. The number of

time steps required for each run is plotted versus the

variable of interest, along with the median over all

100 runs.

We define robustness as the property that the

QS-inspired search strategy will continue to function

effectively in the presence of changes in the

parameters such as field size, # Agents, Quorum

Number, etc., and that team performance degrades

gracefully with increases in task complexity (e.g. by

increasing number of mines, or decreasing number

of agents). To avoid undue influence by pathological

starting conditions we calculate the 5% trimmed

mean over each 100 runs, shown as the red lines in

Figures 2-6.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

200

Figure 1: Screen shot from Matlab simulation of quorum

sensing-based controller showing agents (blue circles)

searching a 50x50 toroidal grid for landmines. There are

10 mines on the grid and 10 agents. Hidden mines are

shown as red stars, and recognized mines are shown as

green squares.

4 EXPERIMENTS

The focus of the experiments was validation of the

QS approach as a method to achieve distributed

collaboration and decision-making for a team of

agents. It is not presented as an optimization

technique, nor is it suggested that this is the only

distributed algorithm for accomplishing this task.

The efficiency of this algorithm compared with other

fully decentralized approaches is beyond the scope

of this paper. Instead we demonstrate the robustness

of QS with respect to variations in the problem

domain and resources as described.

4.1 Expectations and Hypotheses

Our expectations were that QS could be

operationalized for use in decentralized, distributed

mobile multiagent teams to address challenging

problems in computational intelligence such as

decentralized landmine detection using an

autonomous multi-agent team. We hypothesize that

the QS-based approach will be robust to variations

in many of the variables including:

 field size

 number of mines

 number of agents

 sensor performance

 quorum size

4.2 Design of Experiments

The experiments were designed to demonstrate the

application of QS to the mine detection task, and to

test the hypotheses specified above. The measure of

performance for the team was the number of

simulation time steps until all of the mines were

recognized (i.e., quorum number reached for every

mine). Each experiment started by randomly

initializing the positions of the mines and the

positions of the agents.

Each experiment was repeated 100 times with a

different set of starting positions each time. The

parameters not being varied were set at the

following nominal values:

Field size: 50x50 # Mines: 10

# Agents: 10 Walk-length: 3*Field size/#Agents

Quorum Number: 3 Sensor Performance: 0.6

Experiment 1: Test robustness of team performance

with respect to variations in field size.

The field was defined as an mXm unit square

toroidal cellular grid. The value of m was varied

from 20 to 100 in steps of 5 units.

Experiment 2: Test robustness of team performance

with respect to variation in the number of mines.

The number of mines was varied from 5 to 25 in

steps of size 1.

Experiment 3: Test robustness of team performance

with respect to variations in the team size (number

of agents).

The number of agents was varied from 5 to 25 in

steps of size 1.

Experiment 4: Test robustness of team performance

with respect to variations in sensor accuracy.

It was assumed that different sensors have different

performance characteristics in recognizing the

mines. The detection of a mine is modeled as a

stochastic process in which sensor accuracy,

distance to the mine, and random chance determine

the outcome. For all experiments the maximum

range for detection was set at 3 units, with the

probability of detection decreasing with distance

according to a normal distribution with mean zero.

The variance of that distribution was determined by

sensor performance parameter psense. The nominal

value of psense for all experiments except for

Experiment 4 was 0.6.

To test the robustness of team performance with

respect to psense, the value of psense was varied

from 0.3 to 1.0 in steps of size 0.1

Experiment 5: Test robustness of team performance

QUORUM SENSING FOR COLLECTIVE ACTION AND DECISION-MAKING IN MOBILE AUTONOMOUS TEAMS

201

with respect to variations in the quorum size (the

number of hits required for a mine to be fully

recognized by the team).

The quorum size was varied from 2 to 5 in steps of

size 1.

5 RESULTS

Figures 2-6 show the results from the Experiments.

Each experiment was repeated 100 times. Each dot

shows time to completion for the team for a single

run; the line shows the 5% trimmed mean

completion times for the team versus the varied

parameter (shown on the x-axis) over 100 runs.

Figure 2 shows that as the field size increases

from 20X20 to 50X50, a 625% increase in area, the

task completion time increases linearly at

approximately the same rate as the increase in area.

This shows that the QS-inspired search technique is

robust to changes in field size.

Figure 2: Experiment 1: Task Completion Time (y-axis)

vs. Field Size (x-axis).

Figure 3 shows that at the number of mines increases

from 5 to 25, the trimmed mean task completion

time also increases at a gradual (roughly linear) rate,

indicating that the technique is robust to changes in

the number of targets.

