Design of a Multi-Agent System for Hierarchical

Network Management in Wireless Sensor Network

Mubashsharul I. Shaﬁque, Haiyi Zhang and Yifei Jiang

Jodrey School of Computer Science, Acadia University

Wolfville, NS, B4P 2R6, Canada

Abstract. In order to manage a sensor network efﬁciently, we can divide it logi-

cally into disjoint parts, called clusters. As sensor nodes are resource-constraint,

it is desirable to chose suitable cluster heads and do role-switching of the cluster

heads wherever appropriate. In this work, we design a system that selects new

cluster head through collaboration of multiple software agents in each cluster.

Our system allows to pick up cluster heads dynamically based on current net-

work status. Agents in our design use Fuzzy Logic-based controller to ﬁnd new

cluster heads.

1 Introduction

Wireless Sensor Network is a deployment of sensor nodes that do data reporting to in-

terested user via sink node. It offers a great facility to remotely monitor an unattended

environment. Due to this feature, sensor networks are widely used in military surveil-

lance, habitat monitoring, industrial plants, and in many other places. Once a sensor

network is deployed, it is necessary to manage it efﬁciently in order to optimize net-

work life-time.

A sensor network can be managed by hierarchical or ﬂat organization. However,

as hierarchical organization has several beneﬁts over ﬂat network structure, normally a

sensor network is divided into clusters. Here in this work, we design a multi-agent sys-

tem to rotate the role of cluster head nodes. Software agents in our design autonomously

collaborate with each other in the distributed sensor network environment. The remain-

ing part of this paper is organized as follows: section 2 presents some background

knowledge and section 3 deﬁnes the problem that his work addresses. Next, section

5 explains the design of our system and inter-agent communication is presented in sec-

tion 6. We conclude our paper in section 7.

2 Background Knowledge

2.1 Wireless Sensor Network

Wireless Sensor Network(WSN) consists of autonomous sensor devices (known as

nodes or motes) that can sense an environment. For example, MicaZ sensor nodes from

Shaﬁque M., Zhang H. and Jiang Y. (2010).

Design of a Multi-Agent System for Hierarchical Network Management inWireless Sensor Network.

In Proceedings of the 6th International Workshop on Artiﬁcial Neural Networks and Intelligent Information Processing, pages 53-62

 SciTePress

Crossbow

can sense ambient light, barometric pressure, GPS, magnetic ﬁeld, sound,

photo-sensitive light, photo resistor, humidity and temperature.

Often times, these sensor nodes are battery-powered, and equipped with limited

processing capacity and memory. However, WSNs are very effective and efﬁcient at

monitoring remote environment and communicate real-time data. Sensor nodes can de-

tect events or phenomena, collect and process data, and transmit sensed information to

the interested users.

2.2 Virtual Organization in Sensor Network

WSN differs from an IP network in regards to network backbone because, often times a

WSN does not have any ﬁxed infrastructure. Sensor nodes may run out of battery power

or network topology may change due to some mobile nodes. So, in order to support

data routing, WSN relies on virtual infrastructure which forms on-the-ﬂy during system

operation. One way to manage a sensor network, is to logically divide it into some

disjoint clusters. A particular node in each cluster works as cluster head, and gathers

data from other nodes of the cluster. The cluster head then forwards data to the next hop

node towards the data sink, which is the ﬁnal destination of any data. Figure 1 shows

an example sensor network with three clusters. We show data reporting from a normal

sensor node to the sink by arrow heads.

Fig. 1. Virtual organization in WSN.

3 Problem Deﬁnition

In a static network topology, cluster heads of a sensor network die quite early due to

excessive relaying of data stream towards sink. So, in order to prolong network life-

time, a network should rotate the role of cluster heads. For example, if a current cluster

head had drained a signiﬁcant amount of energy, network should keep the provision to

pick a new cluster head for that cluster. This kind of dynamic load balancing can delay

the ﬁrst node death and help decrease data packet loss in the network thereby.

Due to the large number of sensor nodes in a typical sensor network, it is not real-

istic to run any centralized algorithm to switch cluster heads. A network needs to have

distributed architecture which can select new cluster heads with local information, and

further, the role-switching should remain quasi-transparent to the remaining network.

http://www.xbow.com/.

