Dynamic Interference Optimization in Cognitive Radio Networks for Rural and Suburban Areas

,


Introduction
The development of more efficient modulation schemes and forward error correction codes has allowed reducing the gap to the theoretical Shannon capacity limit. For instance, the introduction of LDPC (Low-Density Parity Check) codes allowed the last generation of broadcasting television technologies to get within 2 dB of the theoretical capacity limit (Shannon limit) [1]. The modulation and coding schemes (MCS) developed for ATSC 3.0 digital television standard are 1.3 dB to 1.8 dB from the Shannon limit [2]. This improvement leads to a higher broadcasting spectral efficiency. This allowed releasing hundreds of megahertz for wireless and mobile communications services, i.e., the 700 MHz band has been reassigned in several countries for mobile communications [3,4]. Despite the increasing improvements in the spectral efficiency of technologies, the demand for bandwidth exceeds the availability of spectrum for new communication services and networks. Paradoxically, several spectrum sur-veys demonstrate that the spatial and temporal use of the sub 3 GHz spectrum is less than 20% [5,6,7,8]. This is a direct consequence of the inefficient allocation and assignment of the spectrum. Hence, nowadays, the main problem is not related to technology efficiency itself but how to use it efficiently.
Cognitive radio paradigms over different regulatory constraints have been proposed as a long-term solution for tackling spectrum scarcity for broadband wireless services and the lack of spectrum usage efficiency. Different standards enabling cognitive radio technologies for spectrum allocation have emerged in recent years, e.g., IEEE 802.11af [9], IEEE 802.22b [10], and ECMA 392 [11]. The importance of cognitive radio technologies and dynamic spectrum access for enabling next-generation wireless networks has also been recognized by the European Telecommunications Standards Institute (ETSI) and the Third Generation Partnership Project (3GPP) [12]. ETSI technical report TR 103 067/2013 [13] analyzes the feasibility of Long-Term Evolution (LTE) Cognitive Radio Systems operating in UHF band TV white spaces (TVWS). This technical report also evaluates the coexistence between LTE networks operating on TVWS and television broadcasting services [13].
The large bandwidth requirement for 5G networks is only fully satisfied at the highest spectrum bands. As the cells' coverage is considerably reduced, the application cases for 5G networks are quite limited for small cities, rural scenarios, and in general, in areas with low population density. In [14], the authors investigated the feasibility of 5G network operation based on cognitive radio and carrier aggregation for enabling wideband access in rural areas. In [15], a multilayer routing strategy for enabling 5G services exploiting TVWS is presented. They guarantee a framework for coexistence with heterogeneous networks sharing the same spectrum bands. On the other hand, the Ericsson 5G platform will utilize dynamic spectrum sharing techniques for granting spectrum access to either 4G or 5G user devices [16].
Optimizing spectrum usage efficiency without affecting the users' Quality of Service (QoS) and Quality of Experience (QoE) is critical for the future development of wireless communications operating in TVWSs. The novelty of this paper is a dynamic interference management algorithm for reducing harmful interference to broadcasting services. The algorithm is designed for operating in a centralized spectrum management architecture, rather than the traditional distributed architecture. The designed algorithm seeks a higher spectrum usage efficiency and is validated by measurements in an emulation test with self-developed testbed hardware.
For evaluating the results achieved by our algorithm, we emulate the critical conditions in terms of interference to the broadcasting station. For the first time, we assess the degradation of the broadcasting service Quality of Experience caused by secondary cognitive radio interference. We also compare the interference in a traditional cognitive radio network and the proposed optimized networks. In this research, we consider a Long-Term Evolution-Advanced (LTE-A) network operating on TVWS under the ETSI regulatory conditions for a rural scenario and suburban scenario in Havana, Cuba, and Ghent, Belgium, respectively.
The outline of this paper is as follows. In Section 2, we present the most relevant state-of-the-art related to cognitive radio technologies, in Section 3 we present the method for the dual-objective dynamic network optimization and for the experimental evaluation of the QoE in the broadcasting networks for different interference profiles generated by the optimization software. In Section 4, we present the cognitive LTE-A network optimization results and perceived degradation reduction for the broadcasting users in terms of QoE. In Section 5, the conclusions of our research are formulated.

State of the Art
Interference to and from the primary licensed service has been a major concern related to cognitive radio technologies, although sensing techniques for allocating empty channels have been improved throughout recent years [17,18,19].

