Академический Документы
Профессиональный Документы
Культура Документы
Ad Hoc Networks
journal homepage: www.elsevier.com/locate/adhoc
a r t i c l e i n f o a b s t r a c t
Article history: Context information can be used to streamline routing decisions in opportunistic networks.
Available online 31 May 2011 We propose a novel social context-based routing scheme that considers both the spatial
and the temporal dimensions of the activity of mobile nodes to predict the mobility pat-
Keywords: terns of nodes based on the BackPropagation Neural Networks model.
Opportunistic networks Ó 2011 Elsevier B.V. All rights reserved.
Routing algorithms
Social context-based
Context information
Encounter probability
1570-8705/$ - see front matter Ó 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.adhoc.2011.05.007
1558 H.A. Nguyen, S. Giordano / Ad Hoc Networks 10 (2012) 1557–1569
In this paper, we propose a social context-based routing works, especially in the routing of messages. The Probabi-
scheme, called Context Information Prediction for Routing listic Routing scheme – PRoPHET [8] calculates the delivery
in OppNets (CiPRO), in which a BackPropagation Neural predictability from a node to a particular destination node
Network (BNN) model is used to predict the context of based on the observed contact history. The delivery pre-
nodes, so that the source device knows when and where dictability is the probability of a node encountering a cer-
to start the routing process to maximize the transmission tain destination. It increases when the node frequently
delay and minimize the network overhead. meets the destination and decreases (according to an aging
The structure of this document is as follows: Section 2 function) on the contrary. The context information in Bub-
presents the recent research on routing in OppNets. The ble Rap [9] is the social communities that nodes belong to.
main contribution comes in Section 3 where we present Communities are automatically defined and labeled based
our proposal (CiPRO). In Section 4, we evaluate the perfor- on the patterns of contacts between nodes. When a node
mance of our solution and with simulation results, we wants to send a message to an other node, it looks for
show that it outperforms other solutions. Conclusions nodes belonging to the same community as the destina-
and future work are presented in Section 5. tion. If such nodes are not found, it forwards the message
to increasingly sociable nodes, which have better chance
of getting in touch with the community of the destination.
2. Recent research on routing in OppNets
CAR-Adaptive Routing for Intermittently Connected Mobile
Ad hoc Networks [2,10] uses Kalman filters to combine and
A routing scheme for OppNets has to ensure some reli-
evaluate the multiple dimensions of the context in order to
ability (ideally with full reliability) even when the network
take routing decisions. The context is made of measure-
connectivity is intermittent or when an end-to-end path
ments that nodes perform periodically, which can be re-
might be nonexistent. Moreover, in an opportunistic net-
lated to connectivity, but not necessarily. We can also
work, nodes can arbitrarily have new contacts (neighbors)
cite here some other approaches that fall in this category:
without prior information, meaning that traditional rout-
Meeting and Visits (MV) [11], MaxProp [12], and Moby-
ing approaches cannot be applied.
Space [13], etc.
In this section, we provide a brief description of the pri-
mary routing approaches in OppNets available in the liter-
2.3. Social context-based routing protocols
ature, from naïve to intelligent approaches. Specifically, we
emphasize the role of context information as well as social
The main difference between social context-based pro-
aspects in routing. Based on the context information
tocols and mobility-based protocols is that the context
exploited, we classify routing in OppNets into three clas-
information in the mobility-based protocols is about the
ses: context-oblivious, mobility-based and social context-
mobility patterns of the nodes, the information about the
aware protocols [3,4].
