Abstract
The development of radio frequency identification (RFID) technology generates the most challenging RFID network planning (RNP) problem, which needs to be solved in order to operate the large-scale RFID network in an optimal fashion. RNP involves many objectives and constraints and has been proven to be a NP-hard multi-objective problem. The application of evolutionary algorithm (EA) and swarm intelligence (SI) for solving multiobjective RNP (MORNP) has gained significant attention in the literature, but these algorithms always transform multiple objectives into a single objective by weighted coefficient approach. In this paper, we use multiobjective EA and SI algorithms to find all the Pareto optimal solutions and to achieve the optimal planning solutions by simultaneously optimizing four conflicting objectives in MORNP, instead of transforming multiobjective functions into a single objective function. The experiment presents an exhaustive comparison of three successful multiobjective EA and SI, namely, the recently developed multiobjective artificial bee colony algorithm (MOABC), the nondominated sorting genetic algorithm II (NSGA-II), and the multiobjective particle swarm optimization (MOPSO), on MORNP instances of different nature, namely, the two-objective and three-objective MORNP. Simulation results show that MOABC proves to be more superior for planning RFID networks than NSGA-II and MOPSO in terms of optimization accuracy and computation robustness.
1. Introduction
Academic research into radio frequency identification (RFID) has increased significantly over the last ten years, to the point that RFID is used to build up an “internet of things”—a network connects physical things to the Internet that makes it possible to access remote sensor data and to control the physical world from a distance [1]. An RFID system consists of four types of important components (see Figure 1): (1) RFID tags, each placed on an object and consisting of a microchip and an embedded antenna containing a unique identity, which is called Electronic Product Code (EPC); (2) RFID readers, each having more than one antenna and is responsible to send and receive data to and from the tag via radio frequency waves; (3) RFID middleware, which manages readers, as well as filtering and formatting the RFID raw tag data; and (4) RFID database, which records RFID raw tag data that contains information such as reading time, location, and tag EPC.
In many real-world RIFD applications, such as production, logistics, supply chain management, and asset tracking, a sufficient number of readers are deployed in order to provide complete coverage of all the tags in the given areas [2, 3]. This gives rise to some challenging issues in the deployment of an RFID network, such as optimal tag coverage, quality of service (QoS), and cost efficiency. Therefore, our previous pioneering work pointed out that the RFID network planning (RNP) problem in RFID system is a key issue that has to meet many requires of the RFID system in order to operate the large-scale network of RFID readers in an optimal fashion [4]. In general, we defined that the RNP aims to optimize a set of objectives (coverage, load balance, economic efficiency and interference between readers, etc.) simultaneously by adjusting the control variables (the coordinates of the readers, the number of the readers, and the antenna parameters, etc.) of the system. As a result, in the large-scale deployment environment, the RNP problem is a high-dimensional NP-hard optimization problem with a large number of variables and uncertain parameters.
Obviously, optimization of the RNP is essentially a typical multiobjective problem (MOP). However, the methods used in the previous studies to solve the multiobjective RNP (MORNP) are always weighted coefficient approaches used to transform multiple objectives into a single objective [4–8]. Most of these methods are based on evolutionary algorithm (EA) and swarm intelligence (SI) optimization techniques, such as genetic algorithms (GA) [6, 9], evolutionary strategy (ES) [10], differential evolution (DE) [7], particle swarm optimization (PSO) [4, 5, 10], and bacterial foraging algorithm (BFA) [11]. Notice that these works considered only one object in RFID network planning or a single objective function that linearly composed several planning objectives and none of them can generate the tradeoffs between objectives. However, it is hard for users to determine these coefficients for optimization in RFID network. On the other hand, what can be acquired using a combination of coefficients is a single optimal solution instead of all the optimal solutions, namely, Pareto optimal solutions. Therefore, transformation of multiobjective functions into a single objective function is not the best choice for optimizing the real-world MORNP problem.
