Você está na página 1de 13

1 Introduction

1.1 Problem Addressed and Motivation

The principal problem addressed is Sensor Clustering in wireless sensor networks. A sensor
network can be visualized as a geometric graph with nodes as sensors and weighted arcs as
communications links. The throughput of a node of an unstructured sensor network of randomly
located nodes with fixed transmission range is limited [1]. To maximize sensor network
performance metrics, in the face of limited power, ad hoc deployment, and mobility, one must
structure the network. Sensor node clustering is an important element of such structuring.
Generally, clustering is necessary for

• Data aggregation
• Hierarchical routing
• Optimizing sleep patterns or election of extremal sensors [2]
• Optimizing coverage
• Resource allocation: reuse of frequency bands, codes and transmission schedules
• Conserving energy

We provide a simple example to illustrate the importance of clustering in sensor networks.

Suppose a sensor network is deployed for continuous monitoring of an environmental quantity


for the purpose of generating a map of the variation of the quantity in the environment and
relaying the map at periodic intervals to a base station. In such a scenario, one possible
implementation would be for the base station to communicate directly with all of the sensors
nodes. Having all nodes communicating over long ranges will quickly deplete the energy stores
of the nodes. Nodes far from the base station will die first and the area being monitored will no
longer be coincident with the original target area. Energy is not conserved by the nodes. They
wastefully communicate over long distances when it would be beneficial to have many smaller
communication schedules distributed throughout the network with nodes communicating over
distances that are much less than the full spatial extent of the sensor network. These
communication cells are clusters. Clustering can thus promote energy conservation.

Clustering can promote coverage maximization. It is desirable for the sensor network to be able
to monitor consistently and uniformly over the entire target area for the maximum time. To
evenly distribute energy dissipation by the nodes so that they all die at about the same time,
energy intensive roles, like the role of nodes that make up a routing backbone, must be shared
among nodes in the sensor network. When a node takes on this special role it is called a
clusterhead.

Clustering can promote resource conservation. With all nodes either synchronizing to the same
communication schedule or sharing the same frequency band or codes, the efficiency of the
communication protocol will degrade for larger sensor networks. Reusing frequency bands or
codes within spatially distinct clusters or synchronizing schedules of logically related nodes may
conserve resources.
Clustering makes efficient, intelligent data aggregation possible. Sensor networks are unique in
that they may gather information that is correlated for nearby nodes. Data compression or
aggregation at clusterheads reduces the amount of data to be injected into the routing backbone.
For efficient or maximum compression, it is necessary to form logical clusters of nodes whose
data are most closely correlated.

Efficient routing is not possible without clustering. All nodes would then be a single hop from
the base station. It is possible to enable routing without the use of clusters, but each node would
need a complete routing table which would not scale well to large numbers of sensor nodes. To
maintain scalability, clusters of nodes are formed to reduce the size of the node’s routing tables.
Thus clustering is necessary for scalable routing of data.

1.2 Objectives

The objectives of the proposed research are:


• Develop efficient, distributed, adaptive, clustering approaches that can be applied to static
and mobile wireless sensor networks of varying size and spatial extent (robust and
scalable)
• Investigate local interactions of sensor nodes for global parameter learning.
The global parameters to learn may be related, but not limited to position, routing or data
aggregation clustering

2 Overview of literature

Much theoretical work has already been published on clustering in the graph theoretical
framework [3]-[10]. Most of the work done on clustering has been in the context of developing
efficient routing protocols for wireless, ad hoc, sensor networks. A good survey of routing
techniques for wireless networks can be found in [11]. But clustering need not be for the benefit
of routing alone. If sensed data is correlated over an area covered by multiple sensors, data
aggregation and compression can reduce the number of packets that need to be forwarded to the
end user.

Clustering is now a widely accepted method for optimizing sensor network performance [12]. To
perform efficiently, sensor networks must often optimize global parameters without global
knowledge and control. Examples of optimization objectives relevant to sensor networks include

• Maximizing data packets extracted from the sensor network over its lifetime.
• Maximizing the sensor network’s effective lifetime.
• Minimizing delay between a user query and a sensor network response.
• Maximizing sensor coverage.