devices themselves, or the encounter history between
nodes. Whereas social context-aware protocols do not only
2.1. Context-oblivious routing techniques exploit such kinds of context information as in the mobil-
ity-based protocols, but also take into account the social
The routing techniques in the context-oblivious routing aspect of the nodes as an important parameter to route
group use flooding-based techniques to route data, from‘‘- the messages. This is motivated from the fact that in most
blind’’ flooding to controlled flooding solutions. In the of the cases, the mobility of nodes is decided by the carriers
flooding-based techniques, a sender sends out a request (e.g. humans, animals, and vehicles). Hence, the social rela-
packet on all outgoing links. Each node receives a request tionship of the carriers plays an important role in how they
packet and forwards the packet again on all outgoing links encounter each other. The advantage of this approach is
except the one corresponding to the incoming link on that it is more general than mobility-based approaches. In-
which the packet arrived. Each request packet may reach deed, these routing protocols can be used with any set of
the destination along a different route at a different time. context information, and thus they can be easily custom-
The advantage of flooding-based routing is its simplicity ized to best suit the particular environment. To the best
in finding a route, in particular, it has the best performance of our knowledge, there are two protocols that fall in this
in terms of end-to-end transmission delay. However, flood- category: Propicman [1] and HiBOp [14].
ing causes a huge number of request packets in control The main idea of HiBOp forwarding is looking for nodes
channels, which can result in network congestion. Further that show an increasing match with the known context
examples of context-oblivious routing techniques are Epi- attributes of the destination. A high match means a high
demic routing [5] and network coding based routing similarity between node’s and destination’s contexts and,
schemes [6]. therefore, a high probability for the node to bring the mes-
sage in the destination’s community (possibly, to the des-
2.2. Mobility-based routing techniques tination). One of the main drawbacks of HiBOp is about
its high overhead to the network. Nodes periodically
Routing protocols in the mobility-based category ex- broadcast and exchange their context information. Conse-
ploit the mobility information of nodes to make forwarding quently, the overhead will increase depending on the den-
decisions. Node mobility impacts the effectiveness of rout- sity of the nodes.
ing in opportunistic networks, and Grossglauser and David Our previous contribution Propicman [1] is one of the
[7] proved that it increases the performance of ad hoc net- first Social-context based routing techniques and allows
H.A. Nguyen, S. Giordano / Ad Hoc Networks 10 (2012) 1557–1569 1559
even for routing with zero knowledge. Nodes do not have Propicman [1]), or sends the control packets (the summary
to exchange their information for routing purpose. Unlike vectors) to all the neighbors (as in ProPHET [8]) to compute
Epidemic, the same as the routing techniques mentioned the encounter probability between the neighbors and the
in Section 2.2, in Propicman, the message content is only destination, even though at this time there are no nodes
forwarded to the neighbor that has the highest probability around (for example, during the night). In this situation,
to deliver the message to the destination. Therefore, it if the sender still routes the message, it will have very
eliminates the drawbacks of flooding-based techniques, low probability (even zero) to successfully deliver the mes-
such as high resource utilization, and network bandwidth sage to the destination.
eating. Moreover, in Propicman, the authors also proposed We also observe that most people have some frequent
a technique to hide the information about the destination contacts and other occasional contacts in their social rela-
that is used to select the best neighbor(s) to forward the tionships. The occasional contact persons are the people
message. The privacy of the destination, so far, is not con- that they rarely contact with, whereas they meet the fre-
sidered sufficiently in the recent approaches. quent contact persons very regularly, can be once per
Beside the strong points of Propicman, there is still day, every hour, or once per week, etc. depending on the
room for improvement. In fact, these routing algorithms application. For their frequent contact people, they may
only partially use the spatial and temporal aspects when have some knowledge of these individuals’ periodic behav-
routing messages. For example, nodes still route data even ior. Therefore, if the destination is a frequent contact per-
there is no or very few neighbors that have a high encoun- son of the sender, the sender may know in which period
ter probability with the destination (for instance, during it probably meets the destination or meets the nodes that
the night). As a result, the senders still execute the routing have a high probability to encounter the destination in
process even if the chance of reaching the destination is the future, thus it can have a good routing strategy to en-
very low. sure the delivery. In other words, the sender has some idea
In the proposed routing scheme – CiPRO, we consider when and where to send out messages in order to maxi-
also the temporal and spatial aspects of context informa- mize the delivery probability to the destination. Therefore
tion and uses BackPropagation Neural Network model to we distinguish, in the social relationships of a node, two
predict the encounter probability between the sender classes of persons:
and the destination. Thus, CiPRO provides the sender with
knowledge of when and where it can have a high probabil- Frequent contact persons: the persons that the node
ity to successfully delivery the message to the destination. often meets.
