Information Flow Based Routing Algorithms For Wireless Sensor Networks
Information Flow Based Routing Algorithms For Wireless Sensor Networks
Information Flow Based Routing Algorithms For Wireless Sensor Networks
1 Introduction
Advances in microwave devices and digital electronics have enabled the devel-
opment of low-cost, low-power sensors that can be wirelessly networked to-
gether to give rise to a sensor network. With potential applications in a wide
variety of settings, like military, health and security, sensor networks have wit-
nessed significant attention from the networking research community in the last
few years. Applications of sensor networks range from early forest fire detec-
tion and sophisticated earthquake monitoring in dense urban areas, to battlefield
surveillance [1] and highly specialized medical diagnostic tasks where tiny sen-
sors may even be ingested or administered into the human body [2]. Given this
wide range of applications, wireless sensor networks are poised to become an
integral part of our lives.
Though related, wireless sensor networks are very different from mobile
ad hoc networks. In wireless sensor networks, the sensor nodes are usually de-
ployed very densely, and each sensor is more prone to failure. Each sensor node,
2 Zhang, Mahalingam and Memon
important and no one node must be penalized more than any of the others. This
metric ensures that all the nodes in the network remain up and running together
for as long as possible. In MREP [12], which we shall review later, the authors
try to achieve this by calculating routing paths that postpone the time of death
(running out of battery power) of the first node.
However, the fact that the routing paths are chosen in such a way that all
nodes die at the same time does not automatically imply that the energy uti-
lization is optimal. As an extreme case, we can easily see that if appropriately
selected subset of nodes are forced to be part of the route for every transmis-
sion, it may cause the nodes to “die simultaneously.” But obviously this does
not amount to efficient utilization of resources! Therefore, neither a large ratio
of transmitted bits to the total energy utilized nor the “uniformity” of expend-
ing every node’s resource, by themselves, indicate optimality of the network.
This clearly calls for an alternate metric for the evaluation of the performance
of sensor networks.
sensor 3
sensor 0
sensor 1 sensor 2
Proxy
Fig. 1. Example network configuration that illustrates difference in information flow to the proxy
over lifetime of network for two different routing strategies.
4 Zhang, Mahalingam and Memon
To illustrate our point further, consider the example of Figure 1, where four
sensors are deployed on a 10 10 grid at points 0 414 50, 1 80 50,
2 999 50 and 3 254 700. The four sensors measure and relay the
information to a “proxy” in the center of the grid. Two obvious ways to achieve
transfer of information from the sensors to the proxy are:
If each node is equipped with 500 units of power at the beginning, and each
node transmits one unit of information every unit time, and each unit of trans-
mission through a distance d requires d2 units of power, the direct path algo-
rithm would result in the death of node 2 at 20 units of time, node 3 at 31 units,
node 1 at 55 units and node 0 at 500 units. The shortest path algorithm on the
other hand would cause node 1 to die at time 28, node 2 at 43, node 3 at 55
and node 0 at time 444. It is not immediately obvious as to which scenario is
preferable. The direct path results in the first two nodes dying faster. On the
other hand, the scenario is not bad even after the two nodes die - nodes 0 and 1
on either side of the proxy are still alive. It is therefore still possible to gather
some meaningful information from the remaining nodes. Even though the short-
est path algorithm prolongs the life of the first two nodes, the death of the first
two nodes results in a situation where the proxy is not able to get any measure-
ments from one side (as both 1 and 2 are dead). It is intuitive that after the death
of nodes 2, 3 (direct path) the network retains the capability to provide more
meaningful information, while the situation is different after the death of nodes
1 and 2. This certainly indicates the need for a suitable metric to evaluate the
performance of sensor networks.