The clustering problem can be viewed as a search through a typically NP-Hard solution space.
Clustering to minimize the distance between a node and its clusterhead [13] is one possible
clustering. The quantity to optimize varies with application and some common quantities were
listed in the bullets above. The problem is that although a global quantity optimization is
required, nodes must participate in solving the problem using only local interactions.
The literature dealing with wireless, ad hoc networks is replete with proposed clustering
solutions. Each approach is typically tailored to a subset of the overall problem of optimizing the
operation of a sensor network. One approach may suffice to provide an excellent routing
hierarchy while another may provide for energy efficient data aggregation but none provide a
general framework for optimal clustering. Nearly all clustering approaches are algorithmic
implementations of graph partitioning. Nearly all clustering approaches are reactive and rule
based. None of the clustering approaches appears to be able to dynamically respond to varying
sensing range, communication range, and data correlation length scales.

The lowest-ID Algorithm [14, 15, 16] was devised to introduce Quality of Service guarantees for
multimedia application support, which is a little removed from a simple view of a sensor
network, but the underlying reason incorporates a feature that is important to sensor networks as
well, namely resource reuse. If communication is localized, frequency bands and codes can be
reused in spatially distinct areas of the sensor network. Clusters produced contain nodes that are
at most 2 hops away from other nodes in the cluster. There are several limitations of this
clustering technique. In a static network, only one clustering scheme will be calculated unless
channels fade or nodes die. Clusterheads can be produced using this clustering scheme but
cannot be allocated in such a way as to maximize quantities of interest in a sensor network.
Specifically, if clusterheads are assigned, node die-off will be ongoing which is undesirable in
many envisioned sensor network applications. This algorithm requires either hardware
implementation of node IDs or a startup phase in which node IDs are propagated through out the
network. In the latter case, the protocol would be complicated when introducing new nodes into
the sensor network to replenish spent and failed ones. Since the algorithm operates through
interactions between nodes within communication range, the range of the nodes in relation to the
overall spatial extent of the sensor network impacts performance. In the version of this clustering
where clusterheads are not assigned, each node must maintain its own routing information and
the algorithm will not scale well to large sensor networks.

The highest-connectivity [17, 18] algorithm assigns as clusterheads those nodes with the most 1-
hop neighbors. This algorithm suffers from the same weaknesses as the lowest-ID algorithm and
was shown [18] to re-cluster more frequently for highly mobile nodes. The Max-Min D-Clusters
[19] algorithm extends the lowest-ID and highest connectivity algorithms in which all nodes
were at most 1-hop from their clusterhead, to a more general case where all nodes are at most d-
hops from their clusterhead. Although this algorithm is an improvement in that clusterhead roles
can be rotated, no sensor network optimization is occurring. There is no data aggregation model
built into the clustering. The Maximum Weight Independent Set or Generalized Clustering
Algorithm [20] is a generalization of the lowest-ID and highest-connectivity algorithms. The
claim of this algorithm is that properly chosen weights can be used to adjust clusterhead
selection but there is currently no way to dynamically and adaptively adjust these weights to
optimize the sensor network. The key results of the Landmark Clusters, hierarchy management
algorithm [21, 22] are clusterheads/routers with 3 to 4 children in the steady state. Each
landmark will have a minimum of 1 child and a maximum of 7. Here again, the algorithm is
specifically designed for routing management and so does not incorporate energy considerations
that are so important to the deployed sensor network. The On-Demand Weighted Clustering
Algorithm [23] is a generalization of Maximum Weight Independent Set or Generalized
Clustering Algorithm. Weights are generalized to represent node properties like connectivity,
energy level, mobility and transmission power. This algorithm shows the most promise at being
applicable to sensor networks but performance of the algorithm is still impacted by the
transmission power of the nodes.

