An SDN/ML-Based Adaptive Cell Selection Approach for HetNets: A Real-World Case Study in London, UK

Heterogeneous networks (HetNets) are one of the key enabling technologies for next-generation networks. They aim to provide high capacity, low installation cost, and distributed traffic loads. The cell selection issue is an open research problem in HetNets, due to the different characteristics of base stations and the existence of a large number of them. In this paper, a novel software-defined networking (SDN)/machine learning (ML)-based adaptive algorithm is proposed, called adaptive two-tier, based on the K-nearest neighbor (A2T-KNN) algorithm. It is designed for millimeter wave (mmWave)-based HetNets and it has the ability to adapt to the various movement features of moving vehicles, as well as the different characteristics of the base stations. A real-world case is considered in the center of London. Simulation results demonstrate that A2T-KNN achieves high prediction performance in association with different vehicle features and configuration information. It outperforms other related schemes in terms of average number of handovers by up to 45.83%. Moreover, it was found to enhance the average achievable downlink data rate and network energy efficiency achieved by vehicles by up to 17.18% and 16.86%, respectively.


I. INTRODUCTION
L ONDON is the capital city of the United Kingdom (UK) [1]. It has a geographical area of 607 square kilometers and it is the most densely populated city of the UK. In 2016, the population density per square kilometer was 5590. To make London a smart city, the Smart London Board was established to transform the traditional systems of energy, healthcare, pollution management, transport, and traffic into smart services [2], as figure 1 shows. By 2036, London will become a mega-smart city with 10 million citizens [3].
Smart city refers to a modern city that employs advanced technologies to enhance the lives of its residents [4]. Wireless cellular networks are a fundamental component of smart cities [5]. They are used to enable numerous applications and services, such as the Internet of Things (IoT), virtual reality, and many more. [6]- [8]. Future wireless networks are projected to apply heterogeneous networks enabling technology to fulfill the different requirements of user equipment [9]. HetNets are networks that have a mix of conventional high-power macro and low-power small base stations (BSs) [10]. Small BSs are deployed in HetNets to improve capacity and to distribute traffic loads with a low installation cost [11], [12]. Figure 2 represents a map of London, showing the locations of the macro and small BSs operated by Vodafone. As shown in the figure, there is a high density of small BSs in the middle of London.
Cell selection refers to the process of choosing a serving BS based on certain criteria. However, the cell selection issue in 5G HetNets is a challenging task that affects the performance of users and the network due to different cell sizes, as well as high spectrum reuse [13], [14]. In addition, the movement of vehicles in the small coverage area of the small BSs leads to increasing handover (HO) rates [15]. Handover procedure includes three phases; preparation, execution and completion [16]. HO phases, showing the signaling overhead during these phases. In the preparation phase, a vehicle measures the radio resource management (RRM) parameters and then transmits a measurement report to the source BS which decides whether to perform an HO or not. Then, the source BS sends an HO request to one or more target BSs. These BSs perform admission control and they transmit an HO response to the source BS if they can provide the requested resources. At the end of this phase, the source BS selects the best target BS based on the received HO responses. In the execution phase, the source BS sends an HO command to the vehicle to connect to the selected target BS. When the vehicle associates with the target BS, it sends an HO confirmation to the target BS to finish the execution phase. In the completion phase, the target BS informs the core control plane (C-plane) functions to switch path from the source BS to the target BS by transmitting a path switch request. Then, when the path switch response is received by the target BS, it sends a release resource message to the source BS to complete the handover process [17], [18].
Cell selection schemes vary in term of selection method, network modeling, and problem solving, as shown in figure  4. They are classified based on the selection method into static and dynamic. Static selection refers to choosing the serving base station depending on a fixed method based on one criterion or multiple criteria. Dynamic selection, which is also called adaptive selection, means that the method of choosing a serving cell varies based on predefined values [12]. The dynamic method can be applied to single-tier or multi-tier networks. Wireless networks can be modeled based on hypothetical-based (such as hexagonal grid and random distribution) or real-world-based approaches. Solving the cell selection issue can be achieved by applying machine learning techniques or other approaches. Machine learning is a powerful tool that is based on different algorithms to give computer systems the ability to learn from data without programming them explicitly [19]. Nowadays, machine learning techniques have proven to be effective in many prediction tasks [20]. Software-defined networking is one of the most recent network architectures that aims to facilitate the network management task and to enhance the utilization of network resources in an efficient way [21]. The cell selection decision can be centrally coordinated by using an SDN controller [22], and the combination of SDN and ML creates a new network management solution [23]. Applying the traditional scheme that considers the static method with only a single criterion, the Received Signal Strength Indicator (RSSI), in 5G cellular networks is not appropriate due to the existence of a high density of BSs [24], [25]; this will lead to increasing handover rates and service interruptions with unbalanced loads [12], [26]. Most of the existing works use static methods to select the base station to be associated with [12]. There is a need for an adaptive cell selection scheme that can select the best base station by considering the HetNet's features and user equipment (UE) movement information. In addition, most recent works are based on single-tier selection of base stations. In fact, it is necessary to consider multi-tier, where the macro BS tier is still essential to serve high mobility UEs in order to decrease the handover rate [27]. Moreover, applying a certain cell selection strategy to a real-world case is preferable to study the effectiveness of the proposed algorithm in a realistic environment.
In this paper, the major contributions are as follows: 1) We study the cell selection problem of heterogeneous ultra-dense networks by using supervised learning technology. We propose a novel SDN/ML-based adaptive algorithm called A2T-KNN. It aims to intelligently choose the best BS from the two-tier BSs based on vehicle information and the features of the HetNet. 2) We model the heterogeneous ultra-dense networks based on a real-world dataset that was collected in the UK. The macro and small BSs that are operated by Vodafone and located in the central area of London are selected to be the system model.  and IV, respectively. Section V presents a performance analysis of the proposed A2T-KNN algorithm. The paper is concluded in Section VI. Appendix VI provides lists of the main abbreviation and symbols that are mentioned in this paper. In [29], Tesema et al. developed a fast cell selection (FCS) approach for 5G ultra-dense networks to select serving cells from a set called Active Set (AS). To improve the reliability of the communication system, multi-connectivity VOLUME 15, 2021 was considered. The serving base station was selected based on the signal-to-interference-plus-noise ratio value. The study considered both fast and slow user equipment. Simulation results demonstrated that the proposed approach could overcome the problem of radio link failures. Moreover, the achievable throughput was enhanced in comarison with other single-connectivity strategies.

