OLFAR: ADAPTIVE TOPOLOGY FOR SATELLITE SWARMS
University of Twente,
Faculty of Electrical Engineering, Mathematics and Computer Science,
Telecommunication Engineering Group,
P.O. Box 217, 7500 AE, Enschede, The Netherlands
R. T. Rajan1, S. Engelen2, A. Meijerink3, C. J. M. Verhoeven2 and M. J. Bentum4
Abstract—Low-frequency space observation is one of the de-
veloping directions in radio astronomy, as scientists try to reveal
more of the universe. The Orbiting Low Frequency Array for
Radio Astronomy project is aimed at developing a distributed
radio telescope sensitive to ultra-long waves, by placing an array
of antennas, far away from any terrestrial interference. It will
consist of over 50 small satellites grouped in a swarm capable of
collecting and processing astronomical data. The system will use
its resources to the limit and will need an efficient communication
topology in order to guarantee functionality. In this paper we
present a clustering scheme that reduces data distribution efforts
at the cost of decreasing imaging redundancy.
The Ultra Low Frequency band below 30 MHz is one of
the last unexplored frequency band in radio astronomy, dueto ionospheric distortion, man-made interference and evensolar flares. An unequivocal solution to the problem is toplace antennas far away from Earth to observe at these longwavelengths. The Orbiting Low Frequency Antennas for Radioastronomy (OLFAR)  project is aimed at designing anddeveloping a detailed system concept for an array of scalable
Satellite swarm orbiting around the Moon. Eccentric orbiting causes
autonomous nano satellites in space (not more than 100 km
the satellite cloud to vary its distribution with position.
apart), to be used as a scientific instrument for ultra-low-frequency observations. The OLFAR swarm could either orbitthe moon, whilst sampling during the Earth-radio eclipse
each other, both for radio astronomy observations and com-
phase, or orbit the sun, Earth-trailing or -leading, sampling
munications. The swarm will employ distributed correlation
almost continuously. To avoid single point of failure, OLFAR
, with inter-satellite data rates in excess of 6 Mbits/sec for
will be a distributed system with satellites cooperating with
an instantaneous observation bandwidth of larger than 1 MHz
and a data resolution higher than 1 bit. At the system level,
R. T. Rajan is with Delft University of Technology, Kluyverweg 1, 2629
HS Delft, The Netherlands and ASTRON, P.O. Box 2, 7900 AA Dwingeloo,
each nano satellite  is by definition power constrained. In
The Netherlands (e-mail: [email protected]
addition, the satellites are mobile and thus the topographical
2S. Engelen and C. J. M. Verhoeven are with the Delft University
distribution keeps changing with time. All these requirements
of Technology, Kluyverweg 1, 2629 HS Delft, The Netherlands (e-mail:[email protected]
; [email protected]
make data distribution within the swarm very complex.
3A. Meijerink is with the Telecommunication Engineering Group,
As already described, the OLFAR project will consist of a
Faculty of Electrical Engineering, Mathematics and Computer Science,University of Twente, 7500 AE Enschede, The Netherlands (e-mail:
large number (50 to 1,000)  of small satellites, that will
gather data individually and process it in a collective manner,
4M. J. Bentum is with the Telecommunication Engineering Group, Faculty
ensuring this way the functionality of a low-frequency radio
of Electrical Engineering, Mathematics and Computer Science, University ofTwente, 7500 AE Enschede, The Netherlands and with ASTRON, P.O. Box
telescope. On the whole, the system will act as a large wireless
2, 7900 AA Dwingeloo, The Netherlands (e-mail: [email protected]
sensor network, with a few peculiar requirements.
First of all, the network of satellites will be positioned
for low latency and high data rates makes it an inappropriate
in space, orbiting either the sun or the moon . Little is
solution for the OLFAR network, at least for the permanent
known about the environment from the communication point
regime. Although it cannot guarantee the correct functioning
of view, and, thus, the behavior of the satellites will have a
of the swarm as a radio telescope, gossiping remains a
high degree of uncertainty. Added to this, offline debugging
good solution for the initialization of the system—network
and component replacement is practically impossible. Hence,
discovery and synchronization.
the system designer has to consider the harshest conditions
For large-scale mobile ad-hoc networks (MANETs), a clus-
that may appear.