One of the main motivations of this paper is therefore the choice of a new
metric for evaluation of the performance of sensor networks. We propose the use
of total information delivered by a network, under the constraint of expendable
(battery) power available to each node. It is very important to realize here that
total information delivered is not the same as the total number of bits that are
transmitted. This is due to two reasons. The obvious reason is that the number
of bits transmitted also depend on the number of hops. A bit sent by a sensor
node to the proxy may travel through multiple intermediate nodes and hence get
re-transmitted multiple times. The second, and from our point of view the more
important, reason is that not all bits are equal. Some bits carry more “informa-
tion” than others. This fact can be understood if one recalls that any deployment
of wireless sensors is expected to provide the user with intelligence and a better
understanding of the environment in which they have been deployed. The sen-
Information Flow Based Routing 5
sors for instance may be measuring some field which may be thermal, acoustic,
visual, or infrared. The measurements would then be relayed to a central proxy,
which would then relay the information to the end user. What the user cares
about is the total information the network delivers about the underlying ran-
dom field that is being measured (sensed) (under a given constraint of battery
power in each node). Hence information is a natural evaluation metric for the
performance of a wireless sensor network. The question arises as to how can we
suitably quantify this metric?
Now, it is clear that the total information received by the proxy depends on
the information originating from each node, and the life of each node. Also, the
information originating from a node at any point in time also depends on the
number of nodes that are “alive” at that point in time, and the spatial location
of the nodes. For instance if two nodes are very close to each other (and the
field that is being measured is continuous), then there exists a high correlation
between the data originating from the two nodes. The total information from
both nodes in this case may be very close to the information originating from
just one node. As a more concrete example, in the example of four nodes we
investigated earlier, the information from node 1 becomes more important after
the death of node 2.
2 Problem Setting
proxy
sensor
Fig. 2. Example sensor network of randomly scattered sensors in a square and proxy in the center
of the field.
Usually, it is assumed that the radio dissipates Eelec =50nJ/bit to run the trans-
mitter or receiver circuity and Eamp =100pJ/bits/m2 for the transmitter amplifier
to achieve an acceptable signal-to-noise ratio [14]. Compared to Eamp , the Eelec
is usually very small, and can be ignored. In this paper therefore, we only con-
sider the transmission power, which is proportional to dα , where α is between 2
and 4 [15]. We choose α = 2 in the paper.
In sensor network literature, several different definitions have been proposed
for the “lifetime” of a network. Ref. [14] defines “lifetime” as the time till the
first sensor “dies”. Ref. [13] considers lifetime as the time till all sensors die.
The definition of “lifetime” should obviously depend on the nature of the appli-
cation. For instance, for applications like surveillance, it may be crucial that all
sensors be alive. So even the death of one sensor may end the “useful” life of the
network. In practice, as nodes keep dying, at some point, the total information
that is delivered from the network to the proxy keeps reducing. At some point
when the total information delivered by the network is below some threshold,
it may, for instance, not be worthwhile for the proxy to keep operating. So a
network with only few sensors alive may be useless. As a balance between the
two extreme definitions of lifetime, we define lifetime as the time until only half
of the sensors are alive.
Now that we have defined the framework under consideration, let us exam-
ine the total information originating from a wireless sensor network as the one
shown in Figure 2. We consider the measurement xi of the i’th node as a Gaus-
sian random variable. We shall assume further, without any loss of generality,
that the measurements constitute samples of a unit variance Gaussian distribu-
tion. The covariance matrix K of the n measurements x0 x1 xn 1 is then
In this section we present two routing algorithms, MIR and CMIR, that focus
on maximizing the information flow metric we have defined above. Before we
explain our proposed routing algorithms, we first quickly review the MREP al-
gorithm [12] as it serves as the basis of our constructions.
the energy consumption is balanced among the nodes in proportion to their en-
ergy reserves, instead of routing to minimize the absolute consumed power (as
in [16, 17]). The authors in [12] also showed that (“necessary optimality condi-
tion”) if the minimum lifetime over all nodes is maximized then the minimum
lifetime of each path flow from the origin to the destination with positive flow
has the same value as the other paths. For a path p Pi , where Pi is the set of
all paths from sensor i to the proxy as the destination, the path length Lp is de-
fined as a vector whose elements are the reciprocal of the residual energy for
each link in the path, after the route has been used for a unit flow. The routing
path is therefore calculated for each unit flow. The vector of such link costs is
represented by
ci j E i ei j λ 1 (9)
where E i j is the residual energy at node i, λ is a unit flow, and ei j the transmis-
sion cost (per bit) from node i to node j. A lexicographical ordering was used
in comparison of the two length vectors to enable comparison of the largest el-
ements first and so on. The shortest path from each node i to the destination is
obtained using a slightly modified version of the distributed Bellman-Ford al-
gorithm using the modified link costs. The flow then occurs via the the shortest
path so obtained.
