Interactive Genetic Algorithm For Mobile Sensor Networks

Published Date: 02 Nov 2017

Abstract: Herein, we describe an interactive genetic algorithm (GA)-based technique for the design of mobile sensor networks. Here, the node-mobility aspect requires an online algorithm to determine the optimal network-coverage solution for an area of interest. A real-time GA-based algorithm was developed to estimate the direction of node locomotion, enabling not only coverage of the target area but also estimation of the optimum energy consumption, thus extending the network lifetime. The simulation results indicated that the applied fitness function of the GA attained these objectives while avoiding convergence at local maxima.

Keywords: Mobile sensor networks, WSN, Genetic algorithm, Network life time, Optimum energy consumption

1. Introduction

Wireless sensor networks (WSNs) include numerous unsupervised devices capable of sensing, computation, and communication. These energy-restrained devices are expected to be used for many different kinds of applications [1]. For example, WSNs can be used for environment and habitat monitoring, traffic measurement on roads, vehicle tracking, and personnel tracking inside buildings. Even though WSNs have a variety of applications, their deployment usually has two common objectives: (a) obtaining the maximum area coverage for a specific number of nodes and (b) prolonging the operational life of the individual nodes [2].

A mobile sensor network is a WSN with locomotion capabilities, consisting of several nodes with sensing, computation, and communication functions [1]. This mobility aspect presents a design challenge in unknown environments. A genetic algorithm (GA) is a good approach to this kind of optimization problem. In this article we developed a real-time GA for the design of a network with maximum coverage and minimum energy consumption, thus allowing network lifetime to be extended. Network lifetime can be defined as the time it takes for the first node or a fraction of all the nodes in the network to be depleted of their energies. The aforesaid method is applicable in both a dynamic environment and a dynamic network topology [3]. Here we review the architecture of this type of

network. In most cases a basic sensor node consists of five main components: (a) a power supply that is considered to be the only energy source, (b) a controller with memory, (c) a sensing device, (d) a communications system, and (e) a mobile platform for mobile wireless sensor nodes. All these parts of the architecture consume energy. Especially during the spread of mobile sensor nodes, which can adjust their position to optimize their area coverage, there is a tradeoff between area coverage and energy usage [4].

GAs are heuristic search techniques used to automatically find globally optimal solutions while avoiding local maxima, inspired by the idea of natural evolution [5]. These methods have wide applications in model checking [3]. They are suitable for solving non-linear optimization problems and for finding the global optimization value of a fitness function. A GA fundamentally consists of three important components: recombination, mutation, and a fitness function. Many researchers have concentrated mostly on the fitness function when optimizing the solutions.

The WSN nodes are represented as bits of a chromosome [6]. A network of n nodes is represented by a chromosome of n bits. Each chromosome in a population represents a possible solution to a problem. Chromosomes permute information by crossover and generate new chromosomes. The chance of an individual chromosome surviving a crossover depends on its fitness value. The fitness of a chromosome is defined by several parameters, such as speed, energy consumption, and single hop delay. In a GA, fitness is valuated by the function defining the problem. The fate of an individual chromosome depends on the fitness value. The chances of survival are higher for better fitness values. A population includes several chromosomes and the best chromosome is used to produce the next population. For the initial population, a large number of random nodes are chosen. Based on the survival fitness, the population transforms into the future generation. A new population is produced by replacing one or two members of the population, whereas the next replaces all of them at each generation of evolution. [1].

In general a GA keeps a certain number of the best individuals from each generation and mixes them up to form the new generation. So the new generation will have some of the individuals from the previous population and others that evolved as a result of crossover and mutation [1]. In some GAs, some parent chromosomes are dropped, even though they have high priority, due to incompatible crossover which results in the production of a bad chromosome. For this reason, we suggest that in the case of the best parents with high priority the parent should not be dropped if the result of a crossover would be bad. Such parents are marked with a tag field for use in other crossovers.

Energy consumption is one the most important factors for measuring the efficiency of the network positioning. This energy can be divided into three parts: the energy of transmission for each node, the energy used by the cluster head to receive packets, and the energy of data transmission from cluster head to sink. But in this paper we assumed that the energy of transmission for each node represented the energy usage. Network coverage is another important factor in the fitness function [7]. In this article the fitness of a chromosome is designed to minimize the energy consumption and to maximize the network coverage area.

The remainder of this paper is organized as follows: (a) Section 2 reviews the related work, (b) Section 3 shows our contribution, (c) in Section 4 we propose a solution and elaborate on details of our algorithm, (d) Section 5 presents our simulation and analysis of results, and (e) Section 6 summarizes the conclusions.

2. Related Work

In 1992, Gage [8] proposed a blanket-coverage method to achieve a statistical arrangement of sensor nodes leading to extended coverage of the same area of interest as in the current work. Howard et al. [9] presented a GA-based approach that used a repulsive behaviour to spread the nodes throughout the resource. Here, a virtual "force" algorithm was utilized for node placement to avoid the problem of local-optimum solutions faced by many researchers [10–12]. They considered a complete model network in which only offline planning is required for deployment.