Occasional contact persons: the other persons.
3. Context information prediction for social-based So when the sender has a message to send to a destina-
routing technique in OppNets – CiPRO tion, if the destination is a frequent contact of the sender,
the sender uses CiPRO to route the message, otherwise, it
3.1. Motivation uses Propicman because the sender does not have any
prior knowledge of the periodic behavior of the destination
We observe that the activity patterns of devices (nodes) (see Fig. 1).
are likely to be repeated. For example, in a normal day, In the next sections, we will present the assumptions,
people wake up in the morning, have breakfast, go to some main concepts and definitions that will be used in
work/school/. . . . During a certain period of the day, they our routing techniques.
meet some sets of persons. Obviously, the number of con-
tacts are different by periods. So that if nodes can foresee 3.2. Assumptions
in which period of the day they will be very likely to meet
a certain destination, or meet nodes that have a high prob- In order to evaluate the meeting probability between
ability to encounter the destination in the future, they will two nodes, an important assumption that will be used
initiate the routing process. Otherwise, in low probability through the whole work: if two nodes have the similar node
periods, such as night time, they do not perform the rout- profiles, they will have the very high probability to meet each
ing process.
The published context-based routing solutions in Opp-
Nets do not significantly consider the spatial and temporal
aspects, they either exploit the frequency of contacts with
the destination (Prophet), or exploit the frequency of visit-
ing some pre-known places during a day such as in [15].
Additionally, they do not use social information of
node’carriers to exploit their possible contacts during a
day. The sender – the node holding the message and trying
to choose the best neighbor(s) to forward the message to
the destination - always routes the message in the same
way, regardless of the spatial and temporal context. For
example, when the sender has a message to send to a des-
tination, it sends the control message to its neighbors (as in Fig. 1. Routing strategy.
1560 H.A. Nguyen, S. Giordano / Ad Hoc Networks 10 (2012) 1557–1569
to route messages. So when we write node profiles, it refers Evidence name (e) Value (v) Hashed values
to the profile of the device carriers. ei vi H(ei, vi)
This assumption seems to be intuitive as, for example, if
two nodes work at the same workplace, speak the same
language, having the same habits (practice same kind of
Table 2
sport), participating in the same projects, etc. will have An example of a node profile.
very high probability to meet each other during the day.
However, proving it is out of the scope of this paper. Evidence name (e) Value (v) Hashed values
Rather, it is matter of ongoing experimental work (see Name Alessandro H(Name, Alessandro)
Section 5). Nationality Italy H(Nationality, Italy)
Profession IT researcher H(Profession, IT researcher)
Workplace Cambridge H(Workplace, Cambridge)
3.3. Concepts and definitions in social context-based routing Residence place Luton H(Residence place, Luton)
resents the cardinality of N i , the encounter probability be- probability to send the message, but also some neighbors
tween S and D within period i is computed as follows: that have high delivery probabilities in the top-down order
P to forward the message content, and of course, these val-
A2Ni p½A; D
Pi ¼ ; ð2Þ ues must be higher than the encounter probability of S.