Cell Selection Schemes
In [30], Cacciapuoti, A. S. proposed a mobility-aware cell selection scheme for millimeter-wave networks. The proposed scheme considered the distribution of the loads among the small cells to avoid cell congestion. It reduced the number of frequent handovers between small cells and took into consideration the problems of non-line-of-sight propagation effects, blockage, and directionality. In addition, it could handle changes in the network topology and channel conditions. A polynomial-time complexity figure algorithm was designed to solve the user association issue. The numerical results showed that the proposed scheme outperformed the traditional RSS-based approach.
Wickramasuriya et al. proposed a cell selection algorithm for 5G cellular networks in [31]. A Recurrent Neural Network (RNN) was used to predict the optimal BS that a mobile user would be associated with. To train the proposed RNN model, received signal strength (RSS) values were used. The proposed RNN architecture has three layers; input, hidden, and output. The RNN model has 640 neurons and the activation functions used are sigmoid and tanh. To evaluate the performance of the proposed approach, Google's Python-based Tensorflow library was utilized. The learning rate was set to 0.0003 and the model training took 35 minutes. An area of 36 km 2 was considered, with eight base stations distributed randomly. A mobile node, which can be a pedestrian or a vehicle, can connect with the three nearest base stations. Simulation results show that the proposed algorithm achieved 98% accuracy in predicting the next BS.
In [32], Kishida et al. proposed a cell selection scheme for 5G multi-tier Radio Access Networks (RANs). The selected cell is the cell that has the maximum signal-tointerference-plus-noise ratio (SINR) value and is located in the user's direction. Real locations of base stations are considered in a metropolitan environment in Shinjuku, Tokyo. Two kinds of users are considered: pedestrians, with walking speed 3 km/h, and cars with driving speed 40 km/h. Simulation results demonstrate that the proposed scheme made improvements in terms of the number of handovers by 30%.
In [33], a conditional random field (CRF)-based method was proposed by Zhang et al. for predicting the optimal serving base station. The proposed scheme is called CRFcell selection protocol (CRF-CSP) and it relies on setting up a grid that covers the area of interest in the 3D space. The CRF-CSP is based on localization information to find the nearest grid point. In the model training phase, the optimal cell identifications and SINR values are given. Simulation results demonstrate that the proposed CRF-based method can predict the optimal base station with a high prediction accuracy (90%). In addition, it has superiority over other simple heuristic schemes.
A machine learning-based user association scheme is proposed by Zappone et al. in [34]. It aims to maximize sum date rates achieved by users in massive multiple-input and multiple-output (MIMO) networks. The best serving base station is predicted using a trained model based on a feed-forward artificial neural network (FF-ANN). The geographical locations of users are given to the ANN model as inputs. The FF-ANN is structured as four fully connected layers with neurons of 128, 64, 64, and 40 respectively. The Rectified Linear Unit (ReLU) activation function is employed in the first and the third layers, while a sigmoidal activation function is used in the second layer of the FF-ANN. The adaptive moment (ADAM) estimation algorithm, which is a method for stochastic optimization, is used with Nesterov momentum for training the model [35]. The number of samples is 155,000, which is split into a training set of 140,000 samples and a validation set that includes 15,000 samples. Numerical results prove that the proposed method significantly reduces the computational complexity of the user association compared to the traditional cell selection scheme and it achieves the same performance as the traditional approach.
A user association approach was proposed by Elkourdi et al. in [36] for two-tier HetNets. It solved the cell selection issue by applying a Bayesian game model between two players, i.e., user equipment (UEs) and access nodes (AN). The UE can be either delay-sensitive (DS) or delaytolerant (DT), based on their delay requirement. Simulation results show that the proposed approach outperformed the conventional and cell-range-expansion (CRE) schemes in terms of the probability of proper cell selection and endto-end latency.
Khan, H. et al. introduced an ML-based cell selection approach for vehicles in mmWave networks in [37]. Distributed deep reinforcement learning (DDRL) is used to solve the vehicle association problem. The reinforcement learning problem is formulated as a Markov decision process (MDP). An Asynchronous Advantage Actor Critic (A3C) framework is used that includes actor and critic. Actions are sent from roadside units (RSUs) to a central entity that is responsible for calculating rewards to the RSUs. The proposed scheme decreases the control overhead and the computational complexity compared with other complex methods. Numerical results show that the proposed DDRLbased scheme has superiority over many cell selection schemes in terms of achievable sum rate by up to 15%.
Liu et al. proposed a cell selection approach that integrates the advantages of fuzzy logic and multiple attribute decision-making algorithms to perform the BS selection task in [38]. It depends on the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS). The best serving BS is chosen based on several network parameters; reference signal receiving power (RSRP), SINR, and jitter.
