A green clustering implementation is proposed to be as the first method in the framework of an energyefficient strategy for centralized enterprise highdensity WLANs. Traditionally, to maintain the network coverage, all of the APs within the WLAN have to be powered on. Nevertheless, the new algorithm can power off a large proportion of APs while the coverage is maintained as the alwayson counterpart. The proposed algorithm is composed of two parallel and concurrent procedures, which are the faster procedure based on
Increasing numbers of enterprise offices [
Concerning energy efficiency in WLANs, D’Alessandro et al. introduce a multicarrier infrastructure to implement energy conservation in WLANs [
As an alternative utilization, the energyefficient strategy can be applied in highdensity WLANs for power conservation. The main components of such a strategy usually include procedures, which are initialization and preparation, the access and association management of end users, the analysis mechanism of network information, and the topology management of APs. A green clustering algorithm is introduced and performs the function as the part of initialization, which is mainly responsible for forming AP clusters and selecting the clusterhead APs. Green clustering is the first algorithm for energyefficient strategies and a topology that consists of clusterhead APs and can maintain the coverage of network the same (almost the same) as the traditional WLANs can be formed.
In this paper, a new green clustering algorithm based on DPSMOPSO (Dynamic Population Size Particle Swarm Optimization) is proposed, which will be referred to as GCDPSMOPSO. It consists of two different procedures, which are the faster procedure implemented through
The remainder of this paper is arranged as follows. An overview of related work and background knowledge are given in Section
In many offices and campuses WLANs are deployed consisting of hundreds or thousands of APs, which make the WIFI network highdensity in terms of AP, and the scheme of AP deployment is designed to guarantee peak demand, which is the reason that traditional WLANs mostly cause serious AP redundancy. We propose an energyefficient strategy in highdensity WLANs to reduce the policybased power consumption, and the green clustering based on DPSMOPSO (GCDPSMOPSO) is an important and initial part of our strategy.
SEAR (Survey, Evaluate, Adapt, and Repeat) [
Components of SEAR.
In our energyefficient strategy, as the same as in SEAR, we propose a green clustering algorithm. Firstly, the concept of green clustering is restated and clarified. The strategybased energy efficiency is implemented through powering off redundant APs when few users are online or when the traffic is low. The green clustering is the algorithm to form AP clusters and determine a clusterhead AP for each green cluster. According to the results of green clustering, the clusterhead APs are powered on, whereas other APs (called secondary APs) are off for energy efficiency while maintaining the same coverage as shown in Figure
Illustration of cluster formation.
Besides, it is worth mentioning that the transmitting power of a net device is only a tiny part of the total power consumption. Taking Cisco WAP4410N WirelessN AP of S series, for example, its transmitting power with a single antenna is 13 dBm to 17 dBm (12.6 to 50.1 in mW), while this AP’s total power consumption is about 10 W during normal work. Therefore, even when the transmitting power increases for satisfying the algorithm requirement, the impact on total power consumption could be ignored because the energy savings by powering off APs are much more significant.
In addition, we verified and implemented our green clustering algorithm based on an Evolutionary Algorithm (EA) in our previous work [
To implement green clustering, a method, known as DPSMOPSO (Dynamic Population Size Multiple Objective Particle Swarm Optimization), is introduced. DPSMOPSO is an approach of multiobjective optimization, in which we make improvement based on the classical MOPSO with dynamic population size controlling and mutational operation. Firstly, the classical PSO is introduced as follows.
PSO (Particle Swarm Optimization) is an evolutionary computation technique developed by Kennedy and Eberhart in 1995 and is inspired by the social behavior of bird flocking and fish schooling [
Initialize the population
Initialize the speed of each particle, and
Evaluate each of the particles in
Each particle is evaluated through assigning a fitness value, the value of which is equaled to the objective function value.
Store the particles in the repository
Update the velocity of each particle using the following expression,
where
Update
Maintain the particles within the search space.
Evaluate each of the particles in
Update the repository
Increment the loop counter.
In the above subsection, the classical PSO iterates to search for an optimal or nearoptimal solution. However, one of the main drawbacks is the excessive repetition during the update process of every particle within each iteration. Additionally, it is typical that, at the beginning of a classical PSO, most of the particles will be distributed far away from the potential global optimum. Then, as iterating, the particles will migrate to a region surrounding the global optimum. Therefore, later iterations serve and finetune the approach until the global optimum is reached. To reduce excessive computation complexity, we propose the complementary dynamic population size (DPS) to improve the execution speed of the classical PSO.
During the early phase, distributed positions of the particles are random and mostly are far away from the global optimum; the calculations then will make those far away particles to the region closer to the desired region. However, in this phase the algorithm process is out of order and timeconsuming, and the convergence rate tends to be exponential. Nevertheless, in the latter stages, the convergence rate will be asymptotic as the algorithm approaches the exact global optimal value. To improve the execution speed, instead of the operation in the classical algorithm, a small number of particles (about 10% to 25% of the population size) are generated at the beginning, and, as the increment of loops, additional particles are created and associated with the number of iterations until the number of the particles has reached the presetting population size. Moreover, additional particles are generated and policybased initialized, and three different methods of initialization are introduced. The first approach of initializing is to copy the best particle in
After EPP, the number of the particles remains, which is equal to the presetting population size. This phase of DPSPSO is named Maintaining Population Phase (MPP).
The PSO concept prefers a larger population size to allow exploration of the widest possible region of the solution space to recognize the region containing the global optimum. Once that region is identified, the process will finetune the converge on the global optimum. Not as many as particles required at the beginning of the fine tuning, the search space of particles becomes much narrower in the later process. If the search space is narrow enough, the number of particles is gradually reduced by dropping lower performance particles as measured by fitness function. This phase of DPSPSO is defined as Diminishing Population Phase (DPP). Introducing DPP into PSO process will significantly reduce the number of particles for subsequent iterations.
To conclude, components of proposed population size operator are EPP, MPP, and DPP, the invoking of which is in order, and an example of the population size operator is clearly illustrated in Figure
Diagram of dynamic population size operator.
In Figure
In addition to the improvement of population size, a mutational operation is introduced to maintain the diversity of the trajectory. At the beginning of the search, the mutation operator attempts to explore all the new particles. Then, the mutation rate might be decreased with increasing iterations. Therefore, the mutation range is affected by the rate. The pseudocode of the mutation operator applied is the following.
%
%
%
%
%
%
To verify the performance of complements, the DPSPSO is implemented and compared with the classical PSO, EPPSO (Expanding Population PSO), and DPPSO (Diminishing Population PSO) [
Most of the algorithm parameters of the four PSO methods are configured on the same value. Additionally, the presetting population size is set to 32, which means in the classical PSO 32 particles are employed, In EPPSO 32 will be the value of expanding result, in DPPSO the algorithm will start with 32 particles, and in DPSPSO the value of
To assess the performance, the four PSO approaches are applied to solve the following unconstrained maximization problem, which is introduced to assess the performance of evolutionary algorithms in [
It is worth noting that the global optimum of (
To maintain the same experimental conditions in [
Table
Comparison of execution times.
Run  Exec. time (millisecond)  Exec. speed improvement (%)  