4 Related Works

For Wireless sensor Networks (WSNs), the major design goal is to minimize energy

consumption and maximize the network lifetime. In the last few years, plenty of at-

tempts of exploring advanced power conservation approaches have been used by re-

searchers for wireless sensor networks. Cluster-based routing approach is one of the

famous energy efﬁcient routing approaches in WSNs. LEACH (Low Energy Adaptive

Clustering Hierarchy) in [1], is the ﬁrst hierarchical cluster-based routing protocol for

WSNs. This algorithm uses random and periodic rotation of the Cluster Heads (CHs)

for load balancing, which can evenly disperse the energy load among the sensor nodes

in the network. More speciﬁcally, the cluster heads belonged to the corresponding clus-

ters will only use for certain number of rounds. After a predeﬁned round, new cluster

heads will be randomly generated, which is based on a role that if the random number

is less than a calculated threshold T (n), the corresponding sensor node will be selected

as the cluster-head for the current round. This randomized periodic role rotation en-

sures that all the nodes are equally likely to be cluster head nodes. However, simply

random cluster head rotation will select unfavorable cluster heads, which will turn out

high energy consumption in later rounds.

Due to the above reason, in [9], Energy-LEACH protocol improves the procedure

of CH rotation. Similar to LEACH protocol, the process of CH rotation of E-LEACH

is divided into rounds. In the ﬁrst round, each sensor node has the same probability

to be turned into CH, which means sensor nodes are randomly selected as CHs. In

the next rounds, since the residual energy of each sensor node is different after every

communication round, residual energy of node is considered as the main metric that

decides whether the sensor nodes turn into CHs or not after the ﬁrst round. The sensor

nodes who have more remaining energy will become CHs rather than the ones with less

remaining energy.

As for the aforementioned two LEACH algorithms, there is a shared drawback, that

is, both of them consumes more CPU cycles, since each sensor node in WSNs has to

calculate the threshold and generate the random numbers in each round. To overcome

this shortcoming, a improved LEACH, called LEACH-C protocol, is proposed in [10].

LEACH-C is a centralized clustering algorithm and uses the same steady-state phase as

in LEACH algorithm. This protocol also produces better performance on cluster heads

rotation. During the set-up phase of LEACH-C, each sensor node sends information

about its current location (this may determine by using GPS) and residual energy level

to the Base station (BS). In order to ensure that the energy load is evenly distributed

among all the sensor nodes in WSNs, The average node energy is computed by the BS,

and then, determines which sensor nodes have energy below this average. Once the CHs

and associated clusters are found, the BS broadcasts a message that obtains the cluster

head ID for each node. If a cluster head ID matches its own ID, the node will be selected

as a cluster head; otherwise the node determines its TDMA slot for data transmission

and goes sleep until it’s time to transmit data. The steady-state phase of LEACH-C is

identical to that of the LEACH protocol.

Based on the above approach, in [11], cluster heads election using fuzzy logic

is proposed, which can minimize energy consumption and provide a substantial in-

crease in network lifetime compared with the probabilistically cluster heads selecting

approaches. According to this proposed approach, for a cluster, the node elected by the

base station is the node having the maximum chance to become the cluster-head, which

is based on three fuzzy descriptors: energy level in each sensor node, sensor node con-

centration and node centrality with respect to the entire cluster. The operation of this

fuzzy logic cluster-head election scheme is divided into two rounds with each consisting

of a setup and steady state phase similar to LEACH algorithm. During the setup phase,

fuzzy knowledge processing is used for determining the CHs, and then the cluster is

organized. In the steady state phase, the aggregated data is collected by CHs. After that,

CHs perform signal processing functions to compress the data into a single signal. This

composite signal is then sent to the base station. Fuzzy logic control model is core part

of this proposed approach; it includes a fuzziﬁer, fuzzy rules, fuzzy inference engine,

and a defuzziﬁer. To be speciﬁc, fuzziﬁer is used to take the crisp inputs from each

of variables of energy, concentration and centrality and determine the degree to which

these inputs belong to each of the appropriate fuzzy sets. Then, these fuzziﬁed inputs

are applied to the antecedents of the fuzzy rules. As for the defuzziﬁcation, the input

for this process is the aggregate output fuzzy set chance and the output is a single crisp

number. For fuzzy inference engine, Mamdani Method [13] is commonly used. In [12],

similar cluster heads selections based on fuzzy logic algorithm are also presented.

