power-aware ad hoc cognitive packet networks

Ad Hoc Networks 2 (2004) 205–216

www.elsevier.com/locate/adhoc

Power-aware ad hoc cognitive packet networks

Erol Gelenbe a,*, Ricardo Lent b

a Dennis Gabor Chair, Department of Electrical and Electronic Engineering, Imperial College, London SW7 2BT, UKb School of Computer Science, University of Central Florida, Orlando, FL 32816, USA

Available online 28 April 2004

Abstract

This paper proposes a new energy efficient algorithm to find and maintain routes in mobile ad hoc networks. The

proposal borrows the notion of learning from a previous research on cognitive packet networks (CPN) to create a

robust routing protocol. Our idea uses smart packets that exploit the use of unicasts and broadcasts to search for

routes. Because unicasts impose lower overall overhead, their use is preferred. Smart packets learn how to make good

unicast routing decisions by employing a combined goal function which considers both the energy stored in the nodes

and path delay. The end result is a dynamic discovery of paths that offer an equilibrium between low-delay routes and

an efficient use of network resources that extends the working lifetime of the network.

� 2004 Elsevier B.V. All rights reserved.

Keywords: Cognitive packet networks; Mobile ad hoc networks

1. Introduction

Mobile ad hoc networks are flexible and dy-

namic systems that can operate without the aid of

a fixed communication infrastructure. The topol-

ogy of an ad hoc network is expected to change

over time and it depends on the location of the

nodes and the resources available. Node locations

determine the establishment of links between

nodes whenever the distance and other externalfactors, such as the presence of obstacles and

interference, do not prevent nodes from commu-

nicating. In addition to acting as end systems,

nodes in ad hoc networks also act as transit nodes

* Corresponding author. Tel.: +44-2075-946342; fax: +44-

2075-946274.

E-mail addresses: [email protected] (E. Gelenbe),

[email protected] (R. Lent).

1570-8705/$ - see front matter � 2004 Elsevier B.V. All rights reserv

doi:10.1016/j.adhoc.2004.03.009

for other communications. Their participation in

the process of searching for paths (routing) andforwarding of packets depends on the availability

of internal resources. These resources are typically

scarce because of the mobile nature of the nodes.

One vital component is the stored energy in bat-

teries of mobile nodes, which is limited because of

weight and size restrictions. Furthermore, ad-

vances in battery technology lag behind advances

in computing and storage technologies [1]. Ad hocnetworks inflict extra energy consumption at

nodes, as they need to receive, process, and

transmit packets to assist others communications.

As a result, as nodes consume their resources, they

may quickly become unreliable and contribute to

create an error-prone system. Ad hoc networks are

therefore characterized by unpredictable topolo-

gies that require a highly dynamic routing algo-rithm to cope both with unreliability and mobility

ed.

mail to: [email protected]

206 E. Gelenbe, R. Lent / Ad Hoc Networks 2 (2004) 205–216

of nodes, while attempting to provide good quality

routes.

Flooding is a robust but resource consuming

mechanism that has been extensively employed in

the past as a searching mechanism for routes. A

request packet is addressed to the destination nodeand emitted by the source via broadcast every time a

path is to be discovered. The request is then repli-

cated at most one time by each node in the network.

This mechanism ensures the arrival of the request

packet at the destination provided that at least one

path exists. The process is expensive and pervasive

and, as a consequence, highly energy consuming.

The Cognitive Packet Network (CPN) [2–5] is afast adaptive routing algorithm that exploits learn-

ing to discover and refine routes. Routes are created

and maintained byCognitive or Smart Packets (SP),

which are sent out by source nodes when new des-

tinations are desired. SPs move in the network col-

lecting information and making decisions that take

into consideration what previous packets have

learned. Decisions can be tailored to reflect a desiredquality of service (QoS) on the path, for example,

minimize end-to-end delay or power consumption.

In this paper, we explore an extension to CPN

that enables its operation under mobile ad hoc

conditions. Ah hoc CPN (AHCPN) restricts the

use of flooding as a last-resort alternative and re-

places broadcasts with unicasts as much as possi-

ble. By employing the CPN algorithm, unicastrouting decisions can be adapted to optimize the

use of resources in mobile ad hoc networks that

reduce node unavailability due to power outages

and improve quality of service (QoS).

The rest of the paper is organized as follows.

Section 2 presents an overview of existing propo-

sals for mobile ad hoc networks. Section 3 intro-

duces the AHCPN routing algorithm, which is laterdetailed in Section 4. Section 5 reports on simula-

tion experiments that compares our proposal with

a flood-based algorithm. Finally, Section 6 con-

cludes the paper summarizing our main results.