Protection Limits.
The protection limits, cochannel, and adjacent channel interference between cognitive radio and digital television technologies have been widely investigated. In [20], the authors studied the coexistence between the IEEE 802.22 cognitive radio standard and the DVB-T2 digital broadcasting. The IEEE 802.22 standard itself defined the maximum signal level threshold of ATSC for reusing the spectrum by secondary services without causing harmful interference to the primary ATSC service [21]. However, the IEEE standards (latest update) do not provide protection ratios for other digital television broadcasting services [10]. Further, the FCC and ETSI regulation defined the maximum radiated signal level for minimizing the interference among primary and secondary services [13,22].
Laboratory measurements for assessing the protection of DVB-T2 broadcasting services interfered by IEEE 802.22 Wireless Regional Area Network (WRAN) transmissions in the cochannel and adjacent channels were performed, and protection ratios were defined [20]. Further research is required for studying the maximum device density and distribution [20]. For unspecified transmissions in the IEEE 802.22 standard, the suggested protection limit is -90 dBm. However, this value is quite high for most propagation environments and broadcasting transmitters for Digital Terrestrial Multimedia Broadcast (DTMB) and Digital Video Broadcasting-Terrestrial (DVB-T/T2). For the average DVB-T transmitter configuration, an additional extension of the protected broadcasting transmitter area by 5 to 15 km is recommended. The extended protection is estimated based on the link budget for the digital broadcasting transmitters and the interference protection margin for DVB-T2 and DTMB presented in [20] and [23], respectively. A more appropriate spectrum reusage threshold can be set at -95 dBm for both DVB-T2 and DTMB use cases. Nevertheless, these values do not guarantee that the channel is useful for providing the desired network QoS and QoE for the secondary wireless communication service. For the primary licensed service, the protection margin should be analyzed on a case-by-case basis as the transmitters' setups and propagation conditions significantly vary.
For one of the highest modulations and coding schemes for IEEE 802.22 (64-QAM FEC =3/4), a minimum Signalto-Noise-Ratio (SNR) ratio of 19.7 dB is required for achieving at least 95% of the maximum throughput. The sensitivity of commercial receivers for this MCS is -82.5 dBm. At the broadcasting edge of the protected contour with a mean signal level of -95 dBm, the SNR is 12.5 dB. According to the experimental results presented in [24], this increases the bit error rate beyond the maximum allowable threshold of 2 × 10 −4 . For the worst-case coexistence scenario, being a BS located around such an area, its coverage might be reduced to approximately a third of its maximum radius (6.8 km of 20 km) [24].
The usage of these channels by the secondary wireless devices located on the border of the protected contour of the broadcasting transmitter might require dynamic management of the MCS and interference for satisfying the QoS requirements of all network users while keeping a high spectrum usage efficiency.  [25], and LTE-A cognitive (Frequency Division Duplexing (FDD)) [13].
In [26], authors investigate the coexistence of multiple heterogeneous secondary networks, independently operating in TVWS without a cooperation mechanism. Although interference among heterogeneous networks is also common for other services, e.g., Wi-Fi, Bluetooth, the better propagation conditions in the UHF band, and deployment of regional WRANs increase the complexity of the interference avoidance problem. In [26], authors propose an algorithm for maximizing the desired secondary network throughput by optimally selecting the spectrum channels in the presence of interference from the primary licensed service and other secondary wireless networks. However, interference to other networks is not solved as no cooperation, or priority mechanism is established. According to [27], IEEE 802.22 networks tend to allocate more and better spectrum resources than IEEE 802.11af operating in a common scenario. Our previous research also found slightly better performance for IEEE 802.22 networks [28]. In [27], the authors proposed a centralized cooperative mechanism for allocating spectrum for heterogeneous networks operating on TVWS, considering IEEE 802.11af and IEEE 802.22 networks.
Geolocation databases might lead to lower interference between the primary and secondary services and also might work as a common spectrum distribution mechanism for heterogeneous networks, but with the drawback of a lower spectrum usage efficiency. This is due to the fact that information in the databases is not updated in real-time [28].

Algorithms for Channel Selection and Interference
Mitigation. In [29], authors present a hierarchically based mechanism for granting fairer spectrum access, based on the kind of wireless service and a reputation-based score. Users that access the spectrum in a selfish way receive a lower score and further lower priority by the centralized allocation system [29]. In [30], a channel selection algorithm for maximizing the users' QoS is proposed. However, it is assumed that the channel needs to be configured for maximum capacity transmission. Our research in [28] demonstrates that the highest MCS is not always the optimal solution as it depends on several factors, including the propagation environment, network load, and the ratio between served traffic and network coverage.
Most researches on coexistence and interference in TVWS consider the signal threshold for reusing a specific channel as a fixed value, for interference from the primary service or another cognitive BS or network. However, a better trade-off between interference and spectrum usage could be achieved if a flexible range of interference allowance is con-sidered, and as such, the interference is dynamically managed. For instance, assigning to a specific user(s), a channel with higher interference might result in a better spectrum usage efficiency and lower mean interference when assessing both Key Performance Indicators (KPIs) for the whole network. For this implementation, centralized access control is required, rather than a distributed spectrum allocation system. This is because the data reported by all users and BSs need to be assessed for finding an optimal configuration for each network device and the whole network performance.