tering approach is essential to guarantee basic levels of system
Secondly, the distances between the nodes of the network
performance, such as throughput or delay. In most cases,
will be up to 100 km. This is imposed by the low-frequency
network division offers important benefits. A cluster structure
telescope functionality. In order to achieve sufficient spatial
eases spatial reuse of resources and can increase the system's
resolution, the telescope needs to have an aperture diameter
capacity. Same frequencies or codes can be deployed in
of 10–100 km . All the satellites will then be distributed
disjoint clusters. Added to this, routing is more facile to do in
over a very large cloud, making the network a low-density one.
a hierarchical approach. Cluster heads or gateways can form
This also adds complexity to the communication problem.
a back-bone for the system, simplifying the data distribution
Finally, space observation in low frequencies requires gath-
task. Nonetheless, one of the most important advantages of
ering and processing large amounts of data. What is more, the
clustering is that it makes the network more stable and smaller
satellites need to transmit their observed astronomical data to
from the members' point of view. Local changes in one domain
all the other members of the swarm. The necessary throughput
will not cause disturbances in the entire system .
of every node will be very high. The observed data Dobs
Therefore, clustering might be a good solution for many
rate will exceed 6 Mbit/second/satellite , while the required
WSNs, but, in order to be suitable for the OLFAR satellite
inter-satellite reception rate Din will be
swarm, it should also be power-efficient. Our goal is to reducethe complexity of the communication task and minimize theenergy consumed for transmitting and receiving information.
Din = Nsat − 1 D
In this way, the system's resources will be used more effi-
ciently to achieve the science objective, namely, exploring the
where Nsat is the number of satellites.
universe in the low-frequency band.
Apart from the up-mentioned requirements, the swarm will
also have constraints that most sensor networks encounter. The
power of each satellite will be limited and should be used
To prove the power efficiency of a clustering scheme, the
mainly to serve the purpose of the system—space observation.
following test scenario has been considered: a random sensor
Mobility is another characteristic that has to be taken into
network with N nodes was generated. A full-mesh topology
consideration. Nodes will drift away from the ideal trajectory
and a clustered topology were employed to the system and then
and will move relative to each other . The topographical
the total necessary power for communication was calculated.
distribution of the satellites will not be constant in time.
By necessary power we refer to the sum of power needed for
All these requirements make the communication task very
transmitting the data to all members of the network and the
difficult to fulfill.
power needed for receiving and processing the data.
The rest of this paper is organized as follows: in Sections II,
The following assumptions have been made:
III and IV, we motivate our clustering approach. In Section V,
1) The network terminals are uniformly distributed on a L
a new dynamic clustering algorithm is described. Simulation
by L square surface.
results are presented in Section VI, and concluding remarks
2) The communication environment is contention-free and
are drawn out in Section VII.
error-free, hence, there is no need for data retransmis-sion.
II. PROBLEM FORMULATION
3) In the full-mesh topology, each sensor transmits and
The most straightforward way to solve the communication
receives data from all the other sensors
layer problem would be to employ a full-mesh topology for the
4) Clusters are attained by dividing the initial surface into
network. This copes well with the fact that all the satellites are
M equal squares, M taking only the values 4 and 16.
identical and also provides robustness to the system. However,
This is illustrated in Fig. 2. As the clusters are formed
it comes with high demands of communication resources. It
considering a geometric criterion, the number of nodes
is certainly not feasible for a large network distributed over
may vary from cluster to cluster.
a vast area in space, because the necessary power will tend
5) For each cluster, the the node closest to the center of
to increase exponentially with the number of nodes and the
gravity rG of the cluster is elected as the cluster head,
distances between them.
One may suggest that a Gossip protocol would be a better
solution . It is successfully used in some ad-hoc networks
that implement distributed algorithms. Unfortunately, the need
Fig. 3. Power-efficiency of clustering: the ratio between the necessary powerfor a full-mesh network and the necessary power for a clustered network asa function of dmin/L.
Fig. 2. Clustering example. The area over which the sensor network is spread
was divided into four equal subareas, each one corresponding to a different
cluster. The master nodes are represented by pentagon shapes, whereas the
slaves are represented by circles.