Figure 4 shows that as the number of agents

increases from 5 to 25, the task completion time

decreases monotonically but non-linearly. This is as

expected, since the number of targets is fixed at 10,

adding additional agents after a certain point will not

substantially reduce the search time. This also shows

that the technique is robust to changes in team size.

Figure 3: Experiment 2: Task Completion Time (y-axis)

vs. Number of Mines (x-axis).

Figure 4: Experiment 3: Task Completion Time (y-axis)

vs. Number of Agents (x-axis).

Figure 5: Experiment 4: Task Completion Time (y-axis)

vs. Sensor Performance (x-axis).

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

202

Figure 5 shows that as the sensor performance

increases from 0.3 to 1.0, the mean task completion

time decreases monotonically, indicating that the

technique is robust to changes in sensor

performance.

Figure 6: Experiment 5: Task Completion Time (y-axis)

vs. Quorum Number (x-axis).

Figure 6 shows that as the quorum number

increases from 2 to 5, the mean task completion time

increases linearly and monotonically, indicating that

the technique is robust to changes in quorum

number, the number of collaborators needed to

confirm a decision.

6 CONCLUSIONS

Quorum sensing, a decentralized decision-making

process used by bacteria and by social insects to

coordinate group behavior and to perform collective

decision-making, provides a robust decentralized

team coordination and collective decision-making

paradigm for use in mobile autonomous teams

performing complex tasks. In this effort a quorum

sensing paradigm was used to develop a behavior-

based control strategy for a team of autonomous

mobile robots given the task of surveying a field

containing buried landmines. The quorum sensing-

based search strategy was shown to be robust to

variations in field and team size, number of

landmines, sensor accuracy and quorum size.

7 FUTURE WORK

Quorum sensing offers tremendous potential for

design of robust decentralized control and decision-

making strategies for teams of autonomous systems

and distributed sensing arrays. Mobile autonomous

systems capable of collaboration may provide

significantly enhanced capabilities for recognizing

targets, area search, reconnaissance, and other

critical tasks. Future efforts will focus on refining

QS-inspired approaches to collaborative tasks for

multi-agent teams (such as area search and collective

recognition), implementing these methods on actual

autonomous system hardware, and testing

autonomous teams under real-world conditions. The

form of quorum sensing implemented and studied in

this effort thus far is passive, much like quorum

sensing employed by bacterial colonies, in that the

agents do not practice recruitment to confirm their

classifications. A more advanced form of QS,

Aggressive Quorum Sensing (AQS), akin to that

employed by ants and honeybees, incorporates

recruitment and more behavior states for the agents.

Once an agent makes a successful recognition of a

mine (but still below the quorum threshold), it

begins recruiting other agents to confirm the

recognition. This technique has the potential to

significantly enhance the accuracy of the team on

the mine-clearing task. We will develop AQS and

apply it to the landmine surveying task, comparing

the performance of the AQS-based approach with

that of the QS-based approach described herein.

ACKNOWLEDGEMENTS

This work was performed at the Naval Research

Laboratory and was funded by the US Department

of Defense, Office of Naval Research under the

project titled Heterogeneous Perceptual Sensor

Swarms for Landmine Detection. The views,

positions and conclusions expressed herein reflect

only the authors’ opinions and expressly do not

reflect those of the US Department of Defense,

Office of Naval Research, or the Naval Research

Laboratory.

REFERENCES

Bell, B., 2010. Mobile Meteorological Observation (M2O)

Network. Second AMS Conference on International

Cooperation in the Earth System Sciences and

Services, <http://ams.confex.com/ams/90annual/tech

program/programexpanded_603> {retrieved October

26, 2010}.

Bonabeau, E., Dorigo, M., Theraulaz, G., 1999. Swarm

Intelligence: From Natural to Artificial Systems.

QUORUM SENSING FOR COLLECTIVE ACTION AND DECISION-MAKING IN MOBILE AUTONOMOUS TEAMS

203

Oxford University Press.

Chebrolu, K., Raman, B., Mishra, N., Valiveti, P. K.,

Kumar, R., 2008. BriMon: A Sensor Network System

for Railway Bridge Monitoring. In Proceedings of the

Sixth International Conference on Mobile Systems,

Applications, and Services (MobiSys), Breckenridge,

Colorado.

Dorigo, M., 1992. Optimization, Learning and Natural

Algorithms. PhD thesis, Politecnico di Milano, Italie.