Method
For minimizing the interference between the primary (broadcasting) and secondary (LTE-A) services and reducing the spectrum usage by the cognitive radio network, we developed a dynamic optimization algorithm for these Key Performance Indicators (KPIs). Further, the software optimization results need to be assessed in a real scenario. In this section, we present the dynamic optimization algorithm for cognitive radio-based networks and the methodology for its experimental validation. We emulate the coexistence scenario and perform measurements for quantifying the impact on the broadcasting service QoE. Figure 1 shows a chart illustrating the optimization and validation process.
First, we define a rural and suburban scenario where a broadcasting primary licensed service (DTMB or DVB-T2) coexists with a secondary wireless communication service based on cognitive radio (LTE-A). Here, the service's demands, setup, technological specifications and configurations, and the propagation models are defined (Figure 1(a)). These are the inputs for the LTE-A network dynamic optimization by our multiobjective heuristic algorithm (Figure 1(b)). 'The dynamic optimization of the user's connections to the LTE-A BSs reduces either the spectrum usage, the interference (i.e., LTE-A to broadcasting), or both KPIs. The optimization process generates an interference profile containing the average spatial interference among all transmitters in the whole area. Based on the maximum allowable interference produced by a traditional cognitive radio network for each scenario and user distribution, the optimization gain (dB) in terms of interference mitigation can be calculated. This is the input for an experimental emulation in a real wireless scenario by means of self-developed testbed hardware (see Figure 1(c)). In the context of this research, the emulation is defined as an experimental model where some functional part of it is carried out by a part of the real system [31]. Here, the channel power of a real on-the-air LTE-A signal is reduced in real-time according to the optimization gain at a certain timestamp (obtained from the software dynamic network modelling). For the emulation, we investigate the worst-case scenario at the coverage edge of the broadcasting transmitters in our scenarios (most vulnerable broadcasting signal). The impact of the LTE-A network interference to the broadcasting service QoE is investigated by means of subjective experimental measurements (in Figure 1(d)). In the following subsections, we describe each process (Figures 1(a)-1(d)) in detail.
In the following subsections, each step of the optimization and experimental evaluation process is described. In 3 Wireless Communications and Mobile Computing Subsection 3.1, we detail the characteristics of the scenario, and in Subsection 3.2, the link budget and considered path loss models for propagation calculations (see Figure 1(a)). In Subsection 3.3, the algorithm for the LTE-A dynamic network optimization (see Figure 1(b)) is described. Finally, in Subsection 3.3.1, the broadcasting service experimental interference assessment is explained. This subsection includes the details of the testbed hardware, the emulation conditions, and the QoE assessment (Figures 1(c) and 1(d)).
3.1. Scenario Modelling. We considered two typical scenarios for modelling the interference between the broadcasting and LTE-A network operating on TVWS. Figure 2(a) shows a rural scenario in the outskirts of Havana, Cuba, where a Multifrequency Network (MFN) operates using the DTMB broadcasting transmission standard [32]. Figure 2(b) shows a suburban scenario in Ghent, Belgium, where different Single-Frequency Networks (SFN) operate according to the DVB-T2 broadcasting transmission standard [33].
In the rural area where the LTE-A network provides services (delimited by the blue line in Figure 2(a)), the average population density is approximately 200 people per square kilometer. 10 LTE-A BSs (marked with red dots in Figure 2(a)) serve the traffic generated at the peak time being, on average 135 Mbps. In the suburban scenario in Ghent (delimited by the blue line in Figure 2(b)), the population density is 1700 people per square kilometer, and the peak traffic generated is, on average, 224 Mbps, being served by a total of 27 BSs.
All the broadcasting transmitters in and surrounding the areas where the LTE-A cognitive networks operate (black icons in Figures 2(a) and 2(b)) are included for assessing the interference and dynamically managing the spectrum to be used by the LTE-A network. There are 16 broadcasting towers using 16 different frequencies in the suburban scenario and 15 broadcasting towers using 11 different frequencies in the rural area. In the rural scenario, there is at most a single standard definition stream