Index k denotes a particular cluster, Nk is the number
of nodes in cluster k, and r
i corresponds to node i in
6) A node that is not a cluster head is considered to be a
slave node. All slave nodes send and receive data only
Average Node Power Consumption [dB]
to and from their corresponding cluster head. In other
words, each cluster employs a star topology.
7) Each cluster head sends and receives data to all slaves
in its cluster and to all the other cluster heads.
8) Each sensor consumes energy E for receiving and
processing one unit of data.
Power requirements: average node power consumption as a function
9) Each sensor has a minimum transmission power that
corresponds to a transmission distance dmin. If thedistance dij between the sending sensor i and receivingsensor j is smaller than d
For the scenario described above, we also calculated the
min, energy E is consumed
for transmitting one unit of data. Otherwise, the con-
average power consumed by every type of node: slave and
sumed energy for transmitting one unit of data increases
master for clustered network, and peer node for the full-mesh
quadratically, by the following rule:
case. In Fig. 4, the values are plotted on a logarithmic scale, as
a function of dmin/L. Clustering a network will make most of
the members consume less power for communication, at the
cost of overloading the relay nodes. There is a trade-off, yet,for low values of dmin/L the advantage is clearly in favor of
Considering all the assumptions, we calculate the total
The average power consumed by a slave node is very
necessary power for communication duty for both cases: full-
low compared to a peer node or a master node. Thus, in a
mesh topology and clustered network. In Fig. 3, the ratio
clustered network the slaves could save their resources for
between the two is plotted as a function of dmin/L. The
fulfilling other tasks than data distribution. Master nodes, on
value is higher than one proving that a clustered network is
the other hand, will tend to consume their energy mostly
more power efficient. For low values of dmin/L (low-density
for communication duty. For that reason, masters will not
networks), the required power for a full-mesh topology tends
contribute to the observation task.
to be one magnitude order higher than for the clustered case.
Despite the basic clustering scheme we used, the results
show that dividing a network into multiple clusters increases
nodes that receive only one paging signal are allocated with
communication channels, and, afterwards, other slaves areallocated channels in the decreasing order of their received
V. A DYNAMIC CLUSTERING SCHEME FOR THE
power level. By giving priority to slaves that receive only
one paging signal and employing power control, this scheme
Clustering is a step forward in achieving a functional
can achieve nearly optimum performance. However Ryu's
distributed radio telescope in space. By dividing the OLFAR
algorithm uses pre-defined masters and has no method for
swarm into small groups of satellites, and electing certain gate-
way satellites to route the astronomical data, energy resources
The global k-means algorithm  describes a deterministic
of the system will be used more efficiently. This hierarchical
global optimization method that minimizes the clustering error
approach, though an effective tool for the communication
(sum of the squared distances between nodes and cluster
layer, comes with one major drawback. Some of the satellites
centers). The algorithm is a fast iterative one that solves the
that will act as group leaders, will not be able to actively
clustering problem with M clusters by solving all intermediate
participate in the scientific task of the swarm. However, as long
problems with 1, 2, . . , M − 1 clusters. The basic idea is that
as this trade-off only impacts the redundancy of the system,
an optimal solution for the M clusters problem can be obtained
the improvements are uncontestable.
using a series of local searches with the k-means algorithm.
At each local search, M − 1 cluster centers have their initial
A. Existing algorithms
optimal position according to the M − 1 clusters problem. In
There are many proposed clustering schemes in the litera-
 the method is proposed for a pattern recognition scenario
ture suitable for dynamic wireless sensor networks. Most of
but can be extended for the networking case.
them have the same objective, and that is to optimize the
Both ACE  and ASH  are emergent cluster forma-
resource usage of the network. Nevertheless, for achieving
tion algorithms. In their approaches there is no use of central
this goal, different criteria are used. Dominant-Set-based 
control or visibility over all the network. Added to this, all the
protocols aim to reduce the routing cost by finding a Dominant
nodes communicate with only a limited number of immediate
Set in the network. Other clustering schemes  try to provide
neighbors. In ACE, the objective to create highly uniform
stable cluster architectures so that re-clustering situations are
clusters is achieved using two processes. The first controls the
avoided. In this way, the maintenance cost of the network is
spawning of new clusters by having nodes elected as leaders,
minimized. Mobility-aware clustering  tries to group nodes
and the second controls how clusters migrate dynamically to
by their dynamics as movement is usually the main cause
reduce overlap. For instance, a node can decide by itself to
for changes in the topology. Other used algorithms try to
become a cluster head. It will broadcast a "Recruit" message to
maximize the life time of mobile devices in a network or to
its neighbors that will become the followers of the new cluster.