PSO  EPPSO  DPPSO  DPSPSO  EPPSO  DPPSO  DPSPSO  
1  111.87  38.21  30.20  28.94  48.31  55.48  56.60 
2  104.79  34.02  36.53  29.27  55.58  53.19  60.11 
3  87.23  47.70  32.02  29.11  51.08  69.06  72.39 
4  91.75  35.36  34.75  29.07  62.01  62.68  68.87 
5  81.00  37.23  46.74  29.14  67.94  56.20  77.93 
6  81.17  37.59  40.42  29.26  67.35  63.87  77.62 
7  86.82  52.91  37.95  29.37  45.32  62.56  72.45 
8  98.42  48.80  52.32  29.02  44.16  40.58  64.26 
9  92.67  45.21  41.36  29.25  50.77  54.93  67.99 
10  86.88  35.23  35.06  29.17  65.42  65.84  72.61 


Ave.  92.26  41.23  38.73  29.16  55.82  58.44  69.08 
Table
Convergence comparison.
PSO  EPPSO  DPPSO  DPSPSO  

Mean dist. to global optimum  0.0220  0.0268  0.0177  0.0158 
Standard deviations of dist. (× 
1.54  1.97  1.59  1.30 
Mean iterations to converge  6987  8229  1704  2230 
Convergence speedup over classical PSO  —  −17.78%  75.61%  68.08% 
From the results in Tables
In the above subsection, a modified PSO approach, known as DPSPSO, is proposed, and the results of comparative experiments demonstrate the better performance in terms of both execution time and accuracy.
To implement green clustering, the original problem is rewritten to a multiobjective optimization (detailed in Section
The general mechanism of MOPSO is similar to PSO (already detailed in Section
MOPSO is implemented as the comparison with the proposed DPSMOPSO, and test functions and experimental parameters are set to the same values as used in [
It is worth noting that the average distance to the Pareto optimal set
Table
Comparison between classical MOPSO and the proposed DPSMOPSO.
MOPSO  DPSMOPSO  