A subtractive clustering algorithm is applied to generate a suitable fuzzy membership function. The simulation results demonstrate that the overall performance of the proposed strategy is better than the RSS-based schemes. It enhances the network performance in terms of the number of handovers by almost 90% and the rate of ping-pong handovers by 10%, while providing an optimal Quality of Service (QoS) level.
In [39], Waheidi et al. proposed a distributed multiclass user association algorithm, known as Cell Association based on a Multi-Armed Bandit (CA-MAB) game. Static and mobile environments are considered in this work. User equipment and low-powered IoT devices are the classes that are focused on. The evaluation results demonstrate that the proposed CA-MAB algorithm improves energy saving and the aggregated throughput. In addition, the mobility of devices affects equilibrium, throughput, and energy efficiency.
Zhang et al. introduced a deep learning-based cell selection scheme designed for ultra-dense networks in [40]. A two-tier heterogeneous ultra-dense Network was considered, consisting of macro-and small BSs. It aimed to solve the cell selection problem based on base station load. A U-Net convolutional neural network (CNN) was trained to select the optimal base station by mapping channel gain values onto images. Simulation results showed that the proposed CNN-based approach improved network robustness and decreased the computation time.
In [41], Sun et al. introduced two coordinated multipoint (CoMP) handover schemes for heterogeneous ultra-dense networks. The schemes are called movement-aware CoMP handover (MACH) and improved MACH (iMACH). They are based on estimating a dwell time of a user inside a serving cell. The MACH approach chooses n BSs that have the strongest received power with dwell time longer than a predefined threshold. The iMACH scheme selects n − 1 BSs, as performed by the MACH scheme, and adds the nearest BS. In the MACH scheme, the handover is initiated when the farthest base station in the cooperating BSs set becomes the closest; in contrast, the handover is performed under the iMACH scheme when the closest BS becomes the farthest one. Simulation results shows that the proposed schemes improved the average throughput and the coverage probability and decreased the handover rate.
In [42], a user association method, known as RTP), was developed by Qin et al. for 5G ultra-dense networks. It selects base stations based on estimating the residence time inside a cell, where a predefined time threshold is set. The base station that has the strongest received signal power with a residence time longer than the predefined time threshold will be chosen as a serving BS. Simulation results prove that the proposed HO RTP scheme outperforms the conventional scheme in terms of average throughput achieved by users.
Alablani and Arafah proposed an adaptive cell selection (ADA-CS) approach for two-tier HetNets in [12]. It selects the optimal serving base station based on the various characteristics of network cells and vehicle movements. It performs six phases to achieve the cell selection task: configuration, decision-making, filtering,Handover based on Resident Time Prediction (HO narrowing, selecting, and handover triggering. Simulation results show that the proposed ADA-CS approach outperforms the traditional and recent cell selection schemes in terms of average achievable downlink data rates and spectral efficiency per vehicle by up to 3.98%. Moreover, it decreases the number of handovers by up to 42.39%.
The limitations of recent cell selection works that are presented in this section are: • Most recent works follow a static, non-adaptive, strategy to select the serving base stations. As there are multiple tiers in HetNets, adaptive selection is preferred that can be performed by setting up certain thresholds to switch between the network tiers. For high-speed vehicles, macro-BSs are desired to maintain the network performance. On other hand, low-and medium-speed vehicles will be based on small BSs as the serving BSs. • Most recent works give the highest priority to BSs that have the greatest receiving power to enhance the achievable throughput. In fact, in mobility environments, the closest BS that has the strongest receiving power will be far away when the user is moving forward. Therefore, relying on this principle leads to a degraded network performance due to unnecessary handovers. • Some works depend on the estimation of the cell staying time, which is an essential factor of the serving cell selection. Moreover, these works, for simplicity purposes, estimate the staying time based on the assumption the user is at the edge of the cell. • The number of ML-based works is fewer than the non-ML-based works, whereas predicting serving BSs needs to be based on machine learning algorithms to reduce the computational complexity. Moreover, input features for a machine learning model should be set carefully so that the trained model can solve the cell selection problem efficiently. • Applying a cell selection strategy in a real-world context is preferable to study the effectiveness of the proposed protocol. Some works were tailored to certain typologies and applying them to a real-world case will lead to undesirable network performance.
Based on the limitations mentioned above, there is a need for cell selection methods that are adaptable for selecting the serving cell to be associated with in order to maintain network performance. In addition, relying on machine learning algorithms is a trend nowadays that we should take advantage of to reduce the computational complexity and prediction time. Furthermore, implementing a proposed cell selection algorithm on a realistic environment is desirable to determine the effectiveness and applicability of the proposed strategy. VOLUME 15, 2021