In 2002, Lindsey et al. [13] proposed the PEGASIS protocol, an extension of the LEACH algorithm, in which every node transmits aggregated data to its nearest neighbour and the transmission is repeated until the data packet is delivered to the base station. The PEGASIS method provides an advantage over LEACH in its robustness to node failure. Pan et al. [14] subsequently presented a two-tiered structure that improves energy efficiency by local hierarchical clustering.

Kalpakis et al. [15] proposed the Maximum Lifetime Data gathering Algorithm (MLDA) to find the edge capacities that allow maximum transmission flow by running a linear program. This algorithm is able to maximize the lifetime of a network with fixed locations of the nodes and the base station. One year later, Dasgupta et al. [11] extended the MLDA by applying a cluster-based heuristic algorithm, called CMLDA, where nodes are grouped into several clusters of predefined sizes. The cluster’s energy is determined by summation of the energies of the cluster member nodes. The distance between two clusters is computed as the maximum distance between every pair of nodes in the clusters. Following cluster formation, MLDA is applied as before.

Hussain et al. [12] later proposed an approach using a GA method to obtain an optimum solution including the number of clusters, the cluster heads, the cluster members, and the transmission schedule. Normally, a GA-based algorithm begins to form optimum clusters at the base station, and the same method was followed in this study. The proposed fitness function is based on parameters such as energy consumption, the number of clusters, cluster size, the direct distance to each sink, and cluster distance.

Qu et al. [2] introduced a sensor relocation method based on a multi-objective GA. These objectives were to maximize coverage area and minimize energy usage in the network.

Azadeh et al. [1] described the energy limitation in WSNs and then used a GA to create energy efficient routing. Their proposed algorithm generates a sequence of routing paths that maximizes the system lifetime.

3. Contribution

The nature of sensor nodes and the dynamic aspects of the environment create two main limitations for mobile sensor networks. First, the environment and its geographic features are dynamic, making offline planning through a static map inappropriate. Second, energy consumption is an issue in both static and dynamic networks. As the sensor nodes have limited energy for monitoring the resource and communicating with each other and should ideally function for a long time, a minimum energy consumption approach is highly desirable. To solve these problems, we addressed three network properties: First, the real-time GA-based algorithm can facilitate the interactive solution of the dynamic problem. Second, the algorithm takes power consumption into account as nodes are selected for movement and/or monitoring activities. Third, the GA operates randomly at different levels, yielding a convergence close to the static global optimum in coverage.

In this paper, we will introduce an Interactive Genetic Algorithm for Mobile Sensor Networks. For this algorithm, a powerful processor is needed, which can be installed in the base station. The base station needs to know the location of each sensor. Then the base station runs the optimization algorithm in order to decide where the sensors should move to in order to achieve the maximum coverage and minimum energy consumption. The effect of obstacles is also considered. Also, we considered a case in which each node has a scanning laser range-finder sensor and an omnidirectional camera to monitor its distance from nearby nodes and obstacles.

4. Algorithm

To perform a better comparison of the seven protocols examined in the second section -DSR, AODV, DSDV, TORA, FSR, CBRP and CGSR - the following sections compare them in terms of rate of packet delivery, routing overhead, path optimality and movement speed of nodes.

In this section, assumptions concerning the sensor network are listed and the sensing model and multi-objective GA model are both described.

A. Assumptions

The following assumptions are made for the sensor network:

1. All sensors can communicate with the base station directly or by multi-hopping.

2. Each sensor’s coverage is a circle.

3. All sensors have GPS or other location devices and can move to any position (with known coordinates) within their mobility range.

4. Sensors cannot sense through or move across boundaries and obstacles that are considered walls.

B. Sensing Model

The sensing model used here is a binary model. This means that the area within a certain distance from a sensor can be counted as 100% covered and the area out of the sensing range will be set as less than 50% covered since it cannot be covered by this sensor [1]. The sensing field is considered to be a grid. The coverage of the whole area is proportional to the grid points that can be covered. Considering the grid points (x,y) and (xi,yi), the possibility that they can be sensed by a sensor node Asi(xi,yi) can be described as:

(1)

The other objective of the function is to save the sensor’s energy, which means the distance travelled by the sensors must be minimized. Here, the average distance travelled is used. For a sensor network with Rs sensor nodes, the average distance travelled is equal to:

(2)

Where G(x, y) , and are the corresponding initial and final points of one sensor node. To manage all sensor nodes, we designate a selected subset of association nodes. Initially, all sensor nodes are grouped into the set called Cluster, which in turn is managed by association nodes called Cluster Heads (CHs). The developed algorithm consists of five steps, as explained in Fig. 1, which summarizes the GA-based process for dynamic environmental monitoring. All nodes transmit aggregated data to the common Base Station (BS) through their corresponding CHs. The base station is a constantly powered resource that processes the algorithm and transmits the results to all nodes.