To our knowledge, several attempts have also been used by some researchers to re-

duce energy consumption based on mobile agents [2][3][4][5]. Due to the constraints

of bandwidth in wireless sensor network, the network’s capacity may not satisfy the

transmission of sensory data. In order to handle the problem of overwhelming data traf-

ﬁc, Qi, et al. [6] proposed Mobile Agent-based Distributed Sensor Network (MADSN)

for multi-sensor data fusion. For this proposed approach, it not only achieves data fu-

sion, but also reduces energy expenditure. However, the application of this approach

can only be applied on cluster-based topologies. MADD approach in [7] is introduced

to deal with this problem. Currently, most energy-efﬁcient proposed approaches are fo-

cused on data-centric model, such as the directed diffusion. By selecting good path to

drain quality data from source nodes, directed diffusion approach can achieve substan-

tial energy gain. However, it still allows redundant sensory trafﬁc to ﬂow back to the

Base station. The main advantage of MADD is to reduce the redundant sensory data.

Through using mobile agent, data is aggregated at each source node and is brought back

to sink. This allows substantial energy gain toward the network lifetime.

To explain the process of MADD approach, it starts when the mobile agent is dis-

patched from the BS with the interest and ends when it returns to the sink with the

aggregated data. The processes involved in MADD are divided into three phases. First,

the mobile agent is dispatched from BS to the ﬁrst source node. Second, the mobile

agent shifts from the ﬁrst source node to the last source node, visiting selected source

nodes in between. The drawback for this approach is that it doesn’t always guarantee

the best sequence of nodes to be visited.

To deal with the aforementioned limitations, Shakshuki et al. in [8] proposed a

mobile agent for efﬁcient routing approach (MAER) by using both Dijkstra’s algorithm

and Genetic Algorithms (GAs). As we know, the order of source nodes to be visited by

the mobile agent greatly affects the energy consumption. Although MADD, the work

presented in [7], allows the agent to autonomously select visit sequence of source nodes

for achieving data aggregation, it does not always provide an optimal sequence. To

address this shortcoming, MAER in [8] introduces Genetic Algorithm (GA) to produce

an optimal route.

5 Agent Architecture

An agent in our design consists of four modules: Problem Solver, Knowledge-base

Updater, Scheduler, and Communicator. Further, each agent maintains own knowledge

base in its memory, which gets updated as a result of inter-agent message communi-

cation. Here in this section, we describe different components of individual agent in

detail.

Fig. 2. Agent Architecture.

In our design, the role of an agent software switches between two modes- Cluster

head Mode and Normal Mode. However, at any particular time, within each cluster,

agent software of only one sensor node works in Cluster head Mode, others execute in

Normal Mode.

5.1 Scheduler Module (SM)

This module is in charge of scheduling a new round of cluster head selection. It is active

in Cluster head Mode only. It interacts with Problem Solver Module and Communicator

Module. After periodic interval, it checks the remaining energy of the cluster head node

and initiates new cluster head selection if necessary.

5.2 Problem Solver Module (PSM)

This is the basic working module of an agent. Based on current working mode, PSM

does two distinct tasks:

Weight Calculation in Normal Mode. We deﬁne the weight of a sensor node as a

function of two parameters: Centrality(C) and Remaining Energy(RE). As sensor nodes

are resource-constraint, we use simple Fuzzy Logic based controller in agents to cal-

culate node weight. During Fuzziﬁcation phase, we followed the mapping presented in

[11] to convert crisp values to fuzzy values. Figure 3 portrays this Fuzziﬁcation process.

We deﬁne our own rule-base for fuzzy controller in table 1. PSM uses this rule base to

determine current weight value in fuzzy variable.

Fig. 3. Fuzzy Sets used in Fuzzy Controller.

Table 1. Fuzzy Rule Base.