2. Routing in ad hoc networks

The area of ad hoc routing has been very active

in recent years, and there appear to be two well-

defined trends in the design of routing protocols for

mobile ad hoc networks: table driven protocols

(proactive protocols) and source-initiated on-

demand driven protocols (reactive protocols) [6].

Proactive protocols operate with global informa-

tion about the network so as to maintain routes forevery possible source-destination pair. To acquire

this information, proactive algorithms require

nodes to periodically exchange routing tables.

Reactive protocols, on the other hand, create

routes only as needed. In general, and under low to

moderate network traffic, a periodic exchange of

packets will result in higher energy consumption

than the search for routes on-demand. A few ofthese proposals consider energy constrains in the

establishment of routes. We start this section by

reviewing non-energy aware protocols, and later on

we discuss energy aware protocols.

The Destination Sequenced Distance Vector

(DSDV) [7] is a proactive protocol that implements

the Bellman–Ford distance vector algorithm with

suitable adaptations for wireless environments.DSDV introduces a sequence number to each route

entry in order to solve the count-to-infinity prob-

lem. This sequence number allows nodes to discern

stale from new information so that only the most

recent routes are used. Global State Routing

(GSR) [8] is a proactive protocol that relies on a

variation of a link state algorithm to calculate

shortest paths. GSR requires nodes to periodicallyexchange their entire topology table with their

neighbors and then calculate paths with a variant

of the Dijkstra’s algorithm. Each entry in a topol-

ogy table is tagged with a sequence number to

allow nodes to recognize updates. The Wireless

Routing Protocol (WRP) [9] is another proactive

approach that relies on a link state algorithm to

calculate shortest paths. As defined by the link statealgorithm, nodes exchange link state information

with their neighbors either periodically or when a

change occurs. WRP tracks the length of the pre-

decessor-to-destination in the shortest path calcu-

lation to eliminate the count-to-infinity problem.

A major inconvenience of DSDV, WRP, and

GSR is the large number of control packets that

are needed to circulate routing information to allnodes in the network. Hierarchical-based algo-

rithms partition the network into areas in order to

E. Gelenbe, R. Lent / Ad Hoc Networks 2 (2004) 205–216 207

restrict the flow of routing control packets. Clus-

terhead Gateway Switch Routing (CGSR) [10] is a

proactive algorithm that partitions the network

into clusters, which contain a dynamically elected

clusterhead. Within these clusters, nodes that are

able to reach more than one clusterhead are elec-ted as gateways. CGSR uses distance vector rout-

ing to maintain paths that describe the series of

clusterheads and gateways required to reach a

destination, which can significantly help in reduc-

ing the size of the routing tables. Fisheye State

Routing (FSR) [11] is an improved version of

GSR. FSR periodically exchanges entire link state

tables with neighbors on different frequencies,which depend on hop distances.

Most on-demand protocols use a variation of

flooding to discover and maintain routes. This

usually consists in sending a route request packet

to find a route from a source to a destination.

Flooding is restricted by introducing a sequence

number for request packets and limiting each node

to process one request at most once, thusrestricting the control overhead to OðnÞ for an nnode network.

Dynamic Source Routing (DSR) [12] is an on-

demand protocol that sends request packets to

desired destinations and stores the discovered

paths in source nodes. Request packets record the

path that they need to follow to reach their desti-

nations and the route information is sent to thesource inside route reply packets. DSR takes

advantage of existing routing information in inter-

mediate nodes to accelerate route discovery and

possibly reduce routing overhead. The Ad Hoc

On-Demand Distance Vector (AODV) [13,14]

protocol uses query packets to discover new des-

tinations. A query packet behaves similarly to a

route request packet in DSR. However, AODVstores paths in a distributed fashion. AODV uses

backward learning, which consists in nodes acquir-

ing the path to the source when they receive a

query packet.

A number of proposals discover routes on-

demand but require a periodic transmission of

control packets to check for neighbor connectivity

and other metrics. These protocols are as expensivein terms of energy consumption as proactive pro-

tocols [15] but are able to provide a different type of

route quality. Associativity-Based Routing ABR

[16] is similar to DSR but employs a new metric

in the establishment of routes. Nodes periodically

transmit hello messages to their neighbors and the

count of these messages at a node provides an

indicative measurement of the link strength withthe origin of the hello messages. Signal Stability

Routing (SSR) [17] is another routing protocol that

requires periodic beacons to determine the link

stability of a node. This information is employed to

determine routing tables. The Temporary-Ordered

Routing Algorithm (TORA) is a distributed rout-

ing protocol that is based on local maxima for

propagation. Although TORA is an on-demandprotocol, it is designed to work over the Internet

MANET Encapsulation Protocol (IMEP) [18],

which requires periodic transmission of packets.

