Sie sind auf Seite 1von 12

Ad Hoc Networks 10 (2012) 11341145

Contents lists available at SciVerse ScienceDirect

Ad Hoc Networks
journal homepage: www.elsevier.com/locate/adhoc

Connectivity constrained wireless sensor deployment using multiobjective evolutionary algorithms and fuzzy decision making
Pyari Mohan Pradhan , Ganapati Panda
School of Electrical Sciences, Indian Institute of Technology, Bhubaneswar, India

a r t i c l e

i n f o

a b s t r a c t
Deployment of sensor nodes is an important issue in designing sensor networks. The sensor nodes communicate with each other to transmit their data to a high energy communication node which acts as an interface between data processing unit and sensor nodes. Optimization of sensor node locations is essential to provide communication for a longer duration. An energy efcient sensor deployment based on multiobjective particle swarm optimization algorithm is proposed here and compared with that of non-dominated sorting genetic algorithm. During the process of optimization, sensor nodes move to form a fully connected network. The two objectives i.e. coverage and lifetime are taken into consideration. The optimization process results in a set of network layouts. A comparative study of the performance of the two algorithms is carried out using three performance metrics. The sensitivity analysis of different parameters is also carried out which shows that the multiobjective particle swarm optimization algorithm is a better candidate for solving the multiobjective problem of deploying the sensors. A fuzzy logic based strategy is also used to select the best compromised solution on the Pareto front. 2012 Elsevier B.V. All rights reserved.

Article history: Received 23 October 2011 Received in revised form 26 January 2012 Accepted 5 March 2012 Available online 14 March 2012 Keywords: Multiobjective optimization Sensor network Multiobjective particle swarm optimization Sensor deployment

1. Introduction Wireless sensor network (WSN) consist of sensor nodes which are deployed in a large geographical area for collecting data from the sensor eld. In recent years, WSN has been employed in many applications, such as environment monitoring, war zone monitoring, and weather monitoring. Design of these networks has developed over past few years. These networks employ sensor nodes which dynamically organize themselves into a fully connected network. Deployment of sensor nodes is carried out with the goal of maximizing radio coverage and lifetime of the network rather than minimizing the number of nodes due to availability of inexpensive sensor nodes. Positioning of sensor nodes in the region of interest (ROI) is one of the biggest challenges during the design
Corresponding author. Tel.: +91 8895621359.
E-mail addresses: pyarimohan.pradhan@gmail.com (P.M. Pradhan), ganapati.panda@gmail.com (G. Panda). 1570-8705/$ - see front matter 2012 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.adhoc.2012.03.001

of these networks. Optimum placement of sensor nodes results in the maximum possible utilization of the available sensor nodes. The proper choice for the deployment process based on the application requirement is challenging in nature. In a static network, the positioning of sensor nodes is often either human monitored or random. Random deployment of sensor nodes is adopted in some applications because of cost and time. Self-deployment of mobile nodes provides a reliable network where sensor nodes are uniformly distributed in the ROI. Since most often the WSN is employed in some adverse environment or hostile areas, aerial deployment is mostly performed. Once the sensor nodes are deployed in ground, their initial position (co-ordinates) is transmitted to the data processing unit (DPU) for obtaining optimal position. After the processing in DPU, the nodes are informed about optimal positions to move to their corresponding locations. Wireless sensor nodes perform two types of functions: Sensing and Communicating. They can sense anything within the sensing radius (Rsens) and they can communicate

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

1135

with any node within the communication range (Rcomm). The limited size, energy storage and memory of the deployed sensor nodes prevent them from relaying data directly to the DPU. A high energy communication node (HECN) provides the transmission relay to DPU located in an aircraft or satellite etc. Sensor nodes transmit their data to the HECN which in turn transmits those data to the DPU. All nodes must be able to transmit their data to the HECN directly or indirectly. The next section provides a brief overview of various works related to deployment of sensor nodes. In Section 3, the modeling of a wireless sensor network and the objectives for optimization are discussed. The multiobjective evolutionary algorithm (MOEA) approach for solving the sensor deployment problem is introduced in Section 4. The performance metrics used for evaluating the performance of the multiobjective algorithms are discussed in Section 5 along with a fuzzy logic based strategy for selection of best compromised solution. Simulation results of multiobjective particle swarm optimization (MOPSO) in comparison to another popular MOEA is presented in Section 6. Section 7 provides the sensitivity analysis for MOPSO algorithm. Finally some concluding remarks and future work are listed in Section 8. 2. Literature review There are several research work related to the deployment of wireless sensor nodes. Most of them have focused on optimizing the location of sensor nodes for maximizing the coverage of the network [13]. Although single objective is considered in most of the research papers, another objective i.e. energy consumption minimization is also of practical importance in the choice of deployment process. The connectivity between the sensor nodes is another important aspect in design of wireless sensor networks. Wu et al. [4] have applied particle swarm optimization (PSO) algorithm for optimizing the coverage in sensor network deployment. Based on this result, a distance based energy model is used to reduce the cost based on clustering method. The quality of solution obtained by the clustering method solely lies on the output of the PSO algorithm. The energy consumption which is related to lifetime and the coverage are two conicting objectives which need to be optimized simultaneously. This will provide the end user a set of solutions with different trade offs between the two objectives. Wang et al. [5] have used a parallel PSO algorithm for optimizing coverage in a sensor network. If the environment is sufciently known in advance, the sensor nodes can be manually placed. However, availability of complete knowledge of the environment a priori is virtually impossible. Further, a huge amount of energy is spent while obtaining the global information of a geographical area. Thus, deterministic deployment of sensor nodes become impractical due to hostile environment and cost [6]. When the environment is unknown or hostile, sensor nodes are air-dropped or positioned by other means, generally resulting in a random network [3]. Self-deployment algorithms [1,2] have been proposed for the deployment of wireless sensor network. In [1], the