Rule Centrality Energy Result

1 Close High High

2 Adequate High High

3 Far High High

4 Close Medium High

5 Adequate Medium Medium

6 Far Medium Medium

7 Close Low Medium

8 Adequate Low Low

9 Far Low Low

New Cluster Head Selection in Cluster head Mode. PSM of cluster head agent op-

erates on all Fuzzy weight values of normal sensor nodes. After comparing multiple

Fuzzy weight values, if some nodes are equally suitable to be the new cluster head,

PSM runs following scheme to break a tie:

In order to Defuzzify, we assign numeric values against different Fuzzy values

High/Close=3, Medium/Adequate=2, and Low/Far=1 and bias the decision by putting

60% importance to energy. For example, if two nodes n

and n

are in a tie with Fuzzy

values {Far, Medium} and {Close, Low} respectively, then PSM computes Defuzziﬁed

values as d

= 1*0.4 + 2*0.6 = 1.6 and d

= 3*0.4 + 1*0.6 = 1.8. Here d

and d

are

crisp values for node n

and n

respectively. Even after this computation, if PSM can

not break the tie, it compares node IDs and selects node with minimum ID as next

cluster head.

5.3 Communicator Module (CM)

This module is in charge of sensor radio component and communication with other

agents in the system. If agent software is running in Cluster head Mode, it does two

things: (a) Upon receiving request from SM, it sends out Discovery Message to other

agents; (b) If it receives Weight Message(s), it forwards weight value(s) to PSM.

On the other hand, if the agent is working in Normal Mode, CM does only one thing:

it receives Weight Message(s) and forwards to the Knowledge-base Updater Module.

5.4 Knowledge-base Updater Module (KBUM)

This is the only module that interacts with agent Knowledge-base. KBUM uses a ded-

icated area of sensor node’s memory to maintain the Knowledge-base. Any change in

shared knowledge like- weight updates or rule updates are propagated to it by PSM

and/or CM; and as a result, KBUM synchronizes agent’s local memory with global

state.

5.5 Knowledge-base (KB)

Knowledge-base in an agent holds a local copy of the shared knowledge. An agent’s

KB houses a snapshot of a fragment of global information, that it is interested in. KB

includes rules for Fuzzy controller, latest weights of all neighbor nodes and current

cluster head’s node ID.

6 System Dynamics

6.1 Protocol Message Types

In a collaborative multi-agent system, message passing is an integral part which facili-

tates distributed processing. In this section, we present four different message types that

agents in our design exchange during system operation.

Discovery Message. Once remaining energy of the current cluster head goes below

a threshold value (which is 40% of initial energy in our implementation), cluster head

agent broadcasts Discovery Message. A Discovery Message initiates new cluster head

selection within a cluster.

Weight Message. Weight Message contains latest weight value of a node. This mes-

sage is broadcasted by agent in a normal node as a response to an incoming Discovery

Message from current cluster head.

Control Message. Control Message is used to broadcast any change in routing topol-

ogy. After determining new cluster head, current cluster head agent broadcasts Control

Message indicating new cluster head node ID.

Data Message. Data Message originates from normal node agents and is used to report

current data readings to the cluster head.

6.2 System Work-ﬂow

Fig. 4. Sequence diagram for cluster head mode.

In this section we explain the sequences of operation in our agent-based system.

First, ﬁgure 4 shows action sequences for an agent operating in Cluster head Mode.

After periodic interval, SM measures remaining energy of the cluster head node, and if

it is bellow the threshold (i.e., less than 40% of initial energy), SM starts a new cluster

head selection round. It notiﬁes PSM that a new cluster head selection process is in

effect and also requests CM to send out Discovery Message.

CM then sends out Discovery Message in the wireless medium, and in response, it

receives Weight Messages from other agents. CM forwards these node-weights to PSM

for calculation. In turn, PSM requests KBUM to update Knowledge base with newly

received weight values and once PSM receives weights from all nodes, it uses Rule

base to select new cluster head as explained in section 5.2. After ﬁnding the new cluster

head, PSM requests CM to convey other nodes about new cluster head, and as a result,

CM sends out new Control Message containing newly chosen cluster head ID.