The Zone Routing Protocol (ZRP) [19,20] is a

hybrid approach that combines both proactive and

reactive algorithms. ZRP creates zones and uses

proactivity within these zones. Reactivity is used

when a node needs to reach nodes outside its zone.Although, the use of Global Positioning Sys-

tems (GPS) may introduce additional energy

requirements and possibly extra packet control

overhead to disseminate positional information,

geographic position information can be exploited

to reduce the total number of control packets for

routing purposes. Location Aided Routing (LAR)

[21] utilizes the position of the nodes and a vari-ation of DSR to restrict the area of search for

routes. Similarly, geoTORA [22] is a variation of

TORA using geographic position.

Research in power-efficient techniques em-

braces all seven layers of the OSI model. Power

conserving methods have been investigated in the

design of hardware [23], operating systems, and

applications, ranging from the design of energy-efficient schedulers for processors [24,25] to power-

aware applications for database transactions [26].

A recent commercial product that exemplifies the

need for power-efficient components is the new

Intel Centrino technology.

2.1. Power-aware protocols

Power efficient, routing protocols include the

work of Singh et al. [27] who investigated the use


of power-aware metrics in the calculation of

shortest paths. These metrics describe the power

required for transmitting and receiving a packet

on a link, so as to minimize the end-to-end power

requirements for routing. This proposal did not

take into account the remaining energy in thenodes and it can result in a severe drain of energy

in the batteries of the nodes on the least-cost route.

Other proposals overcame this problem by

using battery lifetime information. Toh [28] pro-

posed a new metric, which calculates the summa-

tion of the inverses of the remaining battery

capacities of the nodes on the path. In addition,

Toh proposed the Min–Max algorithm to main-tain a fair use of resources by avoiding the use of

nodes with the least remaining-battery capacity in

the network. Li et al. [29] propose an algorithm

(denoted by max–min zPmin) that computes the

paths with minimal energy consumption while

maximizing the minimal residual power of the

network.

Power-aware Source Routing (PSR) [30] issimilar to DSR, but the destination calculates the

link cost based on the remaining battery capacity

and transmission power of the nodes. The draw-

back of this approach is that the destination needs

to wait some time after the arrival of the first route

request, so as to receive more than one possible

route, and then selects the one with the minimum

cost.Finally, an alternative approach is used by the

Adaptive Fidelity Algorithm (AFA) [31], which

operates on top of on-demand ad hoc routing

protocols, such as AODV and DSR. AFA saves

battery power by turning off certain transceivers

whenever the applications allow a reduction in the

quality of the connections. The algorithm trades

quality for battery lifetime, network bandwidth, ora number of active sensors.

3. Ad hoc cognitive packet network (AHCPN)

Cognitive packet networks use three types of

packets to accomplish all routing and forwarding

functions: smart packets (SPs), dumb packets(DPs), and acknowledgments (ACK). SPs are

responsible for discovery of routes and for main-

tenance. As more information is learned from the

network, smart packets can refine routes by taking

better routing decisions. DPs employ source

routing with the paths discovered by SPs to move

payload from source to destination. Finally,

ACKs are employed to relay the information ac-quired by SPs of DPs. There are three elements

in the structure of any CPN packet. A header

transports source and destination addresses and

other useful information for the processing of the

packet, such as the packet’s length. A cognitive

map is an area that the packet uses to store net-

work status information. DPs use the data area to

transport payload.

3.1. Route discovery

In ad hoc networks, each node needs to become

aware of its neighbors dynamically and continu-

ously, as opposed to the CPN algorithm where

that information is known a priori and is not ex-

pected to change much over time. A node canlearn the existence of a neighbor simply by listen-

ing to its transmissions, and assuming that the

neighbor is alive as long as it keeps transmitting

packets. In AHCPN, we refrain from forcing each

node in the network to send out periodically a

bacon packet to indicate its existence because of

the energy consumption that this implies [15,

Chap. 9]. Nodes simply listen any transmission todetect neighbors.

To discover routes, SPs may use either broad-

cast or unicast transmissions to move in the net-

work. Broadcasts are obviously more expensive

than unicasts because they may be processed by

all nodes in the communication vicinity, and may

get reproduced as many times as the number

of neighbors. To use a unicast transmission, a SPneeds to select a target neighbor and therefore a

routing decision needs to take place. The process

of taking the routing decision is defined by the

CPN algorithm.