The central idea behind the MREP algorithm is to augment the flow on
paths whose minimum residual energy after the flow augmentation will be the
largest. In the simulations performed in [12], 20 nodes are randomly distributed
in a square of size 5 by 5 among which 5 sensors and 1 proxy are randomly
chosen and the transmission range of each node is limited by 2.5. The energy
expenditure per bit transmission from node i to j is given by
di j 4
ei j max 001
25
(10)
where di j 25 is the distance between nodes i and j. The cases where there is
no path available between the sensor and the proxy are discarded.
di2 j
(11)
exp j I
Though not explicitly shown in the equation above, j I is also a function of time
- as nodes keep dying, j I changes. The distance between i and j is di j . In this
way, we direct the data to the sensor according to not only the power consumed
but also based on (the lack of) information in the originating node of the link.
The algorithm proceeds as follows, in n steps. In each step, we use Dijk-
stra’s algorithm to find the shortest path. After this step the weight of the links
that have been used are increased by a certain factor (this would indirectly cor-
respond to weighing the path based on expended battery power, as in MREP).
The next shortest path is then calculated based on the updated weights, and the
weights of the calculated path are increased again. This process is repeated until
every sensor’s shortest path to the proxy is determined. In our simulations, the
factor used was 18. Since the algorithm entails at most n iterations of Dijkstra’s
algorithm, it results in a worst case complexity of O n2 log n, where n is the
number of sensors.
by each sensor is almost the same, and therefore the remaining battery life of
the nodes is also roughly the same. However, as the algorithm progresses, the
imbalances in the remaining battery life become significant. As MIR does not
amplify the problem of remaining battery life as much as MREP, MIR performs
better when the remaining battery life of the nodes is more even. However, as
the the remaining battery power becomes highly variant, MREP does better. The
CMIR algorithm recognizes this trend, and therefore utilizes MIR initially, and
MREP at the later stages.
For the simulations, random allocation of the sensors were generated to evaluate
the performance of the three algorithms - MIR, CMIR and MREP. The metric
chosen was the total information flow from the network till the death of half the
nodes in the network.
The size of the square field considered was 10 by 10 units. The field itself
was assumed to be a first order Gauss-Markov field with unit variance, and
correlation coefficient ρ 08. The proxy (with unlimited resources) is assumed
to be located at the center of the field. The number of sensors (with random 0
x 10 and 0 y 10 coordinates chosen for the simulations were 10,50,100
and 150. Each sensor node was assigned an initial energy of 1000 units.
The performances of the MIR, CMIR, and MREP are compared in Table 1,
in terms of percentage improvement over MREP. The comparison shows signif-
icant improvement of MIR and CMIR over the MREP algorithm, especially for
large n, the number of sensors deployed.
The choice of ρ 08 and the weighting factors for information exp jI and
the factor (18) for adjusting the weights of computed paths, although reasonable
and intuitive, are primarily arbitrary. Simulations for performance results for
other choices of the parameters and the relationship between the parameters are
in progress and will be presented in the final version of the paper. However, the
results given here are representative of results obtained with different parameters
in an average sense.
In this paper we proposed a new strategy for routing in wireless sensor net-
works. The basis of our work is the realization that the primary metric for the
performance of a network is the information delivered by the network. The basis
translates to the observation that not all nodes are equal, even in a fairly uniform
field, due to the (random) spatial locations of the sensors. All nodes do not con-
tribute the same amount of information. Therefore the routing algorithm tries
to extend the life of nodes that contribute more information, at the expense of
nodes that do not.
We proposed two novel routing algorithms, Maximum Information Routing
(MIR) algorithm and the Conditional Maximum Information Routing (CMIR)
algorithm. Simulations show that the two novel algorithms perform significantly
better than the Maximum Residual Energy Path (MREP) algorithm proposed
in [12].
It is still not clear as to what the “optimal” scheme for maximizing the in-
formation during the “lifetime” of a wireless sensor network. Since the infor-
mation depends on the spatial distribution of the sensors, there may not be a
single scheme that is optimal for all sizes / distributions of a wireless sensor
network. To obtain more insight, our current work is focused on optimal routing
schemes for a fixed allocation of wireless sensors, and spatial allocations that
are inherently suitable for such applications.
References