jNi j
with Pi is the encounter probability within period i, p[A, D] is 3.5. Routing if the destination is a frequent contact
the encounter probability between node A and the destina-
tion, computed as in Eq. (1). As mentioned above, the activities of device carriers are
different by periods of time, for example activities of peo-
ple during 24 hours of a day. In order to analyze the con-
3.4. Routing if the destination is an occasional contact
text information of a node, we divide its activities into
cycles and periods. Dividing the activities of people into cy-
If the destination is an occasional contact of the sender,
the sender does not have any history contact with the des- cles and periods is also used in our first work on the predic-
tion of context information in OppNets (SpatioTempo
tination. In this case, the sender uses Propicman to route
messages. Details are presented below. [19]). The notion of cycles and periods appeared by our
observation that during a normal day, the device carriers
Nodes in the network have node profiles that are com-
posed by evidence/value pairs and the hashed values of (typically people) probably repeat their activities day by
day. We can see that at each period, nodes probably meet
each pair. Each evidence is assigned a weight to classify
which evidence is important to successfully deliver the a similar set of the other nodes. So, in order to send data to
a destination, the source node uses the history encounter
message to the destination. Suppose that the sender S –
the node holding the message and trying to choose the best with the destination at each period to predict the
encounter probability with the destination in the next per-
neighbor(s) to forward the message M to the destination D
– knows some information about the profile of D. S builds iod. As a result, it can decide the frequency of broadcasting
the control packet and the message content in the next
the control message HM which is the concatenation of all
the hashed values of the evidence/value pairs in order period, thus it can minimize the message overhead and
also the resource utilization. Therefore, if the destination
(see Fig. 2). These evidences can have values if S knows
or can be empty in the opposite case. is a frequent contact person, the senders can predict when
it will have a high probability to meet the destination next
HM ¼ Cat ri¼1 ðHðei ; v i Þ; wi ; MS ; qS Þ; time, so prediction is used to find the best interval for
sending the message content in the coming periods.
where MS (MAC address of S) and qS (the sequence number
Obviously, there are always exceptional cases, such as
of M) are used to avoid duplicates, wi are the weights cor-
weekends or holidays. In these cases, they probably meet
responding to each evidence and are put in the same order
other sets of people.
as the corresponding evidences, r is the number of evi-
In the next section, we explain the prediction model
dence/value pairs that S may know in the profile of D. By
that we apply and the reason why we use this for our
putting the weights into the control message HM , the sen-
problem.
der can classify which evidence is important for the inter-
mediate nodes in order to maximize its probability of
3.5.1. Predicting next contacts: BackPropagation Neural
meeting the destination.
Network model
At first, S sends the control message HM to every neigh-
In order to foresee the contact situation of the sender in
bor, without including the message content. Each neighbor
the next period, we must rely on the contact history be-
(for instance, A), when receiving HM , compares each of its
tween the sender and other nodes. As the sender already
own hashed values of evidence/value pairs to the hashed
has some knowledge about the contact history with its fre-
values in HM , then computes its encounter probability
quent contact nodes, this knowledge can be used to predict
with D as in Section 3.3.3, Eq. (1). Then A broadcasts HM
the contact situation of the sender in the next period. We
to its neighbors (second hop of S through A). The second
choose the Neural-based BNN technique as the prediction
hop nodes continue to compute the encounter probability
technique because of the following reasons [20,21]:
and send the result back to A. A chooses the node with
the highest encounter probability from two hop nodes
Adaptive to the current environment, BNNs are flexible.
then sends back to S. Thus, for each control message sent
The parameters (synaptic weights) are automatically
out, the sender S receives the delivery probabilities of its
adjusted after each training process to have a better
two hop neighbors. If the highest encounter probability re-
output accuracy.
ceived from two hop neighbors is higher than the encoun-
It is straightforward to implement.
ter probability of S, S will forward the message content M
It is application-independent.
to this neighbor. To ensure the message delivery, S can se-
lect not only the neighbor that has the highest encounter
BackPropagation technique: A neural network has to
be configured such that the application of a set of inputs
produces the desired output. Various methods to set the
strengths of the connections exist. One way is to set the
synaptic weights explicitly, using a priori knowledge. An-
Fig. 2. Control message structure. other way is to train the system by feeding it with samples,
1562 H.A. Nguyen, S. Giordano / Ad Hoc Networks 10 (2012) 1557–1569
encounter probability within the next period, it will decide weekend days are considered exceptions (behavior is dif-
on the number of broadcasts of the message header in the ferent from the other days of the week).
next period. At the beginning, each node randomly has 8 frequent
Given u the maximum broadcast capacity of S in the destinations. For the training process, both the vertical
period i, the number of broadcasts of the message header and horizontal synaptic weights are 0.5, as well as the
in the period i will be a rounded value as follows: learning rate. During the simulation, the messages were
sent out from arbitrary senders, 75% of them were sent
y ¼ b/Ppred c; 0 < / 6 u: ð9Þ
to the frequent destinations of the sources. The simulation
In Eq. (9), the parameter / is flexible, depends on the con- results are the average results from 16 simulation runs
straints of nodes to the system. In other words, if the mes- with the 95% confidence level.