When nodes have no knowledge about their

surroundings (e.g. when they have just entered the

network), no unicast decision can be made and

broadcast is used to transmit SPs. Under thesecircumstances, a node desiring a path to a desti-

nation emits a SP and a broadcast query and


await-reply cycle takes place. Each SP is tagged

with a packet identifier that is used along with the

source address to avoid processing a smart packet

more than once by any node. Therefore, a node

stores the source address and packet identifier of

each SP that it receives and discards any sub-sequent copy. Also, for any source–destination

(S,D) pair, the identity of the most recently re-

ceived SP is stored. As a SP moves in the network,

it collects the addresses of and arrival times at each

node. The first arriving SP at the destination

generates an ACK that heads back to the source to

inform the discovery and travels on the reverse

path recorded by the SP. Battery information isrecorded by ACKs and not by SPs for conve-

nience, as it will be shown later.

3.2. Information structures

CPN use a learning system whose ability to

create and maintain routes depends on the avail-

ability of accurate network information. Packetsacquire this information as they move, and it is

stored in several structures in a distributed fash-

ion. These structures are: route caches, cognitive

maps, mailboxes, weight parameters, and neighbor

tables.

Route caches are located at source nodes; they

store complete paths for every destination as nee-

ded. As detailed before, SPs create paths as a resultof their exploration of the network. The path

information arrives at the source via ACKs. To

maintain the quality of a connection, SPs are

continuously sent out by the source as a small

fraction of the data traffic, possible resulting in a

route switch to keep or improve the performance

of data packets flow. As a result, only active

connections are maintained.A cognitive map exists within each packet to

store addresses and network metrics. SPs dynam-

ically construct their cognitive maps with the ad-

dresses of the nodes that they visit. Dumb packets

and acknowledgments use their cognitive map to

list the complete path they follow. Also, SPs and

DPs record their arrival time at each node in their

cognitive map. ACKs distribute the cognitive mapinformation along with battery information to the

nodes on an active path.

ACKs visit the nodes on the path listed in their

cognitive map and update their mailboxes. A

mailbox is a structure within a node that maintains

statistics about the performance of active connec-

tions. Mailboxes keep the average round-trip delay

and other metrics to every known destinationthrough every neighbor. This information is re-

quired to operate the CPN algorithm that uses

reinforcement learning to update weight tables of

random neural networks stored within the node.

Neighbor tables are maintained at every node

with entries that define all known neighbors. As

explained before, neighbors are acquired by lis-

tening to their transmissions. Each entry in thetable details the address of the neighbor and an

expiration time that is used to remove nodes that

are no longer present.

3.3. Unicast routing

CPN routing decisions are performed with the

aid of a random neural network, which is trainedwith a reinforcement-learning algorithm. AHCPN

employs a procedure virtually identical to CPN to

make unicast routing decisions, but with a small

difference. CPN replaces a small fraction of deci-

sions with random decisions to avoid trapping the

algorithm in local minima. In AHCPN, we replace

random decisions with broadcasts to allow the

algorithm a better exploration of the mobile net-work. We will omit further details about CPN as

the algorithm is well documented in the literature

[2–5].

3.4. Specific routing goals for ad hoc networks

Ad hoc networks are characterized by a high

transmission error probability, which is caused bymobility, the use of wireless links, and the limited

resources of nodes. We introduce a new routing

metric, in AHCPN, that takes into account the

quality of the links and nodes involved in a con-

nection. We call this metric ‘‘path availability’’. A

combined routing goal, which endeavors to max-

imize path availability and minimize the round-

trip delay of the packets, can offer a good balanceamong the selection of fast paths, the use of stable

paths, and a better use of network resources.


Path availability is defined as the probability to

find nodes and links available for routing on a

path. Formally, suppose that a SP takes the path:

(n1; n2; . . . ; nd), where ni represents the ith node on

that path and (ni; niþ1) represents the link between

nodes ni and niþ1.Assume that node ni is available for routing

with probability PnðniÞ and that link (ni; niþi) is

available with probability Plðni; niþ1Þ. Path avail-

ability Ppðni; ndÞ, from node ni to node nd , is ex-

pressed as

Ppðni; ndÞ ¼Yd�1

j¼i

Pnðniþ1ÞPlðni; niþ1Þ: ð1Þ

A recursive, combined routing goal function Gid

consisting of round-trip delay from ni to nd and

path availability information is defined as

Gid ¼ Ppðni; ndÞDðni; ndÞ þ ½1� Ppðni; ndÞ�ðTo þ GiÞ;ð2Þ

where Dðni; ndÞ is the round-trip delay from ni to ndand To is a predefined timeout interval. To can be

interpreted as the delay before a packet, which has

been lost, can be retransmitted.