III. THE PROPOSED A2T-KNN SCHEME
In this section, the proposed A2T-KNN scheme is discussed in terms of the SDN/ML-based model building process and the framework of the proposed approach.

A. THE PROPOSED SDN/ML-BASED MODEL BUILDING
To build the proposed machine learning model, five main stages have been passed through, as shown in figure 5 which are: • Stage 1: Data Preparation: This stage aims to collect, generate and prepare information from vehicles and from macro-and small base stations. At the end of this stage, data that is essential to train and validate the proposed machine learning model will be ready. 1) BSs Dataset Collecting: In this step, the appropriate dataset for the macro and small base stations should be collected. BSs dataset can be found over the web as a single dataset that saves information related to both of macro and small BSs. On other hand, the macro and small BSs information may be found as two separated databases. The geographical location information of BSs in terms of latitude and longitude coordinates must be exist in the BS dataset. The collected BSs dataset that is used in this work is described in section IV-B. 2) Vehicle Dataset Generation: In this step a vehicle dataset is generated using Google Maps and MATLAB simulator. The explanation of vehicle dataset generation process is given in details in section IV-B. 3) Data Cleaning and Labeling: In the cleaning step, the data that is not used by the proposed cell selection scheme to predict the next BS is removed. It is worth mentioning that the central area of London has a high density of small cells, with a number of the traditional macro-cells, which makes the area suitable for our study. The labeling process is performed based on the A2T scheme that is described in Algorithm 1, where   supervised learning using a training set.
-K-Nearest-Neighbor (KNN): This is a widely used and effective classifier, based on a predefined parameter (K) to determine the number of neighbors used in a similarity calculation that depends on the distance between samples. In this work, to measure the distance between two samples, the Euclidean distance is used.