Test function  1  2  3  1  2  3 
Exec. time (ms)  7.6  4.6  9.8  2.9  2.1  4.5 

20.57  14.74  11.61  17.31  13.92  9.57 
Standard deviation (× 
2.86  2.02  7.21  2.37  1.83  4.36 
In the above section, DPSMOPSO (Dynamic Population Size Multiple Objective Particle Swarm Optimization) is implemented. In this section, green clustering based on DPSPSO (GCDPSMOPSO) is proposed, the framework of which is illustrated in Figure
The algorithm flowchart of GCDPSMOPSO.
The precondition of the green clustering implementation which is denoted by the database of “Loc. (Location) of AP and RP (Reference Point)” in the diagram is that the complete knowledge of the physics positions and status of all APs within the WLAN are achieved. Moreover, the other parts of the framework in Figure
The content in this subsection includes the parts of Simple Modeling, Complicated Modeling, Coverage Radius of AP, and Propagation Model in Figure
In Proc. F., to guarantee time efficiency, a simple and less timeconsuming metric of coverage is introduced, which is based on Friis free space equation. The simple metric
Furthermore, the metric for Proc. A. is more complex; unlike achieving a decision through a numeric comparison, an indoor propagation system based on the combinations of raytracing and FDTD (FiniteDifferent TimeDomain) techniques is introduced. In order to implement this type of estimation of APs’ coverage, the region where the WLAN resides is modeled and meshed for a large amount of RPs. Additionally, TPs (Test Points) are selected among RPs to assess the coverage. To estimate the region covered by an AP, the wave propagation and penetration from the current AP are modeled within a constrained and reasonable area. If the RSS (Received Signal Strength) at a TP is greater than a presetting value which is equal to the Rx sensitivity in accordance with some specific protocol (to be specific 802.11b is introduced here), the current TP is considered as being covered by the corresponding AP. The analogous combinations of raytracing and FDTD have already been proved in [
The content in this subsection includes the parts of Proc. F., OPR. I, and number of AP clusters in Figure
The analogous procedure of Proc. F. has already been studied and implemented in our previous work [
Additionally, before executing
Besides, it is worth noting that the proposed green clustering implementation is designed as a parallel structure for the purposes of speed and accuracy, and the faster procedure is for speed. Therefore, the above branch in Figure
The content in this subsection includes the parts of Proc. A. and OPR. II in Figure
In Proc. A., a set of trial clusterhead APs is considered as a feasible solution, and the region covered by the WLAN is considered as the boundary for feasible solutions, which is the search space as well. Therefore, the solving process of green clustering can be considered as the optimization procedure to achieve the global best solution within the search space. Moreover, the objectives are prerequisite, and the problem of green clustering can be given by
Each AP in the trail feasible solution is represented by the attribution of physics position expressed as
In our earlier work [
The WLAN region is meshed; RPs and TPs are generated in order. Each RP is attributed with the information of corresponding physics position to assess the coverage, and, in order to represent the results of assessment, two additional attributions are assigned for each TP, which are the properties of coverage and overlap levels. Additionally, the concept of resolution is introduced, and lower or higher resolution represents a larger or smaller number of TPs, which are selected from RPs to assess the coverage and overlap.
Proc. A. is as follows.
Initialization:
Obtain the physical positions of all APs in the WLAN.
Calculate and determine the original number of green clusters
Set a smaller number of iterations for DPSMOPSO expressed as
apply the DPSMOPSO and the objective values are calculated with the low resolution and Complicated Modeling;
increase the counter of iterations.
Evaluate the results with
Initialization for another DPSMOPSO processing:
set a higher resolution and a bigger number of iterations to be executed for the current procedure expressed as
apply the DPSMOPSO and the objective values are calculated with the high resolution;
increase the counter of iterations.
Set a bigger ratio
Additionally, once Proc. A. is completed, OPR. II is carried out, which denote operations of powering on APs in
As shown in Figure
In this section, simulations are performed to verify the method of the proposed green clustering implementation GCDPSMOPSO. Three different experimental environments are introduced, and the algorithms in [
Exp. (Experiment) 1 and Exp. 2 are conducted in our previous work as well, and in consideration of comparison, the experimental conditions and environments are set the same. Exp. 1 is in an ideal environment, where 81 APs are deployed uniformly in a 100 m × 100 m region, as shown in Figure
The deployment of APs in Exp. 1.
In GCDPSMOPSO, Proc. F. starts with a bigger number
The results of Proc. F. and A. are shown in Figure
Results of Exp. 1: the former is obtained by Proc. F. and the latter is achieved by Proc. A., where 9 and 8 APs are powered on, respectively.
Unlike Exp. 1, Exp. 2 is in a practical model. To be specific, the floor where our office resides with 27 APs is modeled to verify the proposed algorithm, as shown in Figure
The floor layout and AP deployment in Exp. 2 where 27 APs are deployed, and the numbers denote doors.
The results of Proc. F show that 14 APs are required to cover the interesting region, whereas 9 APs are needed to maintain the coverage according to Proc. A. (shown in Figure
Results of Exp. 2, where 9 APs are powered on and 66.7% of energy can be saved.
In Figure
The environment in Exp. 3 is in another laboratory of ours, and the floor layout where 8 APs are deployed is modeled. Furthermore, in this experiment, a new function of GCDPSMOPSO is discussed and simulated. In Exp. 1 and Exp. 2, positions of APs are previously fixed, so the results of GCDPSMOPSO are to determine that how many and which APs are powered on; therefore, Exp. 1 and Exp. 2 prove the AP selection function. If GCDPSMOPSO is employed before the AP deployment, the results can be taken as reference or guidance for AP deployment and considered as the foundation of the AP deployment, which can provide the basic network coverage. Therefore, another function of GCDPSMOPSO is AP planning that is proved in this subsection. The floor layout and AP deployment of Exp. 3 are shown in Figure
The floor layout and AP deployment in Exp. 3 where 8 APs are deployed, and the numbers denote doors.
From Figure
Furthermore, for performing the AP planning function, we assume that the coverage of all areas (except for the elevator shaft) in Figure
Results of the planning function in Exp. 3: 4 APs can maintain the overall network coverage, where a blue triangle denotes a location to deploy an AP, different colors represent different values of RSS, and only the maximal RSS at each TP is recorded.
Referring to Proc. F., all RPs in the WLAN region are considered as patterns to be classified, and the results are employed as the upper boundary of the clusterhead AP number for Proc. A.
And, referring to Proc. A., the search space for DPSMOPSO becomes the region composed of RPs. Therefore, within each iteration of DPSMOPSO, the algorithm is searching for the constrained space composed of RPs, and objective values are calculated based on the locations of RPs in the feasible solutions and all TPs within a reasonable distance. To be specific, in terms of wave propagation in Exp. 3, if the distance between Rx and Tx nodes is more than 30 m, the Rx region is determined to be uncovered by the current Tx device in an indoor environment. The results of AP planning are shown in Figure
In Figure
The proposed green clustering implementation has two functions, which are AP selection and planning. In this subsection, both functions are compared with green clustering algorithms in [
This function is implemented after AP deployment and for the purpose of choosing specific APs to maintain basic network coverage. Firstly, green clustering in [
The floor layout and AP deployment of the comparison task, where, on 2 floors, each of which is 50 m × 20 m and 14 APs are deployed.
According to the environment given in Figure
Referring to the method in our previous work [
In consideration of fairness, a parameter which can describe the green clustering gain of algorithms in different indoor environments is introduced [
Firstly, the value of
Proc. A. in our previous work is implemented through the modified EA. In this paper, a much more timeefficient algorithm DPSMOPSO is employed. To prove the efficiency of our new algorithm, we test both methods in the same condition (of Exp. 2), where the population size is set to 32, and the number of total generations and iterations is set to 1000. Ten independent experiments are executed for each algorithm, and the results are shown in Table
Comparison of execution time.
Run  Mean execution time (s)  