balance the energy consumption amongst all the nodes. Added
Migration of a cluster is controlled by the leader. Each cluster
to these, combined metrics can also be employed to attain a
head will poll its followers to determine the best candidate for
desired clustering scheme.
the leader of the cluster. Once the best candidate is determined,
When dealing with a swarm of satellites, it is difficult to find
it will be promoted as the new leader. Thus, the position of the
a clustering algorithm that matches all the requirements: power
cluster will appear to migrate in the direction of the new cluster
efficiency, mobility and high data rates. However, there are a
head as some of the former followers of the old cluster-head
few algorithms that are partially fit for the OLFAR project,
will be no longer part of the cluster, while some new nodes
and, out of these, worth mentioning are Ryu's algorithm
near the new cluster head will become new followers of the
for energy-efficient clustering, the global k-means algorithm,
the Algorithm for Cluster Establishment (ACE) for uniform
Based also on emergent behavior, ASH  tackles the
cluster formation, the so-called ASH algorithm for highly
problem of node grouping in networks that exhibit high
dynamic networks, or MOBIlity-aware Clustering (MOBIC)
mobility. Node movement usually introduces a lot of problems
for networks that have group mobility behavior.
in a wireless network, such as routing failure, information loss,
In his paper  Ryu proposes two distributed heuristic
and others. The mechanism described in  handles mobility
clustering schemes that minimize the required transmission
in large-scale networks by employing a diffusion process that
energy in two-tiered MANETs. It assumes that the network
tends to equalize the pressure in the created groups. By using
has only two types of nodes: masters (cluster heads) and slaves
only local interactions, it creates domains whose centers of
(members). A slave node can be connected only to one master,
gravity move around slowly, providing a quasi-static overlay.
and links between slaves are not allowed. Master nodes are
The MOBIC clustering scheme  takes mobility into
selected in advance, and each master node establishes a cluster
consideration for cluster formation, and, especially, for leader
based on its connection to the slaves. The clustering process
election. It uses the premise that cluster head election is a
starts with a paging phase in which every master pages the
local process and should only be determined by the neighbors
slave nodes with the maximum allowed power. A slave that
and itself. The algorithm calculates the variance of a node's
receives this messages replies with an acknowledgement to
speed relative to its neighbors, and increases its probability of
the master corresponding to the strongest signal. First, the
becoming a master based on that. This is suitable for MANETs
in which groups of nodes tend to move with similar speeds anddirections. However, if the network is characterized by randommovement, MOBIC might not show good performance.
All the algorithms described above have good results in
terms of efficiency for different scenarios. Yet, applying themon a complex system such as the OLFAR satellite swarm, will,most probably, cause the system to overload and fail. For in-stance, Ryu's algorithm and the global k-means algorithm bothoptimize the energy consumption for a static network. Thepresence of mobile terminals will cause the cluster structure tobe re-built when events take place, and, thus, the performancewill be degraded. Mobility-aware schemes generate eitherlarge numbers of clusters or multi-hop topologies. For a high-speed network, it is best to avoid these scenarios as much aspossible, because of their need for data aggregation.
The algorithm that we propose combines the aforemen-
tioned advantages, fitting to the necessities of a satellite swarm.
It starts with electing the cluster heads and creating theircorresponding clusters, depending on the distribution of nodes.
Afterwards, it uses two procedures for node migration and forleader election that keep the system stable and minimize therisk of re-clustering. All the decision-making is done usingpower metrics, so that, in the permanent regime, the networktends to evolve to minimum power consumption.
The following assumptions were made when designing the
1) The clustering process starts after a gossiping round
takes place in the swarm. Hence, every node is awareof the spatial distribution of the swarm.
2) The initialization is done very fast compared to the
Initial cluster formation. (NACK is the number of acknowledgment
mobility of the nodes. For the initial cluster formation,
messages that an elected cluster head receives after paging, NK is the number
we assume that the network is static.
of formed clusters, and M is the desired number of clusters)
3) All the nodes are either masters or slaves.