-Feed-Forward Back-Propagation Artificial
Neural Network (FFBP-ANN): This is a popular machine learning algorithm that simulates the human neuron system. It is implemented based on feed-forwarding of data and back-propagation of errors. In this study, an FFBP-ANN is structured as three layers; input, hidden, and output. The number of neurons in the hidden layer is ten, while the number of neurons in the output layer is 658, which equals the summation of the number of small and macro BSs. The activation function used is tansig and the training of the FFBP-ANN took approximately 17 days, using a high-performance computer. -Naive Bayes (NB): This is a simple probabilistic classifier based on Bayesian law. In this work, the NB model is trained to predict the next BS.
• Stage 3: ML Model Validation: The proposed ML classifiers were tested using the testing sample. The performance of the trained ML models is described in detail in section V-A. The KNN classifier outperforms the FFBP-ANN and NB classifiers in terms of prediction efficiency. Therefore, the KNN trained model is selected in this work to be the BS prediction tool.
• Stage 4: ML Model Deployment: ML model deployment is the process of installing the trained ML model on the SDN controller. The SDN controller is the core component of an SDN-based vehicular network, and is physically connected to network elements, including wireless base stations [43], [44]. It is responsible for performing the adaptive cell selection task based on the installed ML model.

B. FRAMEWORK OF THE PROPOSED SDN/ML-BASED CELL SELECTION APPROACH
The framework is based on the combined SDN and ML approach, as shown in figure 6. SDN is applied to perform efficient traffic management and cell selection tasks, based on the installed trained machine learning model. The inputs of the ML-model are vehicle information and network configuration information. The geographical coordinates of vehicles (i.e., latitude (LAT ) and longitude (LON )), azimuth between vehicle direction and the north (AZIM U T H), and vehicle speed in km/h (KSP EED) are given to the ML model. In addition, the predefined BS's load, vehicle's speed, RSSI, and cell radius thresholds are also inserted into the ML model. Based on the given information, the ML model can predict the optimal BS to be associated with, whether it is a macro-or small BS. Algorithm 2 gives the pseudocode for the proposed A2T-KNN scheme.