This paper substantially extends the previous work on RNP and can be distinguished from it from three aspects as follows.(1)A MORNP optimization approach is conducted in this work. In MORNP approach, four objectives, namely, coverage, load balance, economic efficiency, and interference, are considered simultaneously in optimization process. A fuzzy decision-making process for selection of the final solution from the available optimal points on Pareto frontier is also presented here.(2)Due to conflicts between different goals of the existing MORNP model, an efficient solution method should be used to search in the feasible solution space with the hope of finding the ideal RFID network layout while extracting a set of Pareto optimal solutions. Hence, this paper provides recommendations and guidance for the utilization of multiobjective EA and SI optimization techniques, such as the recently developed multiobjective artificial bee colony algorithm (MOABC), the nondominated sorting genetic algorithm II (NSGA-II), and the multiobjective particle swarm optimization (MOPSO). The success of multiobjective EA and SI is due to their ability of finding a set of representative Pareto optimal solutions in a single run.(3)By applying multiobjective approaches for solving the MORNP problem, a new framework was established that could handle different objectives and would enable the planner to find the optimal RFID network plan based on multiobjective EA and SI. Specifically, we formulated MORNP as two types of multiobjective problems, namely, two- and three-objective problems, whereas each two- or three-objective functions in RFID system are optimized simultaneously.
The rest of this paper is organized as follows. Section 2 gives the formulation of the MORNP problem. Section 3 presents the brief review of MOABC, NSGA-II, and MOPSO algorithms. In Section 4, the comparative study is performed for the three nature-inspired algorithms on solving the MORNP problem. Finally, Section 5 outlines the conclusions.
2. Problem Formulation
2.1. Multiobjective Optimization Problems
In many real-world optimization applications, the decision maker is always faced with the presence of multiple noncommensurable and often competing objectives. The solutions for the multiobjective problem (MOP) often result from both the optimization and decision making process [12]. When trying to solve an MOP, a set of trade-off solutions is the target of the solution algorithm and the one that will be chosen depends on the needs of the decision maker.
An MOP can be defined as: where multiobjective function includes objective functions, constraints and are equality and inequality functions, and is control variable.
In order to optimize the vector function, the concept tied to an MOP called “Pareto Optimal” is defined as follows. For (1), let , ( is the objective space) be two vectors, is said to dominate if for all , and . A point ( is the objective space) is called (globally) Pareto optimal if there is no such that dominates . The set of all the Pareto optimal solutions, denoted by PS, is called the Pareto set. The set of all the Pareto objectives vectors, , is called the Pareto front. Illustrative example can be seen in Figure 2.
2.2. Problem Formulation on Multiobjective RFID Network Planning
In this section, a mathematical multiobjective optimization model for the MORNP problem based on RFID middleware is proposed. The model is constructed from several different aspects. The deployment region of hotspots is supposed as a two-dimension square domain. The tags here are passive and based on the Class-1 Generation 2 UHF standard specification [13]. It means that they can only be powered by radio frequency energy from readers. The proposed multiobjective model aims to improve the QoS of RFID networks by simultaneously optimizing the objects including coverage, interference, load balance, and aggregate efficiency via regulating the parameters of RFID networks, including the number, location, and radiated power of readers. Generally the problem is formulated as follows.
2.2.1. Optimal Tag Coverage ()
The first objective function represents the level of coverage, which is most important in an RFID system. In this paper, if the radio signal received at a tag is higher than the threshold dBm, the communication between reader and tag can be established. Then the function is formulated as the sum of the difference between the desired power level and the actual received power of each tag in the interrogation region of reader : where TS and RS is the tag and reader set that deployed in the working area, respectively, and represents the set of readers which has the tag in its interrogation region. This object function ensures that the received power at the tag from the reader in , which is mainly determined by the relative distance and radiated power of the reader , is higher than the threshold , which guarantees that the tag is activated. That is, by regulating the locations and radiated power of the readers, the optimization algorithm should locate the RFID readers close to the regions where the desired coverage level is higher, while the areas requiring lower coverage are taken into account by the proper radiated power increases of the readers.
2.2.2. Reader Interference ()
Reader collision mainly occurs in a dense reader environment, where several readers try to interrogate tags at the same time in the same area. This results in an unacceptable level of misreads. The main feature of our approach is that the interference is not solved by traditional ways, such as frequency assignment and reader scheduling [4, 13], but in a more precautionary way. This objective function is formulated as where is the tag set in the interrogation region of reader . For each tag , this objective considers all the readers except the best one as interfering sources. That is, by changing reader positions and powers according to this functional the algorithm tries to locate the readers far from each other to reduce the interference.