The low-energy adaptive clustering hierarchy protocol (LEACH) [24] is an approach designed
for energy optimization and data aggregation. Each node in the sensor network randomly
determines whether it will take on the clusterhead role. The probability that the node will
announce as a clusterhead depends on whether it has been a clusterhead recently and a system
parameter k , which is the expected number of clusters per round of the protocol. This system
parameter depends on the total number of clusters in the sensor network, the linear spatial extent
on the sensor network and a energy dissipation model for the clusterhead nodes. However, the
protocol is not entirely distributed and autonomous because the protocol cannot scale adaptively
to different combinations of system parameters that determine k .

The Random Competition Based Clustering [25] algorithm is a heuristic approach with a random
aspect akin to the LEACH protocol but is not designed for optimizing parameters of interest to a
sensor network. It is similar to the lowest-ID and highest-connectivity algorithms but is shown to
perform better. The Power Control Clustering [26] approach attempts to improve stability of
least-ID 2–hop algorithm by adding power control. Passive Clustering [27] improves the stability
of the least-ID 2–hop algorithm. Although it performs very well for ad hoc, mobile wireless
networks, it does not address the special needs of sensor networks. Although sleep patterns are
addressed in [28, 29, 30] for the purposes of energy conservation, the implementation is targeted
for applications where a dense sensor network is deployed for single event detection. The
algorithm is therefore special use and not generally adaptable.

Overall, the existing clustering techniques can be grouped as follows:

• Heuristic with target constraints and no optimality guarantees


• Graph Partitioning – greedy search partitioning

Most existing techniques fail to fully explore the special needs of sensor networks. Sensor
networks have at least three important system parameters that are not specifically addressed in
current clustering schemes.

• Communication range
• Sensing range
• Data correlation length scale

This itemization is only for a simple case of scattered sensors measuring a local quantity of
interest. If the sensor network has other functions, such as triangulation, or coverage of an area
for event detection, then other system parameters are introduced. The sensor network must be
able to address optimization issues at the clustering level. The review of current clustering
approaches reveals that many of them are not designed to optimize the energy efficient operation
of the sensor network.
Throughout the literature, biologically inspired search and optimization methods, as applied to
clustering for wireless sensor networks are under-represented. There is a need to investigate
whether, task recruitment like in ant foraging, or social style interaction and learning, can be
applied to the open problem of optimally clustering wireless sensor networks.

3 Approach – Biologically Inspired Approaches

The investigation will focus on developing solutions to the proposed problems building on
systems derived from biological paradigms. Biological systems are extremum machines.
Colonies of bacteria forage and grow to maximize their nutrition intake, in effect maximizing the
colony’s survival probability. Ants recruit nest mates to assist in cooperative movement of prey
that would be unmovable for a single ant. Our immune system minimizes the impact of a foreign
agent in the body. People socialize to propagate good ideas and refute bad ones. Since optimal
clustering is typically an NP-hard problem [13], biological systems, which by nature are
optimizing systems, offer promising avenues of exploration for discovering new approaches to
clustering algorithm and protocol development for wireless, ad hoc, sensor networks.

3.1 Particle Swarm Optimization

NP-hard problem solutions exist in a vast fitness landscape. The shape of the landscape will
determine which search techniques will perform well for a given problem. For a single hill
landscape, a simple gradient following algorithm will suffice. For nearly all other NP-hard
problems of interest, the fitness landscape is sufficiently complex for gradient following to
converge to a local extremum instead of the global extremum. A more adaptive search must be
employed to circumvent local extrema. Particle Swarm Optimization [31, 32] is such a
technique.

Search techniques, such as Random Search [33], Tabu Search [34], Simulated Annealing [35]
can be implemented as an evolution of a single test solution. Genetic Algorithms [36] and
Particle Swarm Optimization evolve populations of test solutions. Particle Swarm Optimization
relies on a pool or swarm of test solutions interacting with each other to encourage both, global
exploration of new areas of fitness space, as well as more thorough searching of promising local
areas. The Particle Swarm Optimization technique is an evolutionary programming technique
with a motivation in biological systems, namely us.