authors have proposed a new approach using virtual force algorithm (VFA) for deployment of nodes in a sensor network, which improves the eld coverage of a sensor node after an initial random placement of sensor nodes. Li and Lei [7] have proposed an improved PSO which provides better coverage than VFA in a wireless sensor network. Subsequently VFA with co-evolutionary PSO was proposed for improving the performance of dynamic deployment optimization [8]. Recently glowworm swarm optimization [9] and articial bee colony algorithm have been adopted for dynamic sensor deployment [10]. Heo and Varshney [2] has proposed an energy-efcient deployment algorithm based on Voronoi diagrams. As mentioned before, most of the work on WSN focuses on maximization of sensing coverage while giving little attention to the energy efciency and connectivity. In most studies, one of the objectives is rst optimized based on which the other one is optimized. Very less number of studies have considered the optimization problem with a full connectivity of the network. This paper attempts to optimize the coverage and lifetime of a wireless sensor network simultaneously using MOPSO algorithm [11] while maintaining a full connectivity of the network. During the optimization process, sensor nodes dynamically organize themselves to form a fully connected network. The preliminary study and the results [12] indicated a wide scope for MOEAs in the eld of sensor deployment. 3. Problem formulation 3.1. WSN modeling It is assumed that each node knows its position in the search space and all sensor nodes are homogeneous and have equal energy and mobility. HECN is assumed to be more powerful than the rest of the sensor nodes. In this paper, a at square surface is considered in which HECN is placed at the center for convenience. The sensing area of each node is assumed to have a circular shape with radius Rsens. It is assumed that for every data transmission, the sensor node loses one unit of energy. The communication range of each node is also dened by the area of a circle with radius Rcomm. Therefore the energy spent by a sensor node is directly related to the distance between the node and its neighboring nodes in a radius of Rcomm. The energy consumption in a sensor node is also affected by the signal strength, interference, retransmission of lost data etc. However these external factors need modeling and detailed analysis in different type of wireless channels and therefore has been kept out of scope of this work. The co-ordinates of the sensor nodes (x1, y1), (x2, y2), . . . are considered as design variables which are to be optimized. 3.2. Calculation of objectives The following two objectives are considered.  Maximize the total coverage of the sensor network: f1  Maximize the lifetime of the sensor network: f2

1136

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

Detailed description of objective functions f1 and f2 are presented below. 3.2.1. Coverage It is one of the measurements of quality of service (QoS). The coverage of each sensor node in terms of probability of detection (Pd) can be dened either by a binary sensor model [6] or a stochastic sensor model [13]. In a binary sensor model, the detection probability of the event of interest is one within the sensing range; otherwise, the probability is zero. Coverage of a network is determined by nding the union of sensing areas dened by the location of each sensor node and Rsens. Although the binary sensor model is simpler, the uncertainty factor in the measurement is not taken into consideration. The binary sensor model is given by

& C ij x; y

1 for dij x; y 6 Rsens 0 for dij x; y > Rsens

The sensor eld is represented by an m n grid. dij(x, y) denotes the Euclidean distance between a sensor node at (x, y) and any grid point at (i, j). Eq. (1) expresses the coverage Cij(x, y) of a grid point (i, j) by a sensor node at (x, y). The coverage for the entire grid is calculated as the fraction of grid points covered. In reality, measurements by a sensor node are imprecise; hence the coverage needs to be expressed in probabilistic terms. In a stochastic sensor model, the probability of detection follows an exponential decaying function of distance from the sensor node. The stochastic sensor model given in (2) is motivated in part by [14],

C ij x; y

8 >1 > < > > : e 0

for dij x; y 6 Rsens Re for Rsens Re < dij x; y < Rsens for dij x; y P Rsens Re 2

kab

where Tfailure is the maximum number of sensing cycles before the failure of any node and Tmax is the maximum number of possible sensing cycles. In every sensing cycle, the data from every node is routed to HECN through a route of minimum weight. To nd this route, the outgoing edges of every node are weighted by the inverse of nodes remaining energy and then Dijkstra algorithm [15] is used to nd out the route with minimum weight. This calculation is repeated for subsequent sensing cycles until energy of at least one node is depleted. This gives the maximum number of sensing cycles before any node fails. The maximum number of possible sensing cycles can be obtained by taking the ratio of total energy of any node and energy lost for one data transmission. This corresponds to a layout when all sensor nodes are directly connected to HECN. In most applications, a large number of sensors are deployed in an area to compensate for different type of losses and damages caused to the sensors due to the terrain, local and environmental factors. The deployment of adequate number of sensors reduces the amount of movement the sensors need to carry out while moving from initial to nal positions. Therefore in this study, the energy spent in the movement has been neglected in comparison to the energy spent on communication. The presence of obstacles in the ROI also affects the connectivity among the sensors and hence the sensors spend more energy for communication resulting in a smaller lifetime. The obstacles in a sensor network can be stationary as well as non-stationary which needs modeling and detailed analysis in different wireless environments and are therefore out of scope of the present work. These two objectives are competing with each other. The coverage objective will try to spread out the nodes for maximizing coverage while resulting in high energy loss and small lifetime. The lifetime objective will try to arrange the nodes as close as possible to the HECN for reducing loss in energy which results in poor coverage.

Re(Re < Rsens) is a measure of the uncertainty in measurement, a = dij(x, y) (Rsens Re), and k and b are parameters that measure the detection probability when there is an uncertainty in sensor node detection. The coverage for the entire sensor eld is calculated as the fraction of grid points that exceeds the threshold Cth. In general, the coverage can be calculated by the following expression:

4. Proposed approach In a multiobjective optimization problem, a set of tradeoffs exists giving rise to multiple solutions. Each solution in this set represents a particular performance trade-off between the multiple objectives and hence considered as an optimal solution. 4.1. Multiobjective optimization problem Assuming a minimization problem for convenience, it can be stated as minimization of a vector function z consisting of M objectives with respect to a vector variable x = (x1, . . . , xn) in a set U, i.e.,

Max Cov eragef1

i1;...;N Ai

where Ai is the area covered by the ith node, N is the total number of nodes and A is the area of the ROI. 3.2.2. Lifetime The second objective considered is maximization of lifetime. Lifetime is dened as the time until one of the participating nodes run out of energy. This objective can be calculated by the subsequent expression:

min zx z1 x; z2 x; z3 x; . . . ; zM x

Max Lifetimef2

T failure T max

A solution to a multiobjective problem (i.e. u) dominates another solution (i.e. v) if u performs at least as good as v with respect to all the objectives and performs better than v in at least one objective.