Fig. 1. GA-based algorithm applied to the mobile sensor network.

4.1. Initialization

Initialization is the first step of the algorithm, in which the value of ti is set to initialize the algorithm, the nodes are clustered, and the CHs are determined. Due to locomotion of the sensor nodes in a mobile network, we introduce the concept of the Direction Vector (DV) to specify the situation of each node, similarly to prior studies [3]. DV is an n-dimensional vector (n) scaled to the interval [0, 1]. This value indicates the probability of the nodes moving in different directions; thus, the summation of this dimension is 1.0.  

Simultaneously, the GA explores node movement using the DV and the locations of the CHs. The cluster heads must know the status of each node, including its remaining power level. Therefore, we introduce another concept, the Member Vector (MV), which consists of a member-node identifier (ID) and its energy level. After monitoring the environment and collecting data from the nodes, every sensor member node periodically transmits its aggregated data to the corresponding CH and the MV is then updated. Therefore, any CH has a set of MVi (i Number of member nodes), which allows the real-time GA to calculate the energy required to deliver optimum network coverage. 

4.2. Selection

During each successive generation, a new population is selected for mating from among the current members based on fitness. Fitter individuals are similarly rated, leading to preferential selection of the best solution. Most of the fitness functions are designed with a stochastic part to choose some smaller, less fit members to help maintain the diversity of the population [15]. Among several available selection methods, the Roulette Wheel was chosen to distinguish appropriate individuals, with a probability given by:

(3)   n j i i i F F P 1

where Fi and n are the chromosomal fitness and population size, respectively, which, according to the Roulette Wheel, are each assigned values in the continuous numerical interval between 0 and 1.

4.3. Fitness Function

The fitness function is one of the most essential components of a GA; using this function, a score can be assigned to any chromosome to distinguish the fittest among the population. Individuals of high quality survive for the next generation, and they, along with a small group of low-value members, are the parents of succeeding generations. In our case, the score awarded by the fitness function is dependent on two different objectives. The first is the coverage criterion, which acts to increase the score, and the second is the amount of energy spent to cover and monitor the area. Higher energy consumption decreases the total fitness value; therefore, the fittest members are those which are able to make an appropriate trade-off between the two objectives of the DVs. The applied fitness function based on the coverage and energy criteria is given as:

Fitness(x) = nj=1 (coveragex,j,e – coveragex,j,e-1) – (energyx,j,e – energyx,j,e-1) –energyCH*p (4)

where coverage is the area covered, energy is the sum of the remaining power in the nodes, and p is the total number of data packets. The index n indicates the number of active nodes taking part in the transmission of data packets and e is the step counter. This function was formulated in an attempt to consider all conditions and so allow for a comprehensive comparison among the chromosomes.

4.4. Termination

The GA algorithm interactively calculates MVi, which should yield tu at the maximum. Further generations can obviously be used to estimate a more accurate solution; however, as the network locomotion is live, iteration should be limited. Sending MVi to the corresponding member nodes requires energy to transmit and load the new MV. This task may be done at long intervals to conserve energy and so extend the network lifetime. Finally, the algorithm is terminated when a given t value is reached or the entire target coverage is obtained

5. Implementation

To implement the algorithm, we utilized a network simulator to assess our proposed algorithm in two respects. First, the GA-based portion was implemented using a Java editor. In this case, we installed the Java Genetic Algorithm Package (JPAC) to test our algorithm in a manner consistent with prior studies. Next, we utilized OMNET++ to trace the movement of the nodes in a virtual environment, as shown in Fig. 2; the parameters are summarized in Table 1.

Table 1. Parameters of the tested virtual environment.

Figure 2. Example virtual environment.

Fig. 3. Energy consumption rate over the lifetime of the virtual environment. Fig. 4. Comparison of coverage with each method.

Figures 3 and 4 show the compared results of the proposed algorithm and those using the LEACH protocol with respect to network energy and lifetime over 200 time units (years). In Fig. 4, the unified energy consumption by the CHs results in a short lifetime in the LEACH protocol. Figure 5 shows that the time to the removal of the first node due to its low-energy status, or the death of the first node, is delayed compared to the LEACH protocol. Additionally, the network remains in working order as long as a minimum number of nodes are alive. Generally, due to the use of the algorithm fitness function that considers the energy status of the nodes and the distance between CHs and the BS, the individuals finally remaining yield a cluster formation with uniform energy consumption, significantly extending the lifetime of the network.

6. Conclusions and Future Work

The presented method utilizes a real-time GA, which has attracted the attention of many researchers. Online properties are used to direct the sensor nodes to the optimum locations by considering the amount of power remaining, leading simultaneously to maximized environmental coverage and minimized power-consumption metrics. The simulation results confirmed that the proposed fitness function fulfils these objectives. In the future, we plan to investigate a GA-based algorithm that requires fewer transmitted directions at longer time intervals to further reduce the energy costs incurred by receiving and loading new directions.