Figure 5 portrays action sequence of an agent working in Normal Mode. If its CM

receives a Discovery Message from current cluster head, CM requests PSM to compute

own node weight. Upon computation, PSM returns own weight to CM. CM then en-

capsulates this weight value in a Weight Message and broadcasts that message in the

neighborhood. Besides, if CM receives Weight Message from a different node, it re-

quests KBUM to update agent’s Knowledge base. In a similar way, KB update is also

performed if an agent receives Control Message from current cluster head.

Fig. 5. Sequence diagram for normal mode.

7 Conclusions

In this work, we design a system to rotate the role of cluster heads in a Wireless Sensor

Network. Multiple agents, each residing in individual sensor node, interact with each

other and participate in new cluster head selection. In our design, agent-based archi-

tecture provides ﬂexibility to network management. At the same time, as we use fuzzy

controller, this design can be easily extended by adding new rules in the rule-base. Our

future work focuses on implementing this design in TinyOS operating system for Wire-

less Sensor nodes.

References

1. W. Heinzelman, A. Chandrakasan and H. Balakrishnan: Energy-Efﬁcient Communication

Protocol for Wireless Microsensor Networks. In: 33rd Hawaii International Conference on

System Science (2000)

2. Y.-C. Tseng, S.-P. Kuo, H.-W. Lee, and C.-F. Huang: Location tracking in a wireless sensor

network by mobile agents and its data fusion strategies. Computer Journal, vol. 47, no. 4, pp.

448-460, (2004)

3. H. Qi, S.S. Iyengar, and K. Chakrabarty: Multi-Resolution Data Integration Using Mobile

Agents in Distributed Sensor Networks. IEEE Trans. Systems, Man, and Cybernetics Part C:

Applications and Rev., vol. 31, no. 3, pp. 383-391, Aug. (2001)

4. Daniel Massaguer, Chien-Liang Fok, Nalini Venkatasubramanian, Gruia-Catalin Roman,

Chenyang Lu: Exploring sensor networks using mobile agents. pp. 323-325, AAMAS (2006)

5. C.-L. Fok, G.-C. Roman, and C. Lu: Mobile agent middleware for sensor networks: An

application case study. In: 4th Int. Conf. on Information Processing in Sensor Networks

(IPSN’05), pages 382–387. IEEE, April (2005)

6. Hairong Qi, Yingyue Xu, Xiaoling Wang: Mobile-agent-based Collaborative Signal and In-

formation Processing in Sensor Networks. In: Proceeding of the IEEE, Vol. 91, NO. 8,

pp.1172-1183, Aug (2003)

7. Min Chen, Taekyoung Kwon, Yong Yuan, Yanghee Choi and Victor C. M. Leung1: Mo-

bile Agent-Based Directed Diffusion in wireless Sensor Networks. EURASIP Journal on

Advances in Signal Processing, Article ID 36871, (2007)

8. Shakshuki, E., Xing, X.Y. and Malik, H: Mobile Agent for Efﬁcient Routing among Source

Nodes in Wireless Sensor Networks. In: 3rd International Conference on Autonomic and

Autonomous Systems (ICAS’07), June(2007)

9. Xiangning Fan and Yulin Song: Improvement on LEACH Protocol of Wireless Sensor Net-

work. International Conference on Sensor Technologies and Applications (SENSORCOMM

2007), pp.260-264, (2007)

10. W. Heinzelman, A. Chandrakasan and H. Balakrishnan: An Application-Speciﬁc Protocol

Architecture for Wireless Microsensor Networks. IEEE Transactions on Wireless Commu-

nications, Vol. 1, No. 4, October(2002)

11. Indranil Gupta, Denis Riordan and Srinivas Sampalli: Cluster-head Election using Fuzzy

Logic for Wireless Sensor Networks. In: 3rd Annual Communication Networks and Services

Research Conference, pp.255 - 260, (2005)

12. Xiaorong Zhu and Lianfeng Shen: Near optimal cluster-head selection for wireless sensor

networks. Journal of Electronics (China), Science Press, co-published with Springer-Verlag

GmbH, Vol. 4, pp.721-725, November(2007)

13. M. Negnevitsky: Artiﬁcial intelligence: A guide to intelligent systems. Addison-Wesley,

Reading, MA (2001)