2.2.3. Economic Efficiency ()
This aspect could be approached from various points of view. For example, due to the stochastic noise, multipath effect, and attenuation in the propagation channel, readers should be located closely to the center of tags in the hotspots. From this perspective, this objective can be reached by weighing the distances of each center of tag clusters from its best served reader. Here we employ -means clustering algorithm to find the tag cluster. It can be defined below: where dist() is the distance between the reader and the tag center and and are the position of cluster center and its best served reader, respectively. In this way the algorithm tries to reduce the distance from the readers to the elements with high tag densities.
2.2.4. Load Balance
A network with a homogeneous distribution of reader cost can give a better performance than an unbalanced configuration. Thus, in large-scale RFID system, the set of tags to be monitored needs to be properly balanced among all readers. This objective function is formulated as where is the assigned tags number to reader and is the maximum number of tags which can be read by the reader in unit time. It should be noticed that the take different values according to the different types of readers used in the network. This object aims to minimize the variance of load conditions by changing the locations and radiated power of readers.
2.3. Objective Constraint
All the tags in working area must be covered by a reader. This constraint can be formally expressed by the following formula: where is a binary variable that if the reader ; otherwise . So this constraint can maintain the power efficiency of network and ensure a complete coverage deployment.
3. Multiobjective EA and SI Algorithms for MORNP
In this section, we detail the representation of the individual solutions and give a brief description of the multiobjective EA and SI algorithms compared in this work.
3.1. Solution Representation
In this work, the task of RFID network planning is to deploy several RFID readers in the working area in order to achieve four goals described in Section 2. Figure 3 shows an example of a working area containing 100 RFID tags and 1 RFID reader, where the following three decision variables are chosen in this work: X: the x-axis coordinate value of the RFID reader, Y: the y-axis coordinate value of the RFID reader, P: the read range (i.e., radiated power level) of the RFID reader.
These variables can be encoded into solution’s representation shown in Table 1. We employ a representation that each solution is characterized by a ( is the total number of readers that deployed in the network) dimensional real number vector. In the representation, dimensions indicate the coordinates of the readers in the 2-dimensional working area, and the other dimensions denote the interrogation range of each reader (which is determined by the radiated power).
3.2. Multiobjective Artificial Bee Colony (MOABC)
ABC is a swarm intelligence algorithm developed by Karaboga motivated by the intelligent behavior of honeybees colony [14]. In ABC model, the colony of artificial bees contains three groups of individuals, namely, the employed, onlookers, and scouts bees [15]. Employed bees go to their food source and come back to hive and dance on this area. The employed bee whose food source has been abandoned becomes a scout and starts to search for finding a new food source. Onlookers watch the dances of employed bees and choose food sources depending on them. Since the original ABC is formulated as a single objective problem optimizer, we defined a new multiobjective algorithm in [16] named as MOABC. This algorithm incorporates two changes that allow its application in multiobjective optimization problems. The first modification applied is based on nondominated sorting strategy. That is, the MOABC algorithm uses the concept of Pareto dominance to determine the flight direction of a bee and it maintains nondominated solution vectors which have been found in an external archive. Secondly, the MOABC uses comprehensive learning strategy which is inspired by comprehensive learning particle swarm optimizer (CLPSO) [9] to ensure the diversity of population. Additionally, MOABC applies the crowding distance concept to calculate the corresponding value for all the solutions of the conflicting Pareto front and choose the sources of the best crowding distances. For further information about the MOABC algorithm please refer to [16].
3.3. Nondominated Sorting Genetic Algorithm (NSGA-II)
Deb developed the nondominated sorting genetic algorithm (NSGA) based on the classification of the population at various levels [17]. In this algorithm, before the selection, NSGA ranks the population by using the dominance concept. All nondominated individuals are classified into a category with a dummy fitness proportional to the population size. In order to maintain the diversity of the population, these individuals are distributed according to their fitness, subject to a distribution parameter (sharing parameter). This classified group is removed of the population and the remaining individuals are reclassified by the same procedure. This process continues until all individuals in the population are classified. Since the first individuals are of best quality, they always get more copies than the rest of the population, allowing the search in nondominated regions. NSGA got promising results; however, it was criticized for computational complexity problem. To repair this limitation, Deb et al. proposed an enhanced version of this method, called NSGA-II [18]. It solves the computational complexity problem through a fast nondominated sorting mechanism and a selection operator to combine the parent and child populations and select the N best solutions taking into account their quality and their distribution in the Pareto front. NSGA-II has become a standard multiobjective algorithm that has solved a lot of multiobjective problems. For further information about the NSGA/NSGA-II algorithms see [17, 18].