4) Every cluster has a star topology. Master nodes are
connected in a full-mesh. A slave node can only be
will suffer a power penalty. This will be detailed in the
connected to one master.
5) A slave node can migrate to another cluster. A master
node cannot migrate, unless it changes its role to a slave.
C. Initial cluster formation
6) For the transmission power, we use the same assumption
In order to choose the nodes that are most suited for the
as in Section III.
leader role, the influence that a node has on all the other nodes
of the network is quantized into a density parameter which is
defined below. Using this value, the clusters are formed as
described in the flowchart in Fig. 5.
The steps of the cluster formation algorithm are as follows:
ij is the distance between sensor i and sensor j.
7) The communication between every pair of nodes is
1) Calculate node densities.
2) The node with the highest density is elected as the
8) Receiving and processing power is ignored. After sim-
ulating the scenario described in Section III, we con-
3) The cluster head pages all its neighbors in a radius
cluded that receiving and processing power is a neg-
dmin. Nodes that do not belong to any cluster respond
ligible quantity when comparing it to the transmission
with an acknowledgement signal and join the newly
formed cluster. If the number of acknowledgements is
9) Each master can have, by default, up to Nch slaves.
larger than the number of channels Nch, the cluster head
If the number of slaves is larger than Nch, the master
chooses only the closest Nch nodes as his followers.
4) The process is repeated until an apriori chosen number
channels a master can allocate. In most of the clustering
M of clusters is created or all the nodes are connected.
algorithms, it is not allowed to have more slaves than the
The density of a node i is a power-related parameter and is
number of channels. Yet, this cannot be the case for our
algorithm. We cannot afford to lose data from any satellite.
According to the Shannon-Hartley theorem:
where C is the communication channel's maximum capacity,
where node i is an unconnected node, and Eij is the energy
B is the bandwidth, and SN R is the signal-to-noise ratio.
required for transmitting one unit of data from node i to node
Taking into account that the communication between every
j, if node j is unconnected, or 0 otherwise.
two nodes is ideal, we can proceed as follows. Let us assume
D. Slave migration
that Cn is the channel capacity for accommodating n users,requiring SN Rn. And let us assume that in order to satisfy
In order to maintain the cluster structure and to keep the
the requirements of an additional m users a channel capacity
power consumption to a minimum, we define a slave migra-
Cn+m is needed, respectively, SN Rn+m. In both of the cases
tion procedure. The migration mechanism will be controlled
we assume that the available bandwidth B is the same, and
entirely by the cluster heads and will work as follows:
also the noise power is the same.
1) Each master node will calculate a cost parameter Cij for
Due to the fact that all users are supposed to have the same
all the slaves in the network, where Cij approximates
requirements, we also have:
the necessary power for node i to be part of cluster j.
2) Nodes that do not belong to any cluster are attributed a
default cost value.
3) If for a slave node i, member of cluster j, the following
Using (8) and (9) we can deduce a relation between SN Rn
condition is true:
and SN Rn+m, which describes, in fact, the amount of addi-tional power needed for a cluster to host n + m slaves when
it is designated for n slaves.
n+m = (1 + SN Rn)(1+m/n) − 1
As a result, in case a cluster will have more than N
then node i cancels its membership to cluster j, and joins
the necessary power will increase exponentially.
cluster k. θmig is a migration parameter and is equal to
Based on (10), using a Taylor series approximation, and
1 in the ideal case. In order to avoid node migrating
back and forth between two clusters, as a result of their
1 and m considerably smaller than n,
we define a new cost function:
random movement, the migration constant can be setlower than 1. This way, a node joins a different cluster
θcost(Nj − Nch)Cij
only if it finds a much better one in terms of cost.
for when the number of slaves in a cluster is larger than the
As already mentioned, the cost function of a slave will
number of allocated channels. In (11) θ
approximate the necessary power for a slave node to be part
cost is a parameter
depending on SN R
of a certain cluster. Initially, we defined the cost C
n, Nj is the number of nodes in cluster j
(including the potential node i), and C
the necessary power for a node i to communicate with another
ij is the cost calculated
according to (7).
hypothetical node positioned in the center of gravity of clusterj.