In can be argued that people adjust their attitudes, actions, and beliefs based on their own
behaviors and those of others. By observing rewards received individually, and rewards received
by others, an individual can adjust its behavior to emulate the more successful individual.
Therefore, an individual moves through a search space of individual fitnesses, by reacting to its
own cognitive component of its motion and a social component inferred by observing others.
The observation of others is an important aspect of particle swarm optimization. It implies
communication between neighboring individuals. It implies that individuals are able to exchange
information on how fit they are as well as the values they internalize that allowed them to
achieve that fitness. This is where one can envision extension of the Particle Swarm
Optimization technique to the clustering of wireless sensors. More generally, the local
neighborhood interactions of sensor nodes may be employed to search a fitness landscape that is
of interest to the whole sensor network.

Particle Swarm Optimization is a topic of current interest and is refined and applied in novel
ways. In [37], the mathematical behavior of a particle in the swarm is analyzed, resulting in an
analytical formulation for stabilizing the swarm. An off-the-shelf version of the method is
formulated in [38] by running experiments with varying population sizes, neighborhoods sizes
and other parameters. A discrete version of the Particle Swarm Algorithm is presented in [39]. A
good review of applications of Particle Swarm Optimization [40] details “function stretching” to
alleviate local minima, experiments on min max problems, multiple minima problems, and
multi-objective function problems. Researchers are exploring using dynamic neighborhoods [41]
and intelligent particles [42]. The Particle Swarm Optimization is clearly a novel and young area
of research capable of yielding useful results in areas from economics to engineering.

The PSO approach can be applied to sensor networks. Individual sensors may internalize a
measure of their fitness. This measure could correspond to the sensor’s battery reserve for
example but in a real implementation would likely be a measure of more than one sensor
attribute. Periodically, as sensors communicate with neighbors, fitness values could be shared
with neighboring sensors. As long as sensors can interpret other sensors’ fitness values in terms
of their own operational parameters, it is possible to have sensors adjust their operational
parameters in favor of a socially measured better value.

3.2 Ant Systems

Colonies of swarming insects, ants, termites and bees have inspired global optimization
algorithms based on local interactions [43]. The basic idea is that insect behavior is determined
in part by environmental cues. Those same environmental cues are affected by the insect
behavior. This feedback loop can lead to emergent global system properties. Ant swarms have
been applied to classical problems, like the traveling salesman problem [44], network routing
[45] and document clustering [46]. For sensor networks, information packets could be interpreted
as the swarming agents/ants. The senor nodes take on the role of environmental cells that
packets can react to and modify. Nodes could use locally deposited information from network
packets to identify clustering parameters.

Mathematical formulation of stigmergic swarms has been studied [47]. Within that framework,
one can identify simulation spaces where emergent behavior is organized and simulation spaces
where the emergent behavior is chaotic. The transition criteria, or edge of chaos [48] can be
derived and interesting behaviors, like trail forming [49], are observed to emerge near the phase
transition. It may be possible to apply the same approaches to clustering in sensor networks to
derive useful quantities that may reveal whether or not a clustering strategy is stable.
Additionally, the stability of the clustering may be characterized as a function of system
parameters.
3.3 Human Immune System

Imitating biological systems has produced several novel computational methodologies such as
genetic algorithms, neural networks, and immune engineering that are useful in solving complex
engineering problems [50]. The natural immune system has proven to be a powerful inspiration
for computer-based solutions [51]. Over the last few years there has been an increasing interest
in the area of Artificial Immune Systems (AIS). AIS use immunological properties to develop
systems that are capable of performing tasks in various engineering applications [52]. The
human immune system is a complex system consisting of cells and molecules distributed
throughout our body, providing us with a basic natural defense mechanism. By defense, it is
meant that the immune system protects our body from infectious agents such as viruses, bacteria,
fungi, and other parasites [53]. Some of the remarkable properties of the immune system include
pattern recognition, sophisticated information processing, autonomous decentralized control,
adaptive learning and memory [50, 52, 53]. AIS have solved several problems that are not
limited to optimization, computer security, neural network approaches, data mining, robotics,
image segmentation, image classification and autonomous agents [54 – 61]. The natural immune
system is a very complex system with several mechanisms for defense against infectious agents
entering our system. In immunology, these agents are called antigens. The immune system
reacts by producing antibodies that attack these antigens. In addition to reacting to the foreign
agents that come from outside and inside the living system, the immune system also plays an
important role to maintain its own system against dynamically changing environments. The
immune system contains thousands of different cells interconnected in a network that serves as a
communication channel [52].