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

1137

4.2. Pareto dominance Given two solutions u and v from a set U, vector z(u) is said to dominate vector z(v) (denoted by z(u) 0 z(v)) if and only if,

zi u 6 zi v ; zi u < zi v ;

8i 2 f1; . . . ; Mg
9i 2 f1; . . . ; Mg

6 7

4.3. Pareto optimality The solutions which are not dominated by any other solution are declared as non-dominated or Pareto optimal solutions. There are no superior solutions to the problem than u, although there may be other equally good solutions. The solution u 2 U is Pareto optimal if and only if,

zv 0 zu;

:9v 2 U

The set of solutions that satisfy (8) is known as the Pareto optimal set and the tness values corresponding to these solutions form the Pareto front or trade-off surface in objective space. The concept of Pareto optimum was originally introduced by Edgeworth [16] and later generalized by Pareto [17]. Goldberg has incorporated the concept of Pareto optimality into an evolutionary algorithm [18] and suggested that the use of non-dominated ranking and selection moves the individuals in a population towards the Pareto front. These ideas of Goldberg have helped the researchers in developing new MOEAs. Although several MOEAs exist in literature, the two popularly used MOEAs i.e. non-dominated sorting genetic algorithm II (NSGA-II) [19] and MOPSO [11] algorithms have been used in this study. These two algorithms have found extensive applications in different elds of engineering. The literature also reveals that these two algorithms have catered the most to the needs of practical optimization problems known to date. These algorithms are also popular for easier hardware implementation. NSGA-II is a popular genetic algorithm (GA) for multiobjective optimization. In NSGA-II, parent population of size N is created which subsequently undergoes selection, crossover and mutation processes to produce an offspring population of size N. The offspring population is combined with the parent population to form a combined population of size 2N which undergoes non-dominated sorting process. This classies the complete population into several non-dominated fronts based on the values of the objective functions. The members of the rst front are completely non-dominant. The members in the second front are dominated by the members of the rst front only. Similarly the other fronts are determined until each member of the population falls into one front. The new population of size N is created by taking the members of the non-dominated fronts starting from the rst level. Since the population size is predened, the combined population cannot be completely accommodated in the new population. Thus, several non-dominated fronts are discarded. If all members of a front can not be accommodated, then the required number of members for the new population are selected

based on the crowding distance technique. The operators such as binary tournament selection, simulated binary crossover, and polynomial mutation are introduced into MOPSO to improve the overall performance. Similar to GA, the PSO algorithm has been successfully extended to multiobjective optimization problems. The large popularity of MOPSO is attributed to its simple implementation, population based approach and success in handling continuous search spaces. Coello et al. [11] have shown that the performance of MOPSO is highly competitive with respect to NSGA-II. The authors have also stated that MOPSO is relatively easy to implement than NSGA-II. Additionally, MOPSO needs exceptionally low computational time which makes it a very promising candidate for engineering optimization problems where the computational cost is a vital issue. These advantages of MOPSO over NSGA-II has encouraged the formers usage in sensor deployment optimization problem. In order to construct a direct relationship between the problem domain and the PSO, each particle represents a set of coordinates corresponding to N number of nodes. So each particle represents a network layout. The proposed MOPSO algorithm is composed of the following steps: 1. Randomly initialize the positions for all the particles i.e. position of ith particle in the 2N-dimensional space as Xi = (Xi1, Xi2, . . . , Xid, . . . , Xi2N) where d(1 6 d 6 2N) represents the dimension. If mapped into the application domain, Xi represents the ith random topology for the sensor network dened by the coordinates of N sensor nodes. Therefore the X- and Y-coordinates of the N sensor nodes are represented as 2N different dimensions of the ith particle. For ith particle Xi, the dimensions (Xi1, Xi2) and (Xi(2N 1), Xi2N) represent the X- and Ycoordinates of the 1st and Nth node respectively. 2. The initial random topology generated in Step 1 may not be a fully connected network which is the basic requirement for evaluating the two tness functions. Therefore the following operations are carried out for each particle until it forms a fully connected network. (a) Add random numbers in the interval [1, 1] to the position of the particle generated in Step 1. This step is carried out to change the position of sensor nodes by small margin to form a possible fully connected network. (b) Constrain the position so that it does not exceed the bound given by the ROI. (c) Check whether the sensor nodes in the particle form a fully connected network. If yes, then go to Step 3. Otherwise repeat Step 2. 3. Initialize the velocity of each particle i.e velocity of ith particle in the 2N-dimensional space as Vi = (Vi1, Vi2, . . . , Vi2N). 4. Evaluate the two objective functions for each particle using (3) and (4). 5. Store the particles representing non-dominated solutions in the elite archive as dened by (7) and (8). 6. Initialize the personal best position of each particle to the position generated in Step 2 i.e. X p X p ; X p ; . . . ; X p . i i1 i2 i2N

1138

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

7. WHILE the maximal number of iterations has not yet been reached. For each particle DO (a) Compute the velocity using (9).

5. Performance metrics and fuzzy decision making With the existence of different MOEAs, it is necessary to quantify the performance of these algorithms. There are two goals in multiobjective optimization: 1. discover solutions as close as possible to the Pareto optimal front 2. discover solutions as diverse as possible in the nondominated front The rst goal guides the solutions towards the Pareto optimal region and the second goal guides along the Pareto optimal front. Various performance metrics to measure these goals have been introduced in the literature [21,22,19]. Some of the metrics require the knowledge of the true Pareto-optimal front which is unknown in our application. Taking this limitation into account, the following three metrics are considered: 1. size of the dominated space, 2. coverage of the Pareto front, 3. non-uniformity of the Pareto front. The rst two metrics measure the convergence of the Pareto front, while the last metric measures the distribution of solutions along the Pareto front. 5.1. Size of the dominated space (D) This metric has been introduced by Zitzler in [23] and was improved in [24]. D is a measure of the amount of objective space weakly dominated by a given non-dominated set. Since sensor deployment involves the maximization of two objectives, zero is chosen as the minimum value for the two objectives to determine the size of the dominated space. Higher value of D indicates better performance. 5.2. Set coverage metric (C) This metric, as suggested by Zitzler [24], measures the relative spread of solutions between two non-dominated sets. The set coverage metric C(A, B) calculates the proportion of solutions in B which are weakly dominated by solutions of A where A, B are two Pareto-optimal sets. C(A, B) = 1 indicates that all the members of B are weakly dominated by A. C(A, B) = 0 indicates that no member of B is weakly dominated by A.