IV. SYSTEM MODEL A. SIMULATION TOOL
The MATLAB 2021a simulator was used for modeling and analyzing the performance of the proposed cell selection scheme due to its powerful capabilities. In addition, THE MATLAB simulator has many toolboxes that can be used to perform the cell selection task in a realistic environment. Figure 7 shows the main MATLAB toolboxes that are installed to perform simulation experiments.    Table 3. Each driving route has a set of geographical points. Using the MATLAB simulator, the following operations were performed:

B. DATASETS
-Generating extra geographical points for each street, based on a predefined distance, which is 1 meter in this work. -Calculating the azimuth for each point according to the street on which the point is located. The azimuth is the angle between north and the street direction, beginning from the located point.   The description of the vehicle dataset fields is given in Table 3. Figure 10 shows a snapshot of the generated vehicle dataset. VOLUME 15, 2021

C. NETWORK MODEL
A two-tier heterogeneous ultra-dense network is considered that is comprised of macro-and small BSs.

D. PROPAGATION MODEL
In this paper, the propagation channel model comprises three main components: path-loss (PL), fading and shadowing. These are the main losses that affect the strength and quality of wireless signals. The 3rd Generation Partnership Project (3GPP) path loss models are used to estimate the received signal strength at a given distance from the serving base station. Table 4 shows the used 3GPP path loss models that are defined in 3GPP technical report (TR) 38.901 version 16.1.0 [47]. As shown in the table, the macro-BSs tier uses the urban macro-cell-non-line-of-sight (UMa-NLOS) PL model, while the small BSs tier uses urban microcellline-of-sight (UMi-LOS) (street canyon) model. The carrier frequency is denoted by f c and it is measured in Gigahertz. The distance between a base station and a vehicle is represented d and it is measured in meters. The definition of break-point distance (d BP ) is given in equation (1).
The heights of a BS and a vehicle are expressed as h BS and h V eh , respectively. The vehicles' height must be between 1.5 and 22.5 meters to apply the 3GPP PL models. The effective height between vehicles and BSs is indicated by h Ef f and it is described in detail in the technical report. The symbol c represents the speed of light in a vacuum, which equals 2.997 x 10 8 ms −1 . Rayleigh fading is considered in our study because it is a good approximation of realistic conditions of a wireless channel. It follows an independent exponential distribution with unit mean [48]. In addition, log-normal shadowing is considered because it is typically used to model the relationship between RSSI and range [49].

V. PERFORMANCE ANALYSIS
In this section, the trained ML models are evaluated to determine how well they classify input data that the models were not trained on. Moreover, the proposed A2T-KNN algorithm is evaluated in terms of the average number of handovers, staying time, number of HO failures and unnecessary HOs, downlink sum-rate, and energy efficiency. Table 5 shows the number of training and testing samples that are used in this work. Root mean square error (RMSE) and mean absolute error (MAE) are common criteria for measuring errors in prediction [27], [50]. The calculation of RMSE and MAE is based on the predicted base stations (ŷ),   [52]. Then, accuracy, sensitivity, specificity, precision, F-measure (F1), and geometric mean (G-mean) were estimated.