The basic components of the immune system are the white blood cells, called self-cells or
lymphocytes in immunological terms. These specialized cells are classified into two types
namely the B- lymphocytes and T lymphocytes. B-lymphocytes are the cells produced by the
bone marrows. T-lymphocytes develop in the bone marrow and mature in the thymus. The
external components to the immune system are antigens, called the non-self cells, as they are
foreign substances to the body. The major responsibility of the B cells is the generation of
receptors called antibodies (Ab) as a response to the antigens (Ag) [62]. The role of these
receptors on the surface of the B cell is to recognize and bind the antigen. These receptors are
called idiotopes and paratopes. Antigens also have receptors called epitopes. The B cell
receptors are complementary to the epitopes of the antigens. The recognition of the antigenic
pattern is achieved by this shape matching between the receptor of the B cell and the epitope of
the antigen [53, 62]. Figure 1 (a) shows the B cells, B cell receptors, antigen, and epitopes.

The main function of the immune system is to kill the antigen, the common goal. It is interesting
to note that antibodies work together to accomplish the common goal even though individually
they do not posses a higher intelligence. This collective behavior makes the immune system a
distributed and decentralized system. The immune system has the following properties: Self/Non
– self discrimination [52, 62], Affinity maturation [3], Clonal selection and expansion [62,63,
64], Immune memory [62], and Immune Network [65, 66]. The immune network presents the
best opportunity to improve the performance of a sensor network even though other properties
may have the same effect. If we think B cells as sensor nodes in a sensor network, it may be
possible to create protocols and/or algorithms based on immune network properties to improve
the data aggregation and self-organization of the sensor network.

B2

Ag Id2 #3

Epitope of the antigen B1 P2


Ab2
B cell Id1 #3
Antige B3 P1
n Ab1 Stimulation
Id3 #3 Suppression
B cell receptor P1, P2, P3 – Paratopes
P3 Id1, Id2, Id3 – Idiotopes
Ab3
Ab1, Ab2, Ab3 - Antibodies
(a) (b)

Figure 1. (a) Components of immune system (b) Jerne’s Immune Network.

Immune Network and its Applicability to Sensor Networks


Jerne’s idiotropic network hypothesis [65, 66] as proposed by the immunologist deals with the
interaction of the antibodies. Jerne’s immune network is a network of B cells that communicate
the shape of the antigenic epitope through idiotopes and paratopes. This communication alters
the structure of the receptors according to the antigenic pattern. This shape transformation is
important for information transfer and communication between the B cells. Figure 1.b shows the
immune network representation. The respective idiotopes and paratopes of the B cells B1, B2 and
B3 are labeled as Id1, Id2, Id3, P1, P2 and P3. On an antigenic stimulus Ag, idiotope, Id1 of antibody
(Ab1) stimulates the B cell B2, with the antibody (Ab2) through the paratope P2. Viewed from the
point of Ab2, idiotope Id1 of the Ab1 works simultaneously as an antigen. As a result, Ab2
suppresses B1. On the other hand, Ab1 stimulates Ab3 since the Id1 of the Ab1 works as an antigen
for the Ab3. These mutual interaction chains between different species of antibodies form the
large-scale closed chain loop, which works as a self and non-self recognizer [65]. Besides the
information transfer between the B cells, the immune network is also necessary for the
collaborative activity of killing the antigen. The antigens have multiple epitopes and the B cells
are monospecific, with a single type of receptor [62]. Therefore, in order to kill the antigen the B
cells with different types of receptors group through the immune network to bind the antigenic
epitopes and eventually kill them. The immune network is the major reason for self-organization
abilities of the B cells.