V id w V id r 1 X p X id r2 X g X id id d

where w is the inertia weight and r1,r2 are random numbers uniformly distributed in the range [0, 1]. The global best position of the particle swarm is rep resented as X g X g ; X g ; . . . ; X g . The global best 1 2 2N particle is selected in the following way: the hypercubes containing more than one particle are assigned a tness equal to the ratio of 10 and number of particles in the corresponding hypercube. Roulette-wheel selection is used to select the hypercube and then, a particle is selected randomly within that hypercube. (b) Compute the new position using (10).

X id X id V id

10

(c) Constrain the position so that it does not exceed the bound given by the ROI. (d) The newly generated topology may not be a fully connected network. Therefore similar to Step 2, the following operations are carried out until the ith particle forms a fully connected network. i. Add random numbers in the interval [1, 1] to the position of the particle generated in Step 1. This step is carried out to change the position of sensor nodes by small margin to form a possible fully connected network. ii. Constrain the position so that it does not exceed the bound given by the ROI. iii. Check whether the sensor nodes in the particle form a fully connected network. If yes, then go to Step 7(e). Otherwise repeat Step 7(d). (e) Evaluate the two objective functions for the ith particle using (3) and (4). (f) Compare the tnesses of the particle with the tnesses of the archive members. (g) Update the elite archive by inserting all the currently non-dominated positions and eliminate any dominated locations from the archive. If the archive reaches its maximum capacity, then the adaptive grid procedure [20] is incorporated. In an adaptive grid scheme, the objective function space is divided into grids where the members of archive are located. If the new solution entering into archive lies beyond the boundaries of the adaptive grid, then the size of the adaptive grid is changed so that the new solution can be located in the new grid. Since the size of the archive is kept xed, one of the solutions from high density region of the grid is removed and the new solution enters into archive as well as located in the adaptive grid. (h) If the new position Xi dominates the personal best X p , i then update the personal best i.e. X p X i . i (i) Increase the loop counter. 8. Return the archive as the non-dominated solution set.

CA; B

jfb 2 Bj9a 2 A : a1bgj jBj

11

where jBj represent the number of solutions in the set B, and a 1 b means that solution a weakly dominates solution b. Therefore, C(A, B) gives the fraction of B dominated by A. Note that C(A, B) is not necessarily equal to 1 C(B, A). If C(A, B) > C(B, A), this indicates that the set A has better solutions than the set B.

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

1139

For comparing more than two Pareto fronts, the metric C can be modied as follows: Consider a set U consisting of M non-dominated sets i.e. U = A1,A2, . . . , AM.

M lj lj PN i1M i P j1

i1

lji

15

CAj

P
i1;...;M;

jfb 2 Ai j9a 2 Aj : a1bgj jAj j ij

12

where jAjj represent the number of solutions in the set Aj. 5.3. Non-uniformity of the Pareto front (S) This metric is proposed by Schott [25].

where M is the total number of objectives. The solution with maximum value of lj is a compromised solution that can be accepted by the decision maker. However this solution is not mandatory for any decision maker. With the change in ROI and physical environment, the decision makers can choose a solution on the Pareto front based on their previous experience and knowledge. 6. Simulation results and analysis

s  2 1 XN di 1 S i1  N1 d

13

where N is the number of solutions in the non-dominated  i  P k set and di mink M fm fm ; k 1; . . . ; N and i k. M is m1  the total number of objectives to be optimized. d is the mean of all di. This metric measures the standard deviations of different di values and provides information about the spread of the non-dominated solutions on the Pareto front. A smaller value of S indicates that the solutions are closer to uniformly spread solutions. Apart from these metrics, the computation time is also measured to show the efciency of our proposed approach. 5.4. Fuzzy decision making The MOEAs provide a set of non-dominated solutions which form the Pareto front. It is assumed that there is fuzziness in each objective due to the imprecision in the nature of judgement of a decision maker. This fuzziness can be dened by membership functions representing the degree of fuzziness. When the number of solutions at the center of the Pareto front is very large and it is difcult to distinguish between the solutions providing almost equal weightage to each objective, then the fuzzy mechanism can be used to nd a compromised solution. The fuzzy mechanism looks at the way the solutions are contributing to each objective and assigns a fuzzy variable. The fuzzy mechanism shows a possible way of nding a compromised solution in case solutions are very close to each other. In this paper, the best compromised solution on the Pareto front is selected by a fuzzy based mechanism [26]. Membership value for ith objective of jth solution in the Pareto front is calculated using the membership function as

j i

8 > 1; > < max


Fi

if F i 6 F min i
F i

> F max F min i > i : 0;

; if F min < F i 6 F max i i if F i > F max i

14

lji indicates how well the jth non-dominated solution is able to satisfy the ith objective. The sum of membership values for all objectives of the jth non-dominated solution suggests how well it satises all the objectives. The achievement of each non-dominated solution with respect to all the N non-dominated solutions can be dened as