1) Key Performance Indicators
The key performance indicators (KPIs) that were used to evaluate the performance of the proposed A2T-KNN algorithm were the average number of handovers, staying time, number of HO failures and unnecessary HOs, downlink sum-rate, energy efficiency, radio link failures, and handover interruption time.
Handover is the process of switching the connection between network cells when a mobile device moves out of the range of the current serving cell [53]. There are two kinds of handovers: horizontal and vertical HOs. A horizontal HO occurs between homogenous base stations, while a vertical HO happens between heterogeneous BSs [54], [55]. Decreasing the number of handovers incurred by a vehicle is an important factor to maintain the network performance [56]. The number of HOs depends on the staying period of a vehicle within a serving cell, as the longer the stay time, the lower the number of HOs [57]. When the handover delay is longer than the staying time within a cell, a handover failure happens. When the sum of HO delays to move into (t i ) and out (t o ) of the network cell is longer than the staying time in a cell, an unnecessary handover occurs [12], [58]. Equations (2) and (3) show the formulas for the probability of a HO failure and unnecessary HO, respectively.
Where vehicle velocity and cell radius are denoted by v and R. The time thresholds of a HO failure and unnecessary HO are represented by T h f and T h u . They can be calculated according to equations (4) and (5), where the acceptable values of P r f and P r u are 0.02 and 0.04.
)− 2 π P r u ) ; 0 P r u 1 (5) The average number of HO failures (N f ) and unnecessary HO (N u ) can be estimated as shown in equations (6) and (7).
Small cell densification is a promising solution that can be used to fulfil the 5G network requirements of network capacity and throughput [59]. The total aggregate throughput (sum rate) is the summation of the achievable data rate across the network when a vehicle moves [41], as equation (8) shows. Channel capacity C ij between base station B i and vehicle V j can be calculated based on the Shannon theorem as given in equation (9).
The signal-to-interference-plus-noise ratio is defined as the power received from the serving BS divided by the summation of the power received from other BSs plus noise [60]. VOLUME 15, 2021 It is represented by ψ ij and its formula is given in equation (10).
where p tx is the maximum transmitting power of a base station, ζ(d) is the path loss function, and g represents the channel gain between BS and vehicle. As previously discussed in section IV-D, the path loss function follows the 3GPP PL models and the channel gain considers the impacts of Rayleigh fading and log-normal shadowing. The noise power spectral density and sub-channel bandwidth are denoted by N 0 and BW , respectively. They are used to estimate the thermal noise level, based on an additive white Gaussian noise model.
The energy efficiency (EE) of the system is a critical evaluation criterion and it refers to the ratio between the sum of the achievable data rate and the total consumed power [61]. Equation (11) gives the formula for the energy efficiency, which is denoted by η EE .
η EE (bits/joule) = Sum of achievable rate (bps) T otal consumed power (W att) The quality of the radio link is modeled in the term signalto-interference-plus-noise ratio. Radio link failure occurs when the value of SINR of a vehicle V j from the serving cell B j falls below the out-of-synchronization threshold (SY N out ) for a Radio Link Failure (RLF) detection period, which is known as T RLF . If T RLF timer, which is also known as T310, has expired and the SINR value does not increase above the in-synchronization threshold (SY N in ), the vehicle faces a problem of RLF. Equations (12) and (13) shows the conditions of failure and recovery of radio link [62], [63].
Handover interruption time (HIT) is an essential metric to evaluate the performance of cell selection schemes. HIT is defined as the duration in which the vehicle's connectivity is interrupted to perform the handover operation [64]. Equation (14) gives the formula of T HIT which is the summation of break time (T Break ), processing time (T P roc ), interruption time (T Interrupt ), radio access channel time (T RACH ) and handover completion time (T HC ) [65].
T HIT = T Break +T P roc +T Interrupt +T RACH +T HC (14)