There are two distinctive properties of the immune network that can improve the operation of a
sensor network. The first property is their way of communicating the non-self information in the
immune network with stimulus. The second property is their suppression behavior when the
stimulus is non-existent. Imitating these properties, algorithms and/or protocols can be generated
to improve the clustering efficiency. Even though communication happens locally in Jerne’s
immune network, the information passes through all the B cells in the system or in a larger
neighborhood since the network is close to fully connected network. It is important to note that
the Jerne’s network is active when there is stimulus between an antigen and a B cell in the
network. If there is no stimulus, the B cells undergo suppression, which is a B cell to B cell
communication or update. This communication (suppression) does not have a lot of overhead
and does not have to happen all the time. This will reduce the amount of energy the sensor nodes
will consume when there is no clustering or routing necessary in the network.

As discussed above, the immune network has two kinds of communication schemes, suppression
and stimulation, depending on whether there is an antigen in the system or not. In a sensor
network, these two schemes could be used. When there is a need for clustering, an antigen, the
sensor nodes can experience stimulation passing clustering information among them. If there is
no need for clustering, the nodes can operate regularly providing sensory information to the
clusterheads. These two schemes can even be set such that if the performance of the sensor
network goes down the network experiences stimulus. For example, if the parameter for
clustering is data aggregation and a node(s) discovers the network performance is low, the
node(s) can create antigenic response and let other nodes know about the performance
degradation.

4 Components of Investigation

4.1 Problem Formulation

The investigation endeavors to perform optimal clustering using biologically inspired


optimization approaches. What may not be clear yet is how can one apply a biologically inspired
approach for optimization of clustering a wireless sensor network. The answer to this question
helps to clarify the methodology to be used in this investigation. For any optimization problem,
there must exist a fitness or metric. In the context of data aggregation, hierarchical routing, and
energy conservation where clustering is crucial, metrics such as maximizing the lifespan of the
sensor network and maximizing data packets extracted from the sensor network over its lifetime
arise naturally. With metrics identified, a problem statement is formulated to bring the
optimization problem into the context of biologically inspired approaches. Two examples are
provided here to illustrate our methodology of applying biologically inspired approaches to
optimization problems for clustering.

Example 1: Clusterhead Identification in an Ad Hoc Sensor Network

The goal in this example is to conserve energy. A biologically inspired approach, Particle
Swarm Optimization, can be applied to conserve energy through a search for an optimal
configuration of clusterheads. A fitness or metric must be constructed.; the fitness chosen is
based on the observation that a clustering that equalizes the node counts in each cluster while
minimizing the distance between a node and its clusterhead will lead to optimal energy
conservation. The problem can now be formulated as follows.

The investigators of this proposal address the problem [67] of finding, in a group of N nodes,
where A nodes are available for the cluster-head role, the M nodes that minimize the average
distance between a node and its assigned clusterhead, while at the same time, equalizing the
number of nodes in each cluster. This is an NP hard problem [13] of a discrete and
combinatorial nature. The problem is solved by recursively splitting groups of sensors by an
optimal line that minimizes
Fitness = ( a1 − f1 A) 2 + ( a2 − f2 A)2 + (c1 − f1 N )2 + (c2 − f2 N )2 .

Two regions are formed in a division. These are referred to as region 1 and region 2. The
subscripts in the fitness function refer to these regions. In the fitness function, N and A are as
previously defined, ai is the number of available cluster-heads in region i , ci is the number of
nodes in region i and

Mi
fi = ,
M

where M i is the number of required clusters in region i . In the event that M is even, then
M 1 = M 2 and f1 = f 2 = 1/ 2 . This formulation allows us to handle the general case where M could
also be odd. The fitness represents the deviation from our prescribed optimum that we balance
the number of nodes and available cluster-heads in each region. Particle Swarm Optimization is
the optimization technique used to calculate the dividing line in each recursive split.