In real systems, many factors contribute to the

values of Pn and Pl. For example, the processing

load of the nodes, type of obstacles, nature of

working environment, etc. In this paper and asexplained next, Pn will be expressed only in terms

of the remaining battery lifetime of the nodes. We

will assume Pl ¼ 1.

3.5. Energy sensitivity

Batteries are the major source of energy in

mobile nodes. To provide greater portability,batteries need to be small and lightweight, which

unfortunately restricts the total energy that they

can carry. Once batteries exhaust their energy,

they need to be replaced or recharged, which typ-

ically reduces the independence of a mobile node

to a few hours of operation.

Energy consumption, in communication-related

tasks, depends on the communication mode of anode. A node may be transmitting, receiving, or in

idle mode. Naturally, transmission consumes more

energy than the other two modes. From the rout-

ing perspective, our interest is in selecting routes in

such a way that the transmission and reception of

packets is intelligently distributed on the network

so as to maximize the overall average battery

lifetime of the nodes. Therefore, we are interested

in getting SPs to select, with greater frequency,

those nodes which have the longest remainingbattery lifetime.

If Bi represents the remaining battery lifetime of

node ni, PnðniÞ can be expressed as

PnðniÞ ¼Bi

Bm; ð3Þ

where Bm is the lifetime of a fully charged battery.

Assume Plðni; niþ1Þ ¼ 1 in Eq. (1), for all

i 2 ½1; d � 1�, to simplify this discussion. Thus,


j¼i

Pnðniþ1Þ: ð4Þ

From (4) and (3), the path availability to destina-

tion nd from a node on the path becomes


j¼i

Bjþ1

Bm:

4. AHCPN protocol description

This section presents a detailed description of

the AHCPN protocol. AHCPN packets are

structured in three main areas: header, cognitive

map, and payload. Payload is only carried in DPs.

Fig. 1 illustrates the format of the packets:

1. Type (4 bits)––identifies the packet type:

0¼ dumb packet, 1¼ smart packet, 2 and 3

indicate acknowledgments originated by dumb

and cognitive packets respectively.

2. QoS (4 bits)––indicates the desired quality of

service for the packet. The quality-of-service

goal determines the reinforcement learning

function to be used to update the random neu-ral network in the CPN algorithm.

3. Header and cognitive map length in words of 32

bits.

4. Cognitive map cursor––indicates the position in

the cognitive map of the node transmitting the

packet.

Reserved

Fig. 1. Packet format in AHCPN.


5. Packet identification (32 bits)––uniquely distin-

guish the packet. Acknowledgments carry the

same packet identification as their originators

(dumb or smart packets). The identification is

used at the source node to remove packets wait-

ing for retransmission, after the correspondingacknowledgment arrives. In addition, smart

packets leave a copy of their packet identifica-

tion in the nodes to identify previously visited

nodes.

6. Destination address (32 bits)––uses an IPv4

addressing space.

7. Cognitive map (variable)––is an area where

packets store information about the network.

A cognitive map in AHCPN consists of the fol-

lowing field:

• Source and intermediate hop addresses––usesIPv4 address format. Smart packets attach a

new record with every node that they visit.

On the other hand, dumb and acknowledg-

ment packets carry the cognitive map given

at the source node.

• Timestamp fields––record the arrival time at

each intermediate node. The entry corre-

sponding to the source node indicates thepacket’s departure time.

• Path availability information field––expresses

the probability to find all nodes and links

available for routing from the present loca-

tion to the destination (calculated from bat-

tery lifetime of nodes).

8. Payload––area to transport IP datagrams.

4.1. Operation

DPs source route datagrams to their desti-

nations, using the paths stored in the route

cache of the source node. If a route to a des-

tination is yet to be determined, datagrams wait

in a queue while SPs seek the route. Fig. 2illustrates the logic in the generation of packets

at source nodes. Until a route to a requested

destination is discovered, source nodes continu-

ously send out smart packets while datagrams

arrive from their upper layer and are locally

stored.

SPs use either unicast or broadcast to propagate

on the network. They decide what method to usedepending on the information available in the

node where they are located. When the available

information is not sufficient to make a unicast

(RNN based) decision, then broadcast is used

(Fig. 3). Note that, at least three neighbors are

required to use the RNN algorithm at any inter-

mediate node. One of the neighbors is simply the

one from which the packet was received and doesnot participate in the computation (split-horizon

principle). At the source node at least two neigh-

bors are required.