sage is important, and the network overhead is not a At period i, if node S has message M to send to destina-
problem, in this case, the communication delay is an tion D, there are two cases:
important parameter, the parameter / is set to maximum
(equal to u). Thus, after the prediction process, if the pre- 1. D is one of the frequent contacts of S, S will predict the
dicted encounter probability within the next period is high, encounter probability within the next period (i + 1).
the sender must increase the number of broadcasts of the 2. D is not a frequent contact of S. In this case, S uses Pro-
message header to maximize the capacity of sending the picman to route the message.
message to the intermediate nodes that have the high
encounter probability with the destination in the future. Note that when there is a message to send to a destina-
Assuming U is the number of units of time in a period, tion, this message can go through some intermediate
for the next period, the sender will broadcast the message nodes. Then each node uses its relationship with the desti-
header in every bUyc ¼ b/PU c units. By predicting the con- nation to decide which routing technique will be em-
pred
text information in the next period, the sender can adap- ployed: CiPRO in the case that the destination is one of
tively have a suitable routing strategy by deciding the its frequent destinations, or Propicman in the opposite
frequency of broadcasting the message headers in the next case.
period.
4.2. Evaluation of the contact prediction
4. Simulation results and performance evaluation
Fig. 6 shows the percent error of the prediction. As
4.1. Mobility model and simulation description shown in the figure, the precision increases by cycles and
periods. During the first cycles (cycles 1–4), the node does
The simulation scenario is a rectangle of 1870 not have enough context information to predict the
1520 m, divided in a 2 6 grid. We simulate 50 nodes, encounter probability within the next period, and as a re-
with the transmission range of 30 m. The mobility model sult, the predicted value and the desired value are very dif-
used in the simulation is the Community-based Mobility ferent. Thanks to the learning process with synaptic weight
Model [22]. Its main idea is to model nodal mobility based adjustments and the impact of the learning rate after each
on human social relationships and thus identify communi- period, in time, period by period, cycle by cycle, these two
ties. The communities are then placed into the celled net- values come closer. This tendency is represented by the
work area, and each community occupies a cell. A node decrement of the percent error shown in Fig. 6. Of course,
moves randomly in its cell. When it is outside of its cell, there are still some points that have quite big relative per-
its moving decisions depend on the social attractiveness centage value (cycle 9, for instance), this happens for the
(importance in terms of the social relationships). We exceptional cases. From cycle 4 onward, the percent error
developed the simulator in Java based on the traces gener- index goes down, from around 20% down to around 7%. It
ated from CMM model.1
The simulation is executed based on some assumptions:
1
The nodes have unlimited power, no constraint about the "percentError.data"
is the result of the learning process, through periods to The sender selects up to 20% of the neighbors that have
periods, with the experience achieved after each learning high delivery probabilities in the descending order. And
process, the synaptic weights are adjusted to have the pre- of course, these values must be higher than the encoun-
dicted outputs as close as possible to the desired values. ter probability of the sender.