Example 2: Building an Optimal Communications Range, Fully Connected Graph, for


Clustering

Assume that a sensor node’s sensing range is fixed. Assume each node knows its position so that
communicating nodes can calculate distances. Also assume that the sensor node can adjust its
power so it can vary its reach or communication range. One framing of a problem statement
could be, “How should nodes vary their communication range optimally such that the number of
links is minimized while ensuring that the sensor network is not partitioned?”. By “not
partitioned”, it is implied that there is a d -hop route from any node in the network to any other
node in the network, where d is an integer. This is a clustering problem based on a single
optimization, specifically, minimizing the number of links. The solution will result in sets of
connected nodes or a graph. In the lowest-ID [14, 15, 16] approach, for example, links are
determined by communication ranges, but here in the biologically inspired solution, the
communication range is optimally adjusted for a specific purpose. Again, our purpose here in
this illustrative example is to minimize the number of links while remaining fully connected.
With this problem statement it is possible to propose a solution in the framework of biologically
inspired systems, namely Particle Swarm Optimization.

To apply Particle Swarm Optimization to this problem, it is necessary to introduce an innovative


extension of the existing algorithm. It is this extension of the Particle Swarm Optimization
algorithm that allows one to identify as particles in the swarm, nodes in the network. The
extension involves how the fitness of a particle, or node is expressed. Traditionally, a single
function takes the particle test solution and returns a fitness. But now there is no single function.
Each node has its own fitness function, which is distinct from the fitness functions of the other
nodes.
For the concrete example presented here, the fitness of node i could be written as

f ∝ ∑( d ij − ri − r j ) 2 ,
j ≠i

where the sum is over neighbor nodes of i that are within communication range, dij is the
distance separating the nodes, and rj is node j ’s currently implemented ‘representative’ range.
‘Representative’ range means to represent the distance you would need to travel from a node
before passing into another node’s ‘representative’ range. The fitness function involves another
term but the important revelation is that each node is trying to optimize its own value of ri , but is
taking cues about what its value of ri should be based on the fitnesses and rj s reported by its
neighbors. The exchanges of fitnesses and values could be executed during normal operation
and nodes could use converged ranges during re-clustering.

Fitness could represent not only communication range, but also energy level, connectivity, target
sensing data density (to indicate how relevant or active the sensor’s data is) and other operational
aspects of the sensor network. Existing clustering techniques are not capable of internalizing
system parameters in this way. Additionally, sensors could have many levels of fitness, evaluated
using different combinations of system parameters, thereby allowing for more than one type of
clustering to exist simultaneously.

4.2 Solution Development and Evaluation


An objective of the proposed research is to develop robust and scalable clustering algorithms. To
ensure that developed algorithms are robust and scalable, simulations are executed. To test
scalability, developed algorithms are tested with varying node number and densities. For each
value of the node number and density pair, robustness is tested by varying node mobility,
communication ranges, sensing ranges and data aggregation length scales. Random failures of
nodes are simulated to test robustness. Adaptability is simulated by randomly or
deterministically introducing additional nodes into the network. Data is collected for multiple
trials and statistical results tabulated. One performance metric may be attacked using one
biologically inspired approach, while another performance metric may be attacked using another.
If more than one biologically inspired approach is envisioned to be applicable, then both will be
investigated and their performances compared.