Packet arriving from IPlayer

Fig. 2. Generation of packets at source nodes.

Fig. 3. Decision logic in smart packets.


After a route to a new destination has been

established, additional smart packets may depart

to maintain (or improve) the quality-of-service of

the connection. The additional smart packets are

generated as a small percentage of dumb packet

Fig. 4. Smart packet

rate. Fig. 4 illustrates the flow of smart packet-

related processes. Note that smart packets may get

destroyed if a node’s resources do not meet their

needs. This happens, for example, when the

remaining battery lifetime of the node is too low or

when a replica of the same request has already

visited the node.

flow diagram.

Fig. 5. Dumb packet flow diagram.


To accelerate the establishment of bidirectional

connections, a smart packet creates a route back

to the source node if no such route exists yet in

the destination’s route cache. As in CPN, an

acknowledgment departs toward the source nodeon the reverse route described in its cognitive map.

As they travel, acknowledgments update mail-

boxes along the path and the route cache at the

source node. If there are packets awaiting for a

route to be discovered at the source, they depart

immediately after the first acknowledgment for

that connection arrives.

A flow control diagram for dumb packets isdepicted in Fig. 5. As in CPN, dumb packets are

given the complete path to transport payload (in

this case, datagrams). A copy of the original

data remains within the source node until the

packet is successfully delivered to its destination.

The acknowledgments that are originated by

dumb packets confirm to the source the deliv-

ery of packets and update mailboxes along thepath.

Dumb packets collect timestamps as they tra-

vel on the network to keep the mailboxes up-

to-date whereas their acknowledgments collect

battery related and link-quality information on

the path.

5. Simulation experiments

An implementation of the AHCPN algorithm

was developed and integrated into the popular

Network Simulator-2 (NS-2). Our experimentsconsisted in observing the establishment and use of

routes in a dynamic network, which incorporated

random node motion. Each experiment simulated

the operation of the network for 15 min.

We employed 50 nodes divided into two popu-

lations. The first population consisted of 10 nodes

that started the experiment with full battery

charge. A full battery charge allows up to 2 h ofoperation in the lack of communication related

tasks. It was assumed that batteries of all mobile

nodes were all identical with a maximum capacity

of 2 Wh. A mobile node was assumed to consume

10 W to uphold non-communication related

operations. The second population (40 nodes)

started the experiment with 1/8 of the full capacity

of their batteries, which provides 15 min of oper-ation excluding communication tasks.

All nodes started the experiment at a random

location within a rectangular working area of

1500 · 500 meters and moved as defined by the

random waypoint model. For this, each node se-

lected a random destination within the working

Fig. 6. Average number of nodes whose battery powder is

completely depleted, as a function of simulated time.

Fig. 7. Average number of data packets that were delivered to

their destination.


area and moved linearly to that location at a speed

of 2 m/s (approximately equivalent to walking

speed). After reaching their destination, the nodes

selected a new random location with no pause.

Five connections were established and removed

during the simulation of each experiment. Theseconnections were performed only among nodes

from the first population to avoid having shortage

of energy at sources or sinks.

Energy consumption for communication re-

lated tasks was recreated after the linear model by

Feeney and Nilsson [32]. The model predicts the

energy consumption for reception and transmis-

sion of packets. The model takes into account thetype of transmission (unicast or broadcast) and

size of packets. To speed up our simulations and

reduce storage requirements, the energy required

to send and receive packets was assumed to be 100

times larger than the value predicted by the Fee-

ney’s model (consequently, assuming the use of

energy inefficient network interface cards).

The reports that follow show the average valueof the observed variable and the 95% confidence

interval of 100 runs of the experiment. We com-

pare the operation of AHCPN with the response

of a source routing, flood-based ad hoc algorithm

where broadcasts packets are emitted to discover

routes while acknowledgments inform the instan-

taneous fastest route (in a way similar to DSR or

AODV).Under AHCPN, smart packets were sent at a

ratio of 0.01 (on average, one SP every 100 DPs).

Their objective function included both round-trip

delay and battery information as detailed in Sec-

tion 3. To obtain a fair comparison, the flood-

based model sent broadcast packets at the same

rate as that of smart packets.

Fig. 6 reports the average number of nodesfrom population 2 whose battery power is com-

pletely depleted, as a function of the simulated

time. A flooding-based algorithm selects fast

routes without considering the availability of en-

ergy in the nodes; this quickly drains the battery

of a large number of nodes. The learning capacity

of smart packets makes a more efficient use of

resources. However, the total number of datapackets that get to arrive to their destinations is

slightly shorter when using AHCPN (Fig. 7),

which can not be avoided because AHCPN re-

quires more time to adapt to changes in topology.