Thus, after some cycles (cycle 5), the predicted result be-
comes more stable. The percent error goes down around Fig. 7 shows that flooding-based routing scheme (Epi-
7%. This is an acceptable level if we consider the typical demic) is very costly in terms of overhead in comparison
characteristics of OppNets. to the three routing algorithms. The figure shows that,
compared to Epidemic, CIPRO gains two orders of magni-
4.3. Impact factors and performance evaluation tude in terms of overhead at the cost of double delay; Pro-
picman and ProPHET have about 20 times less overhead at
In the performance evaluation, we compared CiPRO the price of double and three times delay respectively. This
with three representative routing protocols (context-obliv- is reasonable because in CiPRO, senders only send control
ious, mobility-based and social context-based routing pro- messages and messages when they have a high probability
tocols as mentioned in Section 2): Epidemic, ProPHET and to encounter the destination or when they encounter
Propicman respectively. We ran all these routing solutions neighbors that have a high probability to meet the destina-
in the same testbed with the same scenario to evaluate tions in the future. Whereas, in Propicman, control mes-
their performance. We have focused on comparing their sages are sent regularly regardless of the existence of
performance with regard to the following metrics. First of neighbors, or if these neighbors are useful or not to bring
all, we are interested in the network overhead: how many the messages to the destination. Nodes in ProPHET ex-
additional bytes are sent for successfully delivering a mes- change information whenever they meet each other, there-
sage to a destination. Applications using OppNets should fore, this technique is quite costly in terms of network
be relatively delay tolerant, but it is still of interest to con- overhead in comparison with Propicman and CiPRO. The
sider the message delivery delay to know how long it takes significant advantage of Propicman in comparison with
a message to be delivered. Epidemic and ProPHET techniques is that even if in Propic-
Moreover, we vary the number of nodes and their selec- man the control message is sent regularly, the message
tion for routing to see the impact on the performance of content is only sent when the senders encounter the desti-
the node density and of the number of selected candidates nation or encounter neighbors with a high encounter
at each neighbor selection. As in the simulation scenario, probability.
the speed of nodes and the sending rate are randomly gen- The network overhead parameter depends on the den-
erated, we do not consider their impact on the perfor- sity of the nodes (thus the density of neighbors). Especially
mance. During the evaluation, sometime we use the term in CiPro case, this parameter strongly depends on the den-
‘‘routing process’’. A routing process is started when a node sity of the nodes that have high encounter probabilities
broadcast the control packet and is ended by finding the with the destination. The point of cycle 10 in Fig. 7b can
next hop forwarder. be the case that even the node density is high, but there
are very few neighbors that have high encounter probabil-
4.3.1. Impact of the number of selected candidates ities. Therefore, senders in CiPro do not broadcast the con-
In ProPHET, Propicman and CiPRO, a sender only sends trol packet and message content. As a result, we can see
the message content to the neighbor(s) that have the high- that the overhead of CiPro is very low.
est encounter probability with the destination and this Thanks to the ability to predict the suitable time and
probability must be higher than the encounter probability places for message delivering, CiPRO has significantly less
of the sender. The number of nodes selected as forwarders network overhead than Propicman, without costing much
at each routing process impacts on the routing perfor- more than it in terms of delay, and being much better than
mance. If more nodes are involved in the routing process, ProPHET. This happens because the usage of spatial and
more resources are used but the data delay can be lowered. temporal context information allows a more precise mes-
In order to evaluate the performance, we vary the number sage forwarding (when and where some node with higher
of neighbors selected as forwarders at each routing pro- probability is expected) than ProPHET and Propicman that
cess, and compare the performance between our solution only use context information of current neighbors.
CiPRO with the other ProPHET, Propicman and Epidemic. We expect that the delay parameter of Propicman and
Epidemic is chosen as the basic routing algorithm in the ProPHET is decreased while increasing the number of
comparison to see how much CiPRO, ProPHET and Propic- neighbors selected at each routing process. On the other
man improve/cost with respect to a blind approach. So we hand, we also expect that the network overhead is in-
normalize the results of the three routing protocols creased when more network nodes are involved in the
with the results of Epidemic, and we show that in Figs. 7 routing process. In order to evaluate it, we increase the
and 8. In the implementation, we run the simulation in number of nodes involved in the following way: at each
two scenarios: candidate selection step, if the sender receives the encoun-
ter probability values from its neighbors that are higher
The sender only sends the message content to the than its encounter probability, it sends the message con-
neighbor that has the highest encounter probability tent up to (if there are) 20% of the nodes that have the high
with the destination and this value must be higher than delivery probabilities. Fig. 8, when compared to Fig. 7
the encounter probability of the sender. shows that when there are more neighbors selected to
1566 H.A. Nguyen, S. Giordano / Ad Hoc Networks 10 (2012) 1557–1569
1500
4.0
1000
3.0 Propicman
500
ProPhet
Epidemic
CiPro
0 2.0
0 5 10 15 20 25 0 5 10 15 20 25
Messages Messages
Fig. 7. (a and b) show the delay and the overhead of CiPRO, Propicman, Prophet compared to Epidemic when only the neighbor with the highest encounter
probability is selected. Figure (b) shows the semi-logarithmic plot of overhead comparison. In (b), the overhead caused by CiPRO is around two orders of
magnitude less than Epidemic at the cost of double delay (a), whereas (a) shows that CiPRO has around 2 times. Propicman and ProPHET have about 20
times less overhead at the price of double and three times delay respectively. CiPRO has the best performance in term of network overhead (confidence
level=95%).