The deployment of sensors is a highly non-linear problem. A small change in the position of a sensor can lead to a disconnected network. The MOPSO provides a set of Pareto-optimal layouts from which the end-user can choose depending on the application. The simulation study is carried out in MATLAB environment on a Windows XP based computer with the following specications: 2 GHz Core2Duo CPU, 2 GB RAM. To assess the performance of our proposed approach, the results are compared with the results of NSGA-II. The MOPSO algorithm starts with a swarm of particles randomly generated where each particle represents a network layout represented by sensor co-ordinates. The coverage and lifetime of the particles are then calculated. The archive containing non-dominated solutions is developed according to the concept of Pareto optimal dominance developed by Coello and Lechuge [11]. The velocity of particles are updated using the members of the archive. The simulation parameters used for NSGA-II are as follows: real coded GA, simulated binary crossover with distribution index 20, polynomial mutation with distribution index 20, binary tournament selection based on crowded-comparison operator with pool size as 50 and tour size is 2. The simulation parameters used for MOPSO are as follows: the size of archive is 10, number of grids for objective function space is 20, the inertia weight is 0.4, both the acceleration constants are taken as 2 and the random numbers are chosen in the range [0, 1]. The population size and number of iterations for both the algorithms have been taken as 50 and 10 respectively. The parameters of the sensor network are as follows: ROI = 10 10, number of nodes = 10, Rsens = 2, Rcomm = 2, Re = 1. In this section, the binary sensor model is used to calculate the coverage of the sensor network. Thirty independent runs with different random seeds are performed. NSGA-II nds about 15 Pareto-optimal solutions out of 50 individuals. In contrast, MOPSO leads to a Pareto optimal set with 10 solutions. The quality of the Pareto-optimal solutions obtained with the two algorithms is measured by three qualitative metrics D, C and S. The best average result with respect to each metric is shown in bold font in all the tables. The best, worst, mean, median and standard deviation of the three performance metrics are listed in Table 1. In Table 1, MOPSO leads to a higher value of D, indicating that better solutions are obtained with MOPSO than with NSGA-II. The difference between the means of MOPSO

1140

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

Table 1 Results of different performance metrics for MOPSO and NSGA-II. MOPSO D Best Worst Average Median Std. Dev. C Best Worst Average Median Std. Dev. S Best Worst Average Median Std. Dev. Comp. time (s) Best Worst Average Median Std. Dev. 0.4907 0.3738 0.4496 0.4782 0.053 1 0.8 0.9289 0.9 0.0743 0.2182 0.3876 0.3113 0.3253 0.0687 704.6719 842.5156 746.4992 747.5547 33.9693 NSGA-II

1 0.9 0.8
MOPSO NSGAII

Lifetime

0.4211 0.3367 0.3776 0.3813 0.0279 0.2857 0 0.181 0 0.1283 0.3915 0.6891 0.5169 0.4997 0.1144 617.8906 1763.7969 1404.2625 1522.0469 454.2114

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1

Coverage
Fig. 1. Pareto-optimal solutions obtained with MOPSO and NSGAII.

and NSGA-II with respect to D is as big as the standard deviation of MOPSO. The value C = 0.9289 indicates that 92.89% of the Pareto-optimal solutions obtained with NSGA-II are weakly dominated by the solutions obtained with MOPSO. Similarly, the value C = 0.181 represents that only 18.1% of the solutions obtained with MOPSO are weakly dominated by those with NSGA-II. Furthermore, the standard deviation of MOPSO with respect to C is signicantly smaller than that of NSGA-II, indicating that the performance of MOPSO is statistically more stable. The distributions of the solutions obtained with MOPSO and NSGA-II are evaluated with non-uniformity metric S. A lower value of S indicates a better or uniform spread of solutions. The comparison of S shows that the spread of the solutions obtained with MOPSO is more uniform than NSGA-II. Table 1 also reveals that the computation time for deployment of sensors using MOPSO is approximately half of that of NSGA-II. All the performance metric measurements show that the MOPSO clearly outperforms NSGA-II. More specically, metrics D and C show that MOPSO leads to a better convergence of the Pareto front, and S shows that MOPSO leads to a more uniform distribution of the Pareto front. To illustrate the difference between the Pareto fronts obtained with MOPSO and NSGA-II, the Pareto fronts obtained with one of the 30 runs of MOPSO and NSGA-II are plotted in Fig. 1. The fuzzy logic based mechanism dened by (14) and (15) is used for nding the best compromised solution on each Pareto front. The best compromised solutions among those obtained by the MOPSO algorithm in one of the runs is marked with square bracket in Fig. 1. The end user can use this solution as an initial point for an optimal deployment of sensors. The initial random deployment results in a disconnected sensor network. This disconnected network is converted to

a fully connected network as discussed in Section 4. Fig. 2 shows one of the initial disconnected network layout. Although the sensor nodes are not uniformly spread, still the coverage of this disconnected network is competitive with a value of 0.6371. But the disconnected network can not be accepted as a possible topology for the wireless sensor network. Some out of thirty independent experiments might provide better coverage and connectivity, but it is less likely that the initial layout will be a fully connected network. Therefore the need for an optimized layout exists which paves a way for MOPSO algorithm implementation to this application. Pareto optimal layouts corresponding to the two boundary solutions of the Pareto front obtained in one of the experiments are shown in Figs. 3 and 4. The layout shown in Fig. 3 is the layout with best coverage of 0.6347 obtained by MOPSO algorithm. This value of coverage is less than that of the initial layout, but the network is a fully connected network with minimum lifetime of 0.199. In order to get more coverage, the particles spread out to minimize the overlapping region. This leads to a network where the sensors are far away from HECN. Many sensors transmit their own data as well as act as communication relay for other far away sensors. Sensors

6 4 2 0 2 4 6 6 4 2 0 2 4 6

Fig. 2. Initial disconnected WSN layout with a coverage of 0.6371.

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

1141

6 4 2 0 2 4 6 6 4 2 0 2 4 6

6 4 2 0 2 4 6 6 4 2 0 2 4 6

Fig. 3. Pareto-optimal fully connected WSN layout with lifetime of 0.199 and best coverage of 0.6347.

Fig. 5. Example of another Pareto-optimal fully connected WSN layout with coverage of 0.5617 and lifetime of 0.499.

0.55
MOPSO NSGAII

6 4 2 0 2 4 6 6 4 2 0 2 4 6

0.5

Dominated Space (D)

0.45 0.4 0.35 0.3 0.25

10

20

30

40

50

60

70

80

90

100

Generation Number
Fig. 6. Evolution of dominated space metric forMOPSO and NSGAII.

Fig. 4. Pareto-optimal fully connected WSN layout with coverage of 0.3335 and best lifetime of 0.999.

acting frequently as communication relay lose more energy. Therefore the lifetime of this layout indicates the minimum lifetime that the sensor network can provide for the corresponding application. Fig. 4 shows the layout with best lifetime of 0.999. It can be noted that every sensor node is directly connected to the HECN for maximizing their corresponding lifetime. Therefore no sensor acts as a communication relay in this layout. The sensors try to remain close to the HECN thereby reducing the coverage of the network to the minimum value of 0.3335. Since the sensor deployment problem is a multiobjective problem, one of the solutions is shown in Fig. 5 which is a Pareto optimal layout with good coverage of 0.5617 as well as a good lifetime of 0.499. In addition to the performance at the end of the evolution, the performance of MOPSO and NSGA-II during the evolution is also monitored. Figs. 68 shows the evolution of metrics D, C and S respectively with respect to number of generations. MOPSO shows a faster convergence and a more stable performance than NSGA-II. These results indicate that MOPSO nds better solutions faster and more reliably.