The old  The new  
1  2.3273  0.3650 
2  2.5155  0.8035 
3  2.2572  0.5725 
4  2.2773  0.4937 
5  2.4986  0.5715 
6  2.8065  0.4795 
7  2.3461  0.4876 
8  2.5743  0.6490 
9  2.4549  0.6409 
10  2.7235  0.6417 


Ave.  2.4781  0.5705 
On account of the complexity of calculating coverage for Proc. A., the time consumed for each generation or iteration is high, especially when compared with the execution times in Table
In terms of guidance for AP deployment, Liao et al. propose an optimal set under the diamond pattern to achieve optimal coverage [
In [
The method in neither [
Highdensity WLANs are deployed in increasing numbers of enterprise offices and college campuses at present. To ensure the network coverage, all of the APs in the WLANs have to be powered on, which causes severe energy wastage. In this paper, the energyefficient strategy is discussed and a new green clustering implementation, which is an innovative algorithm, is proposed.
The proposed green clustering is composed of two parallel and concurrent procedures, which are known as Proc. F. and A. In terms of the different procedures, the problem descriptions of green clustering are based on different models. Referring to Proc. F., the green clustering is rewritten to a classification problem, and
Finally, in Section
The green clustering algorithm is proposed as the first method for energyefficient strategy in highdensity WLANs in this paper. The algorithm will be foreshadowing for future study energyefficient strategy, in which the optimization for transmitting power of clusterhead APs will be concerned. Additionally, GCDPSMOPSO will be supplemented comprehensively further. After green clustering, the mechanism of estimating user demands based on Markov Chain Model and queuing theory will also be studied, and access selection scheme based on IEEE 802.11 PCF (Point Coordination Function) for energy efficiency in highdensity WLANs will be proposed in the near future.
The authors declare that there is no conflict of interests regarding the publication of this paper.
This research is supported by the National Science Foundation (Grant no. 61101122), National Science and Technology Major Project (Grant no. 2012ZX03004003), and National High Technology Research and Development Program of China (863 Program) (Grant no. 2012AA120802).