3.4. Multiobjective Particle Swarm Optimization (MOPSO)
Particle swarm optimization (PSO) has established itself as a successful swarm intelligence algorithm in a variety of optimization contexts. The rules of particle dynamics that govern this movement are inspired by models of swarming and flocking [19]. In PSO population, each particle has a position and a velocity and experiences linear spring-like attractions towards two attractors: its previous best position (pbest) and the best position of its neighbors (gbest). Until now there have been several proposals for extending PSO to multiobjective problems and these methods are called multiobjective particles swarm optimization (MOPSO) [20]. The greatest challenge in extending the PSO to multiobjective problems is deciphering the notion of guide and selection of pbest and gbest as in multiobjective scenario. In this work, we considered a Pareto-based MOPSO [21] to solve MORNP. This algorithm incorporates the main mechanisms of NSGA-II to a PSO algorithm. In this approach, once a particle has updated its position, instead of comparing the new position only against the pbest position of the particle, all the pbest positions of the swarm and all the new positions recently obtained are combined in just one set. Then, MOPSO selects the best solutions among them to construct the next swarm (by means of a nondominated sorting). This approach also selects the leaders randomly from the leaders set (stored in an external archive) among the best of them, based on two different mechanisms: a niche count and a nearest neighbor density estimator. For further information about this MOPSO algorithm see [21].
3.5. Optimization Procedure
The overall operating process of MORNP based on MOEA and MOSI algorithms can be described as follows:
Step 1 (initialization). The positions of all individuals of each algorithm are randomly generated, with each being a random point in the working area and being a random value within the transmitted power range of readers .
Step 2 (fitness evaluation). As described in Section 2, the individuals should be evaluated on all the objectives of MORNP including maximizing tag coverage (defined by (2)), minimizing interference (defined by (3)), maximizing economic efficiency (defined by (4)), and maintaining an optimal load balance (defined by (5)), in an order of decreasing importance. Accordingly, the algorithm handles these four objectives in a multiobjective manner in the next step.
Step 3 (population evolution). Compare the evaluated fitness values and update the position of each individual according to specific rules of MOABC, NSGA-II, and MOPSO, respectively.
Step 4 (termination condition). The computation is repeated until the maximum number of iterations is met or the system requirement is researched.
4. Experiment
4.1. Experimental Setup
We consider an idea example shown in Table 2 [4]. That is, the proposed algorithm is evaluated against a test working area: a 30 m × 30 m working space with 100 tags that are distributed by a specific topology (shown in Figure 4). Ten RFID readers, whose radiated power is adjustable in the range from 0.1 to 2 watt, are considered to serve this area. Here the interrogation range according to the reader radiated power is computed as in [4].
In this experiment, the performance of three multiobjective EA and SI algorithms, namely, the MOABC, NSGA-II, and MOPSO, is compared on two- and three-objective MORNP cases. The maximum generation for each algorithm is 1000. The initialized population size of 100 individuals is the same for all tested algorithms. For MOABC, archive size , elitism probability is 0.4. For NSGA-II, crossover probability , mutation probability (where is the number of decision variables), and distribution indices for crossover and mutation operators and , respectively. For MOPSO, the inertia weight started at 0.9 and ended at 0.5 and the learning rate , and for the other parameters set for MOPSO please refer to [21].
4.2. Best Compromise Solution Based on Fuzzy Decision
Upon having the Pareto-optimal set of nondominated solution, the proposed approach presents one solution to the decision maker in RFID middleware as the best compromise solution. In this work, a fuzzy-based mechanism is employed to extract the best compromise solution over the trade-off curve and assist the decision maker to adjust the generation levels efficiently [22]. Due to imprecise nature of the decision maker’s judgment, each objective function of the th solution is represented by a membership function defined as follows: where and are lower and upper bounds of th objective function. The higher the values of the membership function are, the greater the solution satisfaction is.
For each nondominated solution, the normalized membership function is calculated as where is the number of nondominated solutions and is the number of object. The best compromise solution is the one having the maximum of .