E. Cluster head re-election
Similar to the slave migration process, we define a mecha-
nism for changing the leader of a certain cluster. As the nodes
move, a master can turn up to be inefficient as a cluster head,so that another member of the cluster should take its role. The
dij is the distance between node i and the center of
master node should always be the node closest to the center of
gravity of cluster j. The coordinates for the centers of gravity
the group, moving in similar direction and at a similar speed
are calculated according to (2).
as its members.
Calculating the cost relative to the center of gravity instead
The re-election process is a local activity, in which all the
of relating to the position of the master node makes the system
members of a certain domain are involved. Let there be a
more stable. The movement or re-election of the cluster head
cluster j with master node i. Let G be the center of gravity
has less influence on the behavior of the slave nodes. The
of the cluster with its position vector r
G and its speed vector
migration process will be controlled by the entire cluster.
G, calculated according to (2), respectively:
The cost function described by (7) is valid only if the
number of slaves in a cluster is less than the number of
Power Ratio: transmission power needed for a full-mesh topology
divided by the transmission power necessary when applying the dynamicclustering scheme.
The movement pattern was chosen to be a random walk
pattern: each node was given a speed of L/200 per time stepand could change its movement direction at every simulationstep with a probability of 0.1. Finally, we set the parametersθcost, θmig, θCH and α with the values 2, 0.6, 0.6 and 0, thelast one corresponding to a random movement scenario.
Master Node Task.
The results of the numerical simulations are statistic results,
which are the mean values of 1,000 random configurationsof the network. As performance metrics, we use the total
consumed power for transmission and the average cluster size.
i are the speed vectors of nodes i in the cluster j.
With these definitions, the decision to attribute the role of
We ignored the receiving and processing consumed energy as
the leader to a new node is made based on the following
it is very low comparing it to the transmitting energy .
The total consumed power consists of the power necessary forall the existing links in the considered network: slave-master,
− α) kr
i − r
Gk + α i − v
Gk 6 θ
master-slave, and master-master links. We compare this value
j − r
j − v
with the total necessary power in case a full-mesh structurewould be used. In Fig. 7, we plotted the ratio between the two
where i denotes the slave node that candidates for a leader
values, i.e. the power needed for a full-mesh topology divided
position, and j is the actual master node. α is a tunable
by the power needed when we apply our clustering scheme.
parameter for strengthening one of the two terms, and θCH
After the initial cluster formation, the value of the ratio will be
is a parameter that has the same role as θmig defined in the
very high because some of the nodes will be unconnected. In
time, the coverage of the clusters grows to 100% and the value
In Fig. 6 a flowchart of the entire process of a master node
tends to stabilize. As anticipated in Section IV, the clustered
network is more power-efficient.
The average number of nodes in a cluster and the standard
deviation are plotted in Fig. 8. The proposed clustering scheme
We simulated the algorithm using a Netlogo environment
generates quasi-equal-sized domains in term of number of
and Matlab. The test scenario consisted of a network of 100
nodes. The mean value will remain constant once every node is
nodes uniformly distributed on square plane. The number
assigned to a cluster. However, the standard deviation indicates
of clusters was selected to be 6, and for every cluster 16
that cluster sizes will vary slightly, in order to keep the cost
channels were allocated without any penalty. For all the nodes,
functions to a minimum.
a minimum transmission range of L/4 was chosen. The default
The proposed algorithm comes with a few drawbacks, some
cost value for unconnected nodes was set to four. According
of which are shown in Fig. 9 and Fig. 10. Node mobility will
to this value, it is possible for unconnected terminals to join a
generate node transitions from one cluster to the other, and
cluster only when the distance to the center of gravity is less
role-transitions from master to slave and vice-versa. Changes
in the node distribution will not generate re-clustering, but
Cluster Size: average number of nodes per cluster and standard
Power Requirements: maximum transmission power for master
nodes (clustering approach) and peer nodes (full-mesh topology).
Power consumption for communication will mainly be con-
centrated in only a few satellites, leaving the other members of
the swarm with enough resources to fulfill the low-frequencyobservation task.
Although we designed the presented mechanisms for a
particular application, they can be used for most WSNs with
similar characteristics. By simply tuning the parameters, the
scheme may be a solution for many distributed systems that
 M. J. Bentum, C. J. M. Verhoeven, A. J. Boonstra, A. J. van der Veen
and E. K. A. Gill, "A novel astronomical application for formation flyingsmall satellites," 60th International Astronautical Congress
, Daejeon,Republic of Korea, 12–16 October 2009.