1 0.9

Set Coverage Metric (C)

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 10 20 30 40 50 60 70 80

MOPSO NSGAII

90

100

Generation Number
Fig. 7. Evolution of set coverage metric for MOPSO and NSGA-II.

7. Sensitivity analysis In this section, the impact of different parameters on the performance of MOPSO algorithm is studied. Four

1142

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

0.7 0.65 0.6


MOPSO NSGAII

0.55 0.5 0.45 0.4 0.35 0 20 40 60 80 100

Generation Number
Fig. 8. Evolution of non-uniformity of the Pareto front metric for MOPSO and NSGA-II.

coverage probability greater than the threshold is fewer. Practically measurement by a sensor node involves some uncertainty which is not considered in a binary sensor model. Although stochastic sensor model gives less coverage, it is physically realizable. Table 2 shows that the binary sensor model provides a higher value of D than the stochastic sensor model. 78.6% of the non-dominated solutions obtained with the stochastic sensor model are weakly dominated by the solutions obtained with the binary sensor model. Similarly, the value C = 0 indicates that no member of the non-dominated set obtained with the binary sensor model are weakly dominated. The lower value of S shows that the distribution of solutions along the Pareto front obtained with the binary sensor model is more uniform than that obtained with the stochastic sensor model. The computation time for obtaining the Pareto optimal solutions using the binary sensor model is less than that of the stochastic sensor model which well agrees with the discussion in Section 3. 7.2. Experiment 2: Effect of variation in population size The number of particles are varied for studying the impact of population size on the performance of the proposed approach. The variation in performance is measured with respect to different metrics discussed in Section 5. Different runs using 25, 50, 75, 100 and 125 particles are carried out. Table 3 shows that the use of 50 particles provided the best average result with respect to dominated space and spacing and the use of 100 cats provided the best average result with respect to set coverage metric. However the average result obtained with the use of 50 particles are competitive in terms of set coverage metric. The small va-

experiments are carried out to demonstrate the sensitivity of different parameters. 7.1. Experiment 1: Effect of sensor model In order to demonstrate the effect of different sensor models on the coverage of the sensor network, two different experiments are carried out with the two models explained in Section 3. The parameters of stochastic sensor model are as follows: k = 0.5, b = 0.5, Cth = 0.7. It is interesting to look at the Pareto fronts obtained using two different sensor models as shown in Fig. 9. The upper bound for coverage in case of a stochastic sensor model is lower than the upper bound in case of a binary sensor model. This is due to the fact that coverage for the binary sensor model is the fraction of the grid points covered by the circles. For the stochastic sensor model, even though there are a large number of grid points that are covered, the overall number of grid points with

Spacing (S)

Table 2 Results of experiment 1. Stochastic model D Best Worst Average Median Std. Dev. C Best Worst Average Median Std. Dev. S Best Worst Average Median Std. Dev. Comp. time (s) Best Worst Average Median Std. Dev. 0.3627 0.293 0.329 0.3334 0.0273 0 0 0 0 0 0.2812 1.0396 0.5835 0.5471 0.2215 895.6094 1049.6875 964.9945 963.1953 39.9839 Binary model 0.4907 0.3738 0.4496 0.4782 0.053 1 0.5714 0.786 0.7889 0.1545 0.2182 0.3876 0.3113 0.3253 0.0687 704.6719 842.5156 746.4992 747.5547 33.9693

1
Binary Sensor Model

0.9 0.8 0.7

Stochastic Sensor Model

Lifetime

0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1

Coverage
Fig. 9. Pareto optimal solutions obtained with different sensor models.

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145 Table 3 Results of experiment 2. Particles 25 50 0.4907 0.3738 0.4496 0.4482 0.053 1 0.8 0.9289 0.9 0.0743 0.2182 0.3876 0.3113 0.3253 0.0687 75 0.4916 0.4125 0.4464 0.4497 0.0355 1 0 0.9028 0.9107 0.1377 0.2933 0.3868 0.3283 0.312 0.0257 100 0.4871 0.4127 0.4438 0.4488 0.034 1 0.5492 0.9463 0.9413 0.1276 0.2503 0.3679 0.3174 0.3133 0.0289 125

1143

Keeping this in mind, a value of 10 is adopted for our application. 7.4. Experiment 4: effect of variation in archive grid size
0.5128 0.4263 0.4422 0.4447 0.0335 1 0 0.9381 1 0.1508 0.2989 0.4138 0.3329 0.3321 0.0867

Dominated space Best 0.4144 Worst 0.3574 Average 0.3823 Median 0.3877 Std. Dev. 0.0227 Set coverage Best Worst Average Median Std. Dev. Spacing Best Worst Average Median Std. Dev. metric 0.779 0.3 0.6453 0.5264 0.1557 0.2684 0.3998 0.3962 0.3537 0.0217