4.3 Resources

RIT is home to two computer clusters with another larger cluster currently under construction.
The investigators are actively using these resources to produce results from distributed
applications much faster than is possible on a single PC platform. Current applications include
biologically inspired, data-mining and optimization applications. With the addition of the newest
computer cluster, approximately 105 distributed computers will be available for message passing
applications.
5 Timeline and Task Distribution

During the clustering algorithm development, promising candidates will be implemented as


protocols. Implemented protocols will be tested using the Network Simulator (ns-2) [68].
Therefore protocol development and simulations will be ongoing throughout the life of the grant
period. There is a period of time of pure algorithm development during which initial good
protocol candidates are sought. Only after a candidate protocol is identified can the protocol
development phase begin. Likewise, until the first protocols are written, no simulations will be
possible. New protocols will be written as needed and simulated once completed. Results from
simulations will be the measure of the performance of the sensor network. Table 1 illustrates the
timeline and tasks to be completed during the grant period.

Table 1. Timeline
2003 2004 2005

task Sept Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug

algorithm development

student involvement

protocol development

ns-2 simulations

ns seminar development

quarterly reports
final report

Dissemination - Conference
Dissemination - Journal

broader impact activities


intellectual merit activities

As detailed in Section 4, the investigators have extensive experience [67] formulating


appropriate fitness functions for applying the biologically inspired Particle Swarm Optimization
method to problems directly applicable to the clustering of wireless sensor networks. The
investigators have also developed ant based and immune system based systems that operate to
minimize the time required to de-mine an area [69, 70, 71, 72].

Professor Sahin’s expertise is in biologically inspired learning, intelligent agents, decision-


making, and mobile robots [73, 74, 75]. The primary responsibility of the PI is to develop and
work on the biologically inspired models and/or algorithms to optimize the system parameters in
sensor networks. Professor Rao's expertise is in statistical signal processing and
communications. He has an extensive background in network models and data compression.
His expertise will be most useful in the proposed work as it relates to optimal data aggregation
and wireless traffic issues. Dr. Tillett, research staff member, will work on algorithmic
development and software implementation. He has 10 years of experience in software
development. His latest endeavor is developing a Message Passing Interface (MPI) program for
structural Bayesian network learning using Particle Swarm Optimization [76] on a 34-node
Solaris computer cluster. The proposed algorithms will be simulated in a Network Simulator
[68].

6 Impact of Proposed Research

6. 1 Impact in the field of sensor networks

The proposed research will result in innovations in approaches for optimal cluster formation of
sensors in wireless sensor networks. The innovations will come through novel techniques that are
based on biological problem solving paradigms. The techniques will offer robustness and
scalability with respect to changing operational conditions and impairments, traits that follow
naturally from the biologically inspired basis for the approaches. The resulting algorithms will
lead to efficient implementation of sensor network functions through quick convergence to
nearly optimal solutions. The distributed nature of the resulting algorithms will provide fault
tolerant capabilities. The work will enable more reliable, long-lived sensor networks for
application in industry, environmental monitoring, defense, law enforcement and other areas.
Once the approach has been well tested, dedicated hardware implementation on microsensor
platforms can follow enabling widespread deployment of such sensor networks.

6.2 Impact on PI institution

The grant will have a broader impact beyond the immediate contribution to science and
engineering. It will be a significant contributor to RIT’s current efforts to position itself as a
leader in Information Technology and Micro/nanosystems. The Kate Gleason College of
Engineering has recently instituted a Ph.D. program in Microsystems Engineering. The institute
has initiated a program called First-in-Class to enable it to become the leading institution of
choice for government and industry to carry out applied research. The New York State Science,
Technology and Academics Research Department has chosen RIT to be a “Star” Center with a
grant of $14 million to develop the infrastructure for the Information Technology Collaboratory.
The institute is committed to seeking external research funds to support research within the
collaboratory. The PIs are also part of the Laboratory for Autonomous Cooperative
Microsystems (LACoMS). The areas of sensor research as well as multi-agent network
communication, which come under the activities of LACoMS will benefit from the proposed
research on sensor networks. The graduate and the senior undergraduate curricula will benefit
from the results of the research material being incorporated in advanced courses and electives.

6. 3 Dissemination of research results

This will be accomplished through presentations at research symposia, journal articles and
technical reports. Data generated at the PI’s host institution will be made available to other
researchers through either the Internet or other appropriate arrangements.

Você também pode gostar