Propicman
ProPhet Propicman
1000 Epidemic ProPhet
CiPro Epidemic
CiPro
800 5.0
seconds
600
400 4.0
200
0 3.0
0 5 10 15 20 25 0 5 10 15 20 25
Messages Messages
Fig. 8. (a and b) show the delay and the overhead of CiPRO, Propicman, Prophet compared to Epidemic when only the neighbor with the highest encounter
probability is selected. Fig. 7b shows the semi-logarithmic plot of overhead comparison. In (b), the overhead caused by CiPRO is around two orders of
magnitude less than Epidemic at the cost of double delay (a), whereas (a) shows that CiPRO has around 2 times. Propicman and ProPHET have about 20
times less overhead at the price of double and three times delay respectively. CiPRO has the best performance in term of network overhead (confidence
level=95%).
send the message content at each routing process, delay is by the sender to forward the message content. Thus, there
reduced at the price of higher overhead. However, as the is no significant change in terms of overhead as each time,
gain in terms of delay is not high, we can conclude that only the neighbor with the highest encounter probability is
there is no evident advantage in selecting more than one selected (in other words, the maximum number of neigh-
neighbor (assuming no losses). Clearly, if we release the bors that will receive the message is fixed by the protocol).
assumption that no message is lost, the problem changes, We can only appreciate some improvement in terms of de-
and this is matter of ongoing work. lay as we increase the probability of finding a better for-
warder as we increase the population. Similar reasoning
4.3.2. Impact of the node density applies to CIPRO, where the overhead also remains quite
We investigate the scalability of the routing algorithms stable, and we observe some improvements in terms of de-
by increasing the number of nodes from 50 up to 250 lay (Fig. 9).
nodes in steps of 50. On the contrary, if we increase the node density, there
In Propicman, senders regularly broadcast the control are more nodes involved, messages in the flooding-based
message to find the best forwarder(s). Only the neighbor routing scheme will quickly invade the entire network.
that has the highest encounter probability will be selected As a result, the message will reach the destination more
H.A. Nguyen, S. Giordano / Ad Hoc Networks 10 (2012) 1557–1569 1567
Delay with different node densities Overhead with different numbers of nodes
8
Epidemic Epidemic
2 0.1
1 0.05
0 0
50 100 150 200 250 50 100 150 200 250
Number of Nodes Number of Nodes
Fig. 9. (a and b) show the delay and the overhead of CiPRO, Propicman, Prophet and Epidemic in the case of varying the number of nodes. The delay and
overhead of Epidemic are normalized to 1 and the other algorithms are normalized by the Epidemic curves. With more nodes involved, the delay caused by
Epidemic is dramatically decreased, whereas the delays caused by ProPHET, Propicman and CiPRO are slightly decreased. The network overhead caused by
Epidemic is much increased, the overhead caused by ProPHET is slightly increased. The overhead caused by CiPRO and ProPHET shows to be less dependent
on the number of nodes (confidence level=88%).
quickly (Fig. 9a), thus the message delay is decreased. The other interesting work is to investigate the possibil-
However, because of the flooding of messages on the net- ity of leveraging long term repetitive patterns in the
work, this routing technique is very costly in terms of net- behavior of nodes to minimize the overhead of the learning
work overhead in comparison with the other techniques as process and therefore the computational cost.
shown in Fig. 9b. Increasing the number of nodes is also
important in ProPHET, if we increase the number of nodes, Acknowledgments
the number of contacts that a node may have will increase.