Dorigo, M., V. Maniezzo, V., Colorni, A., 1996. Ant

System: Optimization by a colony of cooperating

agents. In IEEE Transactions on Systems, Man, and

Cybernetics – Part B, 26:1:29–41, IEEE.

Galle, B., Johansson, M., Rivera, C., Zhang, Y., Kihlman,

M., Kern, C., Lehmann, T., Platt, U., Arellano, S.,

Hidalgo, S., 2010. Network for Observation of

Volcanic and Atmospheric Change (NOVAC)—A

global network for volcanic gas monitoring: Network

layout and instrument description. In J. Geophysical

Research 115, D05304.

Garnier, S., Gautrais, J., Theraulaz, G., 2007. The

Biological Principles of Swarm Intelligence. In Swarm

Intelligence, 1:3–31, Springer.

Lawless, W. F., Rifkin, S., Sofge, D., Hobbs, S.,

Angjellari-Dajci, F., Chaudron, L., Wood, J., 2010.

Conservation of Information (COI): Reverse

Engineering Dark Social Systems (DSS). In Structure

and Dynamics: eJournal of Anthropological and

Related Sciences, University of California

eScholarship Repository.

Mehta, P., Goyal, S., Long, T., Bassler, B., Wingreen, N.,

2009. Information processing and signal integration in

bacterial quorum sensing. In Molecular Systems

Biology, 5:325.

Merkle, D., Middendorf, M., Schmeck, H., 2002. Ant

colony optimization for resource-constrained project

scheduling. In IEEE Transactions on Evolutionary

Computation, 6:4:333–346, IEEE.

Ng, W., Bassler, B., 2009. Bacterial Quorum-Sensing

Network Architectures. In Annual Review of Genetics,

43:197–222.

OSD, 2009. 2009-2034 Unmanned Systems Integrated

Roadmap. U.S. Office of the Secretary of Defense.

Peysakhov, M., Regli, W., 2005. Ant inspired server

population management in a service based computing

environment. In Proceedings Swarm Intelligence

Symposium, pp. 357–364, IEEE.

Pham, D. T., Ghanbarzadeh, A., Koc, E., Otri, S., Rahim,

S., Zaidi, M., 2005. The Bees Algorithm. Technical

Note, Manufacturing Engineering Centre, Cardiff

University, UK.

Polikar, R., 2006. Ensemble based systems in decision

making. In IEEE Circuits and Systems Magazine,

6:21–45, IEEE.

Pratt, S., Mallon, E., Sumpter, D., Franks, N., 2002.

Quorum Sensing, Recruitment, and Collective

Decision-Making During Colony Emigration by the

Ant Leptothorax Albipennis. In Behavioral Ecology

Sociobiology, 52:117–131, Springer-Verlag.

PTWC (The Pacific Tsunami Warning Center), The

Tsunami Warning System <http://www.geophys.wash

ington.edu/tsunami/general/warning/warning.html#ov

erview> {retrieved October 26, 2010}.

Robinett, R. III, Wilson, D., 2010. Hamiltonian surface

shaping with information theory and exergy/entropy

control for collective plume tracing. In International

Journal of Systems, Control and Communications,

2:1–3:144–169, InderScience Publishing.

Sahin, E., Franks, N., 2002. Measurement of space: From

ants to robots. In Proceedings of WGW 2002:

EPSRC/BBSRC International Workshop Biologically-

Inspired Robotics: The Legacy of W. Grey Walter, pp.

241–247, Bristol, UK.

Seeley, T., P. Visscher, 2004. Quorum Sensing During

Nest-Site Selection by Honeybee Swarm. In

Behavioral Ecology Sociobiology, 56:594–601,

Springer-Verlag.

Sofge, D., Harper, M., 2009. Autonomous Route Planning

and Navigation for UxV Teams. ONRG Workshop on

Machine Intelligence for Autonomous Operations,

Lerici, Italy.

Spears, W., Spears, D., Hamann, J., Heil, R., 2004.

Distributed, Physics-Based Control of Swarms of

Vehicles. In Autonomous Robots, 17:2–3:137–162,

Springer Netherlands.

Wiegand, R., Potter, M., Sofge, D., Spears, W., 2006. A

Generalized Graph-Based Method for Engineering

Swarm Solutions to Multiagent Problems. In Parallel

Problem Solving from Nature - PPSN IX, Lecture

Notes in Computer Science, pp. 741–750, Springer-

Berlin.

Wokoma, I., Sacks, L., Marshall, I., 2003. Clustering in

Sensor Networks using Quorum Sensing. In London

Communications Symposium, University College

London.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

204