4.3. Two-Objective MORNP Optimization Results
In this case, the RFID network planning is handled as a multiobjective optimization problem, where each two objective functions are optimized simultaneously. According to the tag coverage-reader interference, tag coverage-load balance, tag coverage-economic efficiency, reader interference-load balance, reader interference-economic efficiency, and load balance-economic efficiency pairs, all obtained Pareto fronts by the MOABC, NSGA-II, and MOPSO algorithms are shown in Figures 5, 6, 7, 8, 9, and 10, respectively. Table 3 shows the best compromise solutions for each objective in the two-dimensional Pareto front.
From the results, we can observe that the trade-off among two selected competing objectives is obtained by emphasizing on nondominated solutions and getting a well-distributed set of solutions, respectively. From Figures 5–11, it is clear that the MOABC algorithm is able to obtain well-distributed Pareto-optimal fronts. From Table 3, we can see that MOABC gets the best convergence solutions for most of objective pairs.
4.4. Three-Objective MORNP Optimization Results
In this case, three competing objectives are optimized simultaneously by the MOEA and MOSI algorithms. According to the three-objective combination set, namely, tag coverage-reader interference-load balance, tag coverage-reader interference-economic efficiency, tag coverage-load balance-economic efficiency, and reader interference-load balance-economic efficiency, all obtained Pareto fronts by the MOABC, NSGA-II, and MOPSO algorithms are shown in Figures 11, 12, 13, and 14, respectively. Table 4 shows the best compromise Pareto-optimal solutions for each objective in the three-dimensional Pareto front, respectively.
It is clear that each RNP objective cannot be further improved without degrading the other related optimized objectives. Figures 11–14 clearly show the relationships among all presented objective functions. Between the obtained Pareto-optimal solutions, it is necessary to choose one of them as a best compromise for implementation.
Figures 15(a)–15(l) show the reader locations and radiated power contours for the four three-objective MORNP instances, in which the best results obtained by MOABC, MOPSO, and NSGA-II are compared in a visible way. A contour in the figures represents the points with the same radiated power (which equals the value assigned to that contour). It can be seen in Figure 15 that the power peaks in the working area are the points where the readers are placed. Then, the signal strength decreases with respect to the distance to the readers. The figures clearly show that all the MO algorithms try to (1) generate an optimal reader network layout with high tag coverage rate; (2) maintain sufficient distances between RFID readers to reduce interference; (3) provide a satisfactory economic efficiency by increasing the best-server areas; and (4) configure the network in a load balance scheme so that each reader in the network serves the optimal amount of tags according to its capacity.
(a) MOABC
(b) MOPSO
(c) NSGA-II
(d) MOABC
(e) MOPSO
(f) NSGA-II
(g) MOABC
(h) MOPSO
(i) NSGA-II
(j) MOABC
(k) MOPSO
(l) NSGA-II
It can once again be proved that the MOABC algorithm is giving better performance for four three-objective MORNP cases, while the other two algorithms sometimes have redundant readers and cannot provide full coverage. All the results confirm that the MOEA and MOSI methods are impressive tools for solving the multiobjective RFID network planning problem where multiple Pareto-optimal solutions can be obtained in a single run.
5. Conclusion
In this paper, we have compared three state-of-the-arts evolutionary and swarm intelligence based multiobjective algorithms, namely, MOABC, NSGA-II, and MOPSO, to solve the multiobjective RFID network planning (MORNP). This work differs from previous approaches to RFID network planning, because our new MORNP model focuses on use of multiobjective algorithms to find all the Pareto optimal solutions and to achieve the optimal planning solutions by simultaneously optimizing four conflicting objectives, instead of transforming multiobjective functions into a single objective function in the previous works on RNP. By applying multiobjective approaches for solving MORNP, a new framework was established that could handle different objectives and would enable the planner to find the optimal RFID network plan based on multiobjective EAs and SI.
To summarize, some of the contributions of this work are the formulation presented and applied to solve the MORNP and the comparison made among MOABC, NSGA-II, and MOPSO, where we analyze the behavior of each of them in two- and three-objective MORNP totally composed of 10 instances of different nature. As we have seen, MOABC is the algorithm that best results have obtained in most instances.
Evaluating new algorithms for this MORNP problem is a matter of future work. In particular, we have planned comparisons with other known multiobjective evolutionary and swarm-based algorithms that have received a lot of attention in the literature. Furthermore, we will also investigate the application of parallel or distributed techniques for solving the MORNP.
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publishing of this paper.
Acknowledgment
This research is partially supported by National Natural Science Foundation of China under Grants 61105067, 71001072, 61174164, 71271140, 51205389, and 61203161.