As a consequence, there are more control packets ex- We would like to thank Prof. Luca Maria Gambardella
changed, thus more overhead uploaded to the network. (IDSIA-Istituto Dalle Molle di Studi sull’Intelligenza Artifi-
To conclude, if we increase the node density, the mes- ciale) for his valuable suggestions in prediction with Neu-
sage delay is much improved in the case of Epidemic, ral Networks. We also would like to thank the reviewers
slightly improved in the case of ProPHET, Propicman and for their insightful comments on the paper.
CiPRO as there may be more nodes with high delivery
probabilities. However, Epidemic is very costly in terms Appendix A. Training data with BNN model: synaptic
of network overhead, this parameter is slightly increased weights, learning rate adjustments
in the case of ProPHET as there are more control packets
exchanged between nodes. In the case of Propicman and For any given training sequence, this value is a function
CiPRO, increasing the node density does not significantly of the synaptic weights of the network. So, to reduce the
impact the performance of the algorithms. error, it is necessary to control the gradient of the error
function with respect to each synaptic weight. One may
5. Conclusion and future work then move each synaptic weight slightly in the opposite
direction to the gradient. The gradient is defined as the
We applied the BackPropagation Neural Network model vector of partial derivatives of the multivariate function
to predict the behavior of device carriers. The evaluation with respect to each variable. Because the error function
analysis and the simulation results indicate that for the is a function of the network output, we first need to
context-based routing algorithms in OppNets, CiPRO – calculate the partial derivative of the network output
the BNN-based solution – outperforms other routing solu- with respect to each associated connection weight.
tions thanks to its ability to maximize the delivery ratio Since all other variables but the one of interest are held
and to minimize the network overhead. constant when we calculate the partial derivative, only
This work is the first contribution to date that employs one linear term is left in the calculation of the partial deriv-
the BNN model to predict the mobility pattern in OppNets. ative of the output, and leaving the coefficient - which is
In our future work, we will continue to evaluate the perfor- just the corresponding input. So we compute the synaptic
mance of CiPRO by evaluating the impact of the number of weight adjustment of the vertical and horizontal inputs
frequent contacts to the CiPRO performance. Especially, in as follows:
order to better evaluate our social context-based routing The synaptic weight of the vertical input – wv:
techniques, currently we are doing a real experiment to dE dE dPpred dm
capture the real social contact with the participants wear- ¼ : ðA:1Þ
dwv dPpred dm dwv
ing sensor nodes (Telosb). The collected traces will be the
first case of contact traces with context information. From Eqs. 4, 5, 7, we have:
1568 H.A. Nguyen, S. Giordano / Ad Hoc Networks 10 (2012) 1557–1569
European Science Foundation (ESF) activities. Since 1999 she serves as She will be general chair of WoWMoM 2009, program co-chair of IEEE
Technical Editor of IEEE Communications Magazine, and is currently PERCOM 2009, was program co-chair of IEEE VTC-Fall 2008, IEEE MASS
editor of the series co-editor of the new Series on Ad Hoc And Sensor 2007, workshop chair of IEEE WOWMoM 2007, tutorial chair of MobiHoc
Networks of the IEEE Communication Magazine. She is also editor of Ad 2006, general chair of IEEE WONS 2005, organizer of IEEE Persens 2005–
hoc networks journal by Elsevier, Ad Hoc & Sensor Wireless Networks 2009 workshop, IEEE AOC2005–2009 workshop and ACM Mobihoc SANET
journal, Ocpscience, Journal of Ubiquitous Computing and Intelligence workshop 2007–2008 and is/was on the executive committee and TCP of
(JUCI) and Journal of Autonomic and Trusted Computing (JoATC) both by several international conferences, and serves as reviewer on transactions
American Scientific Publishers (ASP), and Mediterranean Journal of and journals, as well as for several important conferences. Prof. Silvia
Computer and Networks, SoftMotor. She was already co-editor of several Giordano is a member of IEEE Computer Society, ACM and IFIP WG 6.8.
special issues of IEEE Communications Magazine and Baltzer MONET and Her current research interests include wireless and mobile ad hoc net-
Cluster Computing on mobile ad hoc networking and QoS networking. works, QoS and traffic control.