Flooding provides faster adaptation but have to

pay a high price in consumption of resources.

Figs. 8 and 9 report the average total number of

packets that were transmitted and received during

the simulation. The figures illustrate the aggre-gated number of data packets and control packets.

Flooding produces many more packets than

AHCPN, so a higher consumption of resources is

to be expected with flooding.

Fig. 8. Average number of packets (data packets and control

packets) transmitted during the experiment.

Fig. 9. Average number of packets (data packets and control

packets) received during the experiment.


6. Conclusions

This paper has specified as evaluated the

AHCPN protocol, which uses our previous re-

search on Cognitive Packet Networks (CPN) to

produce an innovative solution that supports the

operation of a mobile, ad hoc network.

In AHCPN, SPs use broadcasts to create a total

or partial flooding that allows nodes acquire

neighboring information while SPs move on thenetwork since flooding is expensive in terms of

resource utilization. Whenever possible, SPs use

unicast-based transmissions based on the CPN

routing algorithm.

We have introduced a new routing metric to

cope with the limited power resources of mobile

networks. Path availability, which models theprobability to find available nodes and links on a

path, was employed to take into account the en-

ergy available at nodes. The end result is that

packets choose nodes which have longer remaining

battery lifetime with a higher probability than

nodes with shorter remaining lifetime. By com-

bining path availability and round-trip delay in a

single goal function, selected paths are still of alength close to the shortest path length.

We have discussed a formal definition of an

Ad Hoc Cognitive Packet Network protocol that

provides an implementation framework to test our

ideas, and a simulation model was developed and

integrated into the Network Simulator 2 (NS-2).

Our results show that the CPN protocol (1) is able

to dynamically discover neighbors and routes, (2)can discover and maintain routes without the need

of a large number of broadcasts, (3) will distribute

network traffic so as to extend the battery lifetime

of the nodes and (4) maintains a comparable per-

formance to more energy consuming, broadcast

based approaches.

Acknowledgements

The authors gratefully acknowledge the support

that made this research possible via the Engineer-

ing and Physical Science Research Council (UK)

under Grant GR/S52360/01, by US Army Stricom

via NAWC under Contract No. N61339-02-

C0117, by NSF under Grant No. EIA0203446,and by the US Army Research Office under Con-

tract No. DAAD190310135.

References

[1] C.E. Jones, K.M. Sivalingam, P. Agrawal, J.-C. Chen, A

survey of energy efficient network protocols for wireless

networks, Wireless Networks 7 (4) (2001) 343–358.

[2] E. Gelenbe, R. Lent, Z. Xu, Networks with cognitive

packets, in: Proceedings of the Eight International


Symposium on Modeling, Analysis and Simulation of

Computer and Telecommunication Systems (IEEE Com-

puter Society) San Francisco, CA, 2000, Opening Key-

Note Paper, pp. 3–12.

[3] E. Gelenbe, R. Lent, Z. Xu, Toward networks with

cognitive packets, in: K. Goto, T. Hasegawa, H. Takagi,

Y. Takahashi (Eds.), Performance and QoS of Next

Generation Networking, Springer, London, 2000.

[4] E. Gelenbe, R. Lent, Z. Xu, Measurement and perfor-

mance of cognitive packet networks, Computer Networks

37 (2001) 691–701.

[5] E. Gelenbe, R. Lent, Z. Xu, Design and performance

of cognitive packet networks, Performance Evaluation 46

(2–3) (2001) 155–176.

[6] E. Royer, C.-K. Toh, A review of current routing protocols

for ad-hoc mobile wireless networks, IEEE Personal

Communications 6 (2) (1999) 46–55.

[7] C.E. Perkins, P. Bhagwat, Highly dynamic destination-

sequenced distance-vector routing (dsdv) for mobile com-

puters, in: Proceedings of the ACM Conference on

Communications Architectures, Protocols and Applica-

tions, London, England, 1994, pp. 234–244.

[8] X. Chen, L. Qi, D. Sun, Global and superlinear conver-

gence of the smoothing Newton method and its application

to general box constrained variational inequalities, Math-

ematics of Computation 67 (222) (1998) 519–540.

[9] S. Murthy, J. Garcia-Luna-Aceves, An efficient routing

protocol for wireless networks, Mobile Networks and

applications 1 (2) (1996) 183–196.