In this experiment the grid size is varied while all other parameters are kept constant. Different runs are carried out using 10, 20, 30 and 40 divisions for the adaptive grid. Table 5 shows that a value of 20 divisions provided best average results for all three performance metrics. The smaller values of standard deviation indicates the higher stability in the performance. Hence an adaptive grid with 20 divisions is recommended for our optimization problem. 7.5. Experiment 5: effect of variation in number of sensor nodes In this experiment the number of sensor nodes is varied while all other parameters are kept constant. Different runs are carried out using 5, 10, 15, 20 and 25 number of sensor nodes. Table 6 shows that with increase in number of sensor nodes, the performance of the algorithm improves. However the trade off between the extra cost due to sensor nodes and the performance has to be considered. Above a particular number of sensor nodes, the extra addition will provide minor improvement in performance while adding huge cost to the sensor deployment process. 7.6. Experiment 6: effect of variation in size of ROI In this experiment the ROI is varied while all other parameters are kept constant. Different runs are carried out using 5 5, 10 10, 15 15, 20 20 and 25 25 as the ROI. Table 7 shows that the computational time requirement for optimization increases with increase in the size of the ROI which is also theoretically true. However the increase in computational time is noteworthy when the size of ROI

lue of the standard deviation with respect to set coverage metric indicates that the performance obtained using 50 particles is more stable. Hence the use of 50 particles is a reasonable choice for our optimization problem. 7.3. Experiment 3: effect of variation in archive size In this experiment the size of the archive is varied while all other parameters are kept constant. Different runs with archive size of 5, 10, 15 and 20 are carried out. Table 4 shows that an archive size of 10 provided the best average results for all three performance metrics and signicantly smaller standard deviation with respect to set coverage metric and spacing. It is interesting to notice that while moving from an archive size of 10 to 20, the performance with respect to spacing degrades by a signicant amount of 18%. The search effort and complexity, involved in the selection of a new member of archive, increases with increase in the size of the archive.

Table 4 Results of experiment 3. Archive size Dominated space Best Worst Average Median Std. Dev. 5 0.4862 0.3619 0.401 0.3888 0.039 10 0.4907 0.3738 0.4496 0.4482 0.053 1 0.8 0.9289 0.9 0.0743 0.2182 0.3876 0.3113 0.3253 0.0687 15 0.496 0.3901 0.424 0.4165 0.0343 0.9472 0.3 0.8597 0.8458 0.1858 0.2048 0.7616 0.3817 0.3638 0.1574 20 0.4851 0.3711 0.4295 0.4145 0.0445 0.9227 0.2679 0.8736 0.852 0.1878 0.3199 0.6602 0.3588 0.3269 0.1338

Table 5 Results of experiment 4. Grid divisions Dominated space Best Worst Average Median Std. Dev. 10 0.4838 0.392 0.4306 0.4249 0.0344 20 0.4907 0.3738 0.4496 0.4482 0.033 1 0.8 0.9289 0.9 0.0743 0.2182 0.3876 0.3113 0.3253 0.0687 30 0.4843 0.357 0.4314 0.4288 0.0447 0.9778 0.2198 0.8738 0.7958 0.2759 0.2353 0.6292 0.3268 0.3229 0.1399 40 0.49 0.367 0.4288 0.4131 0.0425 0.9381 0.2018 0.8715 0.8292 0.2583 0.29 0.6763 0.3733 0.3786 0.1426

Set coverage metric Best 0.979 Worst 0.1 Average 0.8225 Median 0.8556 Std. Dev. 0.1788 Spacing Best Worst Average Median Std. Dev. 0.2099 0.9319 0.5688 0.5138 0.2467

Set coverage metric Best 0.9583 Worst 0.1454 Average 0.8646 Median 0.8548 Std. Dev. 0.2509 Spacing Best Worst Average Median Std. Dev. 0.3623 0.6266 0.46 0.4371 0.0906

1144 Table 6 Results of experiment 5. Sensors 5 10 0.4907 0.3738 0.4496 0.4482 0.033 1 0.8 0.9289 0.9 0.0743 0.2182 0.3876 0.3113 0.3253 0.0687 15

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145

20 0.4954 0.3930 0.4599 0.4504 0.0384 1 0.8192 0.9497 0.9152 0.0747 0.2076 0.3478 0.3103 0.3226 0.0831

25 0.4982 0.3945 0.4599 0.4567 0.0365 1 0.8387 0.9522 0.9273 0.0871 0.2006 0.3414 0.3100 0.3210 0.0998

Dominated space Best 0.4006 Worst 0.3822 Average 0.3919 Median 0.3912 Std. Dev. 0.0358 Set coverage Best Worst Average Median Std. Dev. Spacing Best Worst Average Median Std. Dev. metric 0.5583 0.2739 0.5558 0.5137 0.0766 0 0.4449 0.3559 0.3563 0.0363

0.4945 0.3833 0.4498 0.4399 0.0312 1 0.8118 0.9458 0.9123 0.0627 0.2169 0.3498 0.3112 0.3240 0.0418

Table 7 Results of experiment 6. ROI Computation Best Worst Average Median Std. Dev. 10 10 time (s) 704.6719 842.5156 746.4992 747.5547 33.9693 30 30 782.9861 864.0747 806.3678 792.3514 29.4260 50 50 908.6648 979.3412 944.8631 951.2178 22.7599 100 100 1592.0588 1676.4858 1632.6482 1637.3845 28.8860

choose from, depending on the application. The performance of MOPSO and NSGA-II is compared to identify the contribution of the proposed approach. Three different performance metrics were used such as size of the dominated space, set coverage metric and non-uniformity of the Pareto front. The performance difference between MOPSO ans NSGA-II also appeared in the variation of the metrics with respect to generation number. The proposed approach leads to a better Pareto front more efciently and reliably. The simulation results also indicated that the binary sensor model gives better coverage than stochastic sensor model but it is difcult to realize practically. A fuzzy based mechanism is also used to nd out the best compromised solution on the optimal Pareto front. In practice, a WSN is divided into multiple sub-regions for easy layout, organization and management. Since stochastic sensor model is accepted practically, the size of sub-region and their corresponding density and edge effects should be considered. In future, the energy consumption due to sensor movement in presence of obstacles will be taken into account while calculating lifetime. In addition the analysis of energy consumption model considering signal strength, retransmission and interference will also be undertaken in the future work. Other objectives, such as time and distance for sensor movement and uniformity of the network will be considered for optimization. The energy consumption due to sensor movement in presence of obstacles will also be taken up in future work. Role of the funding source One of the authors, P.M. Pradhan, acknowledge the generous funding received from Ministry of Human Resource Development, Government of India for carrying out this work. References
[1] Y. Zou, K. Chakrabarty, Sensor deployment and target localization based on virtual forces, in: Proceedings of IEEE INFOCOM Conference, vol. 2, pp. 12931303. [2] N. Heo, P.K. Varshney, Energyefcient deployment of intelligent mobile sensor network, IEEE Transactions on Systems, Man, and CyberneticsPart A: Systems and Humans 35 (2005) 7892. [3] S.S. Dhillon, K. Charabarty, Sensor placement for effective coverage and surveillance in distributed sensor networks, in: Proceedings of IEEE Wireless Communication Network Conference, pp. 16091614. [4] X. Wu, S. Lei, J. Yang, H. Xu, J. Cho, S. Lee, Swarm based sensor deployment optimization in ad hoc sensor networks, in: ICESS05, pp. 533541. [5] X. Wang, S. Wang, J. Ma, Dynamic deployment optimization in wireless sensor networks, in: Intelligent Control and Automation, Lecture Notes in Control and Information Sciences, vol. 344, Springer, Berlin, Heidelberg, 2006, pp. 182187. [6] S. Slijepcevic, M. Potkonjak, Power efcient organization of wireless sensor networks, in: Proceedings of Fifth International Conference on Commun., vol. 2, pp. 472476. [7] Z. Li, L. Lei, Sensor node deployment in wireless sensor networks based on improved particle swarm optimization, in: Applied Superconductivity and Electromagnetic Devices, 2009, ASEMD 2009. International Conference on, pp. 215217. [8] X. Wang, S. Wang, J.-J. Ma, An improved co-evolutionary particle swarm optimization for wireless sensor networks with dynamic deployment, Sensors 7 (2007). [9] W.-H. Liao, Y. Kao, Y.-S. Li, A sensor deployment approach using glowworm swarm optimization algorithm in wireless sensor networks, Expert Systems with Applications 38 (2011) 12180 12188.