Cumulative number of transitions.
 R. T. Rajan, S. Engelen, M. J. Bentum and C. J. M. Verhoeven, "The
orbiting low frequency antenna array," IEEE Aerospace Conference
, BigSky, MT, 5–12 March 2011.
 S. Engelen, E. K. A. Gill and C. J. M. Verhoeven, "Systems Engineering
will have a major impact in the roles of the swarm members.
Challenges for Satellite Swarms," IEEE Aerospace Conference
, Big Sky,
Initially, the cluster structure will not be very stable, as a
MT, 5–12 March 2011.
 S. Jester and H. Falcke, "Science with a lunar low-frequency array: From
large number of nodes will tend to join the newly formed
the dark ages of the universe to nearby exoplanets," New Astronomy
clusters. This results in a high number of master-slave and
slave-master transitions (dashed line). After a transient phase,
 N. Saks, A. J. Boonstra, R. J. Rajan, M. J. Bentum, and F. Belien,
cluster head re-election process will occur less often, yet the
"DARIS, a fleet of passive formation flying small satellites for lowfrequency radio astronomy," Small Satellite and Services Symposium
number of slave migrations will increase. Added to this, as
Funchal, Portugal, 31 May–4 June 2010.
anticipated, the power requirements for master nodes will
 D. Kempe, J. Kleinberg and A. Demers, "Spatial gossip and resource
prevent them from having other tasks than communication.The
location protocols," Journal of the ACM
, vol. 51, no. 6, November 2004.
 J. Yu and P. Chong, "A survey of clustering schemes for mobile ad hoc
main drawback of the algorithm is common for most of the
networks," Communications Surveys Tutorials, IEEE
, vol. 7, no. 1, pp.
clustering schemes. The power requirements for the master
node will prevent cluster heads from carrying any other tasks.
 P. Basu, N. Khan, and T. D. C. Little, "A Mobility Based Metric for
Clustering in Mobile Ad Hoc Networks," Proc. IEEE ICDCSW01
413–18, April 2001.
 J. H. Ryu, S. Song and D. H. Cho, "New Clustering Schemes for Energy
Conservation in Two-Tiered Mobile Ad-Hoc Networks," Proc. IEEE
In this paper, we presented an adaptive topology that
, vo1. 3, pp. 86266, June 2001.
matches the demands of the OLFAR satellite swarm. By
 A. Likas, N. Vlassis and J. J. Verbeek, "The global k-means clustering
employing power cost functions and a node migration mecha-
algorithm," Pattern Recognition
, vol. 36, pp. 451-461, 2003.
 H. Chan and A. Perrig, "ACE: An emergent algorithm for highly
nism, the proposed algorithm creates a two-layer hierarchical
uniform cluster formation," European Conference on Wireless Sensor
structure that improves the data distribution in the system.
,Berlin, Germany, 19–21 January, 2004.
 A. Pruteanu, S. Dulman and K. Langendoen, "ASH: Tackling node mo-
bility in large-scale networks," IEEE International Conference on Self-Adaptive and Self-Organizing Systems
, Budapest, Hungary, September27–October 1 2010.
 G. J. Pottie and W. J. Kaiser, "Wireless Integrated Network Sensors,"
Communications of the ACM
, vol. 43, no. 5, pp 51-58, May 2000.
MONOGRAPHIE Pr ARIMIDEX® comprimés à 1 mg Inhibiteur de l'aromatase non stéroïdien AstraZeneca Canada Inc. Date de révision : 1004 Middlegate Road 27 avril 2011 Mississauga, Ontario L4Y 1M4 www.astrazeneca.ca Numéro de contrôle : 143918 ARIMIDEX® est une marque de commerce du groupe AstraZeneca.
2016–2017 TBOR Basic International Student Injury and Sickness Plan Endorsed by Middle Tennessee State University Who is eligible to enroll? How do I Enroll? International students or other persons with a current To enroll visit www.pghintlstudent.com, and follow passport who: 1) are engaged in educational activities; 2) are temporarily located outside his/her home country as a