2) Simulation Results
In this section, the simulation results are presented and discussed. The performance of the proposed A2T-KNN is compared with the traditional max-RSSI, HO RTP [42], and Zappone et al. ANN-based [34] schemes. Table 7 displays the simulation parameters used to evaluate the cell selection schemes.   68% and 45.83%, respectively, when the speed threshold is not exceeded. The HO RTP scheme selects the nearest small BS that has a residence time greater than a specific time threshold; as a result, the proposed A2T-KNN outperforms the HO RTP scheme in terms of the average staying time and number of HOs by 35.12% and 38.1%, respectively. This is because the HO RTP considers the residence period within the small cell, but it gives the highest priority in selection to the RSSI value. Moreover, the proposed AT2-KNN achieves additional enhancements regarding the average staying time and number of handovers with vehicles that exceed the speed threshold, due to the adaptation characteristic of the proposed algorithm. under a specific speed (10 km/h in this example) is represented in Figure 16. As the figure shows, the proposed A2T-KNN protocol reaches peaks in the data rates, while the max-RSSI, Zappone Figure 17 illustrates the CDF of network energy efficiency for a vehicle during the simulation time. As the energy efficiency is the achievable sum data rate divided by total consumed power, the A2T-KNN algorithm is superior to the other cell selection strategies because it outperforms them in terms of the total achieved downlink data rate, as illustrated in figure 16. The percentage of improvement in terms of average network energy efficiency achieved by vehicles is 13.99% compared with the conventional max-RSSI and Zappone et al. ANN-based methods, while A2T-KNN outperforms the HO RTP method by 16.86%.  Figure 18 shows the CDF of received SINR values by a vehicle during simulation time. According to the result, downlink SINR improves by applying our A2T-KNN scheme because it achieves high SINR values that are not reached using the other cell selection methods. We found that RLFs occur when SINR values drop below SY N out for T RLF , which does not exceed 1.67% using our A2T-KNN scheme and the other cell selection scheme. The reason for this is that all simulation experiments depend on setting an RSSI threshold and if the received RSSI value is less than this threshold, the cell selection algorithm is executed. In fact, radio link failures can be avoided by enhancing the SINR values by applying interference mitigation techniques. In addition, relying on soft handover, which means connecting to the next BS before breaking the old one, helps in reducing the RLF rate. Figure 19 shows the relation between vehicle speed and average cumulative handover interruption time. As the figure shows, our proposed A2T-KNN scheme outperforms other methods because it aims to prolong the staying time of vehicles within serving cells and therefore decreases the cumulative HIT. In addition, the proposed scheme applies switching between the small BS tier and macro BS tier if the vehicle speed exceeds the predefined speed threshold to avoid frequent HOs. We find that the worst cell selection methods in terms of cumulative HIT are the Max-RSSI and Zappone et al. ANN-based schemes. The reason is that they depend on the strongest received RSSI value when they select the next BS, and therefore the number of HOs increases and the cumulative HIT gets worse. Our protocol outperforms them in terms of cumulative HIT by 45.83%. The HO RTP method achieves better performance in terms of cumulative HIT compared with the max-RSSI and Zappone et al. ANN-based schemes because it sets a residence time threshold, but it gives high priority to the cell that has the strongest signal power. Consequently, the A2T-KNN has superiority over the HO RTP method by 38.1%.

VI. CONCLUSIONS AND FUTURE WORK
In this paper, we study the cell selection problem of HetNets by using SDN and machine learning technologies. In this paper,an SDN/ML-based adaptive cell selection scheme called A2T-KNN is proposed. It is designed for two-tier HetNets and it can adapt to the characteristics of the network and the features of moving vehicles. The proposed A2T-KNN algorithm uses a KNN model that are trained based on realistic vehicle and BS information in the central area of London. Simulation results show that the trained model has a high prediction performance using the testing sample. In addition, our proposed A2T-KNN algorithm is superior to the traditional and HO RTP schemes in terms of the average number of HOs by 45.83% and 38.10%, respectively. Moreover, it enhances the average achievable DL throughput and network energy efficiency achieved by vehicles by up to 17.18% and 16.86%, compared with the other methods. For future work, the A2T-KNN algorithm will be applied in other cities, such as the city of Riyadh, the capital of Saudi Arabia, based on the availability of macro and small BSs information. In addition, other machine learning algorithms can be applied and prediction performances can be compared. The cell selection issue in HetNets can be solved in subsequent studies based on reinforcement learning (RL) techniques, such as Q-learning and deep Q-networks (DQN). Licensed Assisted Access (LAA) deployment can be considered in our future work. It is a service that combines the use of licensed and unlicensed spectrum to enhance the achievable data rates and to improve the user response.

APPENDIX A: LISTS OF ABBREVIATIONS AND SYMBOLS
The following table gives a list of abbreviations. Tables 8  and 9 give lists of the main abbreviations and symbols. Signal to interference plus noise ratio between BS Bi and vehicle Vj ptx Transmission power of base station g The channel gain N0 Power spectral density of noise BW Channel bandwidth ηEE Energy efficiency of the system SY Nout Out-of-synchronization threshold SY Nin In-synchronization threshold TRLF Timer of radio link failure