changes from 50 50 to 100 100. This is due to the large change in the size of ROI while the number of sensors remains the same. Due to large search space, MOPSO takes large time to converge to the Pareto front. Therefore depending on the size of the ROI, the total number of nodes should be changed in order to get the best performance. 8. Conclusion and future work Several papers have been reported in the literature showing the importance of wireless sensor network and the potential applications that are emerging with the development of this new technology. An efcient deployment technique for wireless sensor nodes has been a challenging problem for the network designers. In this paper, an efcient deployment technique with good coverage and lifetime was considered. The application of multiobjective particle swarm optimization to maximize coverage and lifetime simultaneously was discussed. Thus, the aim of the proposed algorithm is to locate good non-dominated solutions. To optimize multiple objectives simultaneously, the proposed algorithm maintains an elite archive and uses the archive members to dynamically lead the particle swarm in searching for more and better nondominated solutions. While applying the algorithm, the connectivity of the network was considered as a constraint. The mobility of the nodes provides a way to avoid time consuming, expensive deployment techniques. The end user is provided with a set of Pareto-optimal solutions to

P.M. Pradhan, G. Panda / Ad Hoc Networks 10 (2012) 11341145 [10] C. Ozturk, D. Karaboga, B. Gorkemli, Articial bee colony algorithm for dynamic deployment of wireless sensor network, Turkish Journal of Electrical Engineering and Computer Sciences 20 (2012). [11] C.A. Coello, G.T. Pulido, M.S. Lechuge, Handling multiple objective with particle swarm optimization, IEEE Transactions on Evolutionary Computation 8 (2004) 256279. [12] P.M. Pradhan, V. Baghel, M. Bernard, G. Panda, Energy efcient layout for a wireless sensor network using multiobjective particle swarm optimization, in: Proceedings of IEEE International Advance Computing Conference (IACC 2009), Patiala, India, pp. 6570. [13] Y. Zou, K. Chakrabarty, Sensor deployment and target localization based on virtual forces, in: Proceedings of IEEE INFOCOM Conf., vol. 2, pp. 12931303. [14] A. Elfes, Sonar-based real-world mapping and navigation, IEEE Journal of Robotics and Automation RA-3 (1987) 23492365. [15] T.H. Cormen, et al., Introduction to algorithms, MIT Press, Cambridge, MA, 2001. [16] F.Y. Edgeworth, Mathematical Physics, P. Keagan, London, England, 1881. [17] V. Pareto, Cours DEconomie Politique, P. Keagan, London, England, 1896. [18] D.E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning, AddisonWesley Publishing Company, Reading, Massachusetts, 1989. [19] K. Deb, S. Agrawal, A. Pratab, T. Meyarivan, A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II, Technical Report 200001, 2000. [20] J. Knowles, D.C. Corne, Approximating the nondominated front using the Pareto archived evolution strategy, Evolutionary Computation 8 (2000) 149172. [21] K. Deb, Multiobjective Optimization Using Evolutionary Algorithms, Wiley, 2001. [22] E. Zitzler, K. Deb, L. Thiele, Comparison of multiobjective evolutionary algorithms: empirical results, Evolutionary Computation Journal 8 (2000) 125148. [23] E. Zitzler, L. Thiele, Multiobjective optimization using evolutionary algorithms a comparative case study, in: Fifth International Conference on Parallel Problem Solving from Nature (PPSN-V), Berlin, Germany, pp. 292301. [24] E. Zitzler, Evolutionary Algorithms for Multiobjective Optimization:

1145

Methods and Applications, Ph.D. thesis, 1999. [25] J.R. Schott, Fault Tolerant Design using Single and Multi-criteria Genetic Algorithms, Masters thesis, 1995. [26] B. Panigrahi, V.R. Pandi, R. Sharma, S. Das, S. Das, Multiobjective bacteria foraging algorithm for electrical load dispatch problem, Energy Conversion and Management 52 (2011) 13341342.

P.M. Pradhan is currently a graduate student working towards his PhD in Electrical Sciences at the Indian Institute of Technology Bhubaneswar. He has received B.E. degree in electronics and telecommunication engineering from the University College of Engineering, Burla in 2006 and M.Tech degree in Telematics and Signal Processing from the National Institute of Technology, Rourkela in 2009. His research interests include sensor network, cognitive radio, evolutionary computing and signal processing.

G. Panda is currently a professor of Electrical Sciences at the Indian Institute of Technology Bhubaneswar. He has received PhD degree from the Indian Institute of Technology, Kharagpur in 1982 and carried out post doctoral work at the University of Edinburgh, UK during 19841986. His research interests include digital and adaptive signal processing, intelligent instrumentation, sensor network, soft and evolutionary computing, data clustering, radar signal processing and acoustic noise cancellation.

Das könnte Ihnen auch gefallen