[10] C. Chiang, H. Wu, W. Liu, M. Gerla, Routing in clustered

multihop, mobile wireless networks, in: Proceedings of the

IEEE Singapore International Conference on Networks,

1997, pp. 197–211.

[11] G. Pei, M. Gerla, T.-W. Chen, Fisheye state routing: A

routing scheme for ad hoc wireless networks, in: Proceed-

ings of ICC 2000, New Orleans, IEEE, 2000.

[12] D.B. Johnson, D.A. Maltz, Y.-C. Hu, J.G. Jetcheva, The

dynamic source routing protocol for mobile ad hoc

networks, IETF Draft, March 2001.

[13] C.E. Perkins, E.M. Royer, Ad hoc on demand distance

vector routing, in: 2nd IEEE Workshop on Mobile

Computing Systems and Applications, 1999.

[14] C.E. Perkins, S.R. Das, Ad hoc on-demand distance vector

(aodv) routing, IETF Draft, March 2001.

[15] C.E. Perkins, Ad Hoc Networking, Addison-Wesley,

Reading, MA, 2000.

[16] C.-K. Toh, A novel distributed routing protocol to support

ad-hoc mobile computing, in: Proceedings of the 1996

IEEE Fifteenth Annual International Phoenix Conference

on Computers and Communications, 1996, pp. 480–486.

[17] R. Dube, C. Rais, K. Wang, S. Tripathi, Signal stability

based adaptive routing (ssa) for ad hoc mobile networks,

IEEE Personal Communication 4 (1) (1997) 36–45.

[18] S. Corson, S. Papademetriou, P. Papadopoulos, V. Park,

An Internet manet encapsulation protocol (imep) specifi-

cation, IETF Draft, August 1998.

[19] Z.J. Haas, M.R. Pearlman, P. Samar, The intrazone

routing protocol (iarp) for ad hoc networks, IETF Draft,

January 2001.

[20] Z.J. Haas, M.R. Pearlman, P. Samar, The interzone

routing protocol (ierp) for ad hoc networks, IETF Draft,

January 2001.

[21] Y.-B. Ko, N.H. Vaidya, Location-aided routing (lar) in

mobile ad hoc networks, in: Proceedings of ACM/IEEE

Mobicom, Dallas, TX, 1998, pp. 66–75.

[22] Y. Ko, N. Vaidya, Geotora: A protocol for geocasting

in mobile ad hoc networks, in: Proceedings of the 2000

International Conference on Network Protocols, Osaka,

Japan, 2000.

[23] M. Bhardwaj, R. Min, A. Chandrakasan, Power-aware

systems, in: Proceedings of 34th Asilomar Conference on

Signals, Systems and Computers, November 2000.

[24] A. Sinha, A. Chandrakasan, Energy efficient real-time

scheduling, in: Proceedings of the International Conference

on Computer Aided Design (ICCAD), San Jose, Novem-

ber 2001.

[25] H. Mehta, R. Owens, M. Irwin, R. Chen, D. Ghosh,

Techniques for low energy software, in: Proceedings of the

International Symposium on Low Power Electronics and

Design, August 1997.

[26] L. Gruenwald, S.M. Banik, A power-aware technique to

manage real-time database transactions in mobile ad hoc

networks, in: 4th International Workshop on Mobility

in Database and Distributed Systems, International Con-

ference on Database and Expert Systems Applications

(DEXA), September 2001.

[27] S. Singh, M. Woo, C.S. Raghavendra, Power-aware

routing in mobile ad hoc networks, in: Proceedings of the

Fourth Annual ACM/IEEE International Conference on

Mobile Computing and Networking, 1998, pp. 181–190.

[28] C.-K. Toh, Maximum battery life routing to support

ubiquitous mobile computing in wireless ad hoc networks,

IEEE Communications Magazine 39 (6) (2001) 138–147.

[29] Q. Li, J.A. Aslam, D. Rus, Online power-aware routing

in wireless ad-hoc networks, in: Mobile Computing and

Networking, MobiComm 2001, 2001, pp. 97–107.

[30] M. Maleki, K. Dantu, M. Pedram, Power-aware source

routing protocol for mobile ad hoc networks, in: Proceed-

ings of the 2002 International Symposium on Low Power

Electronics and Design, 2002.

[31] T. Xu, J. Heidemann, D. Estrin, Adaptive energy-conser-

ving routing for multihop ad-hoc networks, USC/ISI

Research Report 527, October 2000.

[32] L.M. Feeney, M. Nilsson, Investigating the energy con-

sumption of a wireless network interface in an ad hoc

networking environment, in: IEEE INFOCOM, 2001.

power-aware ad hoc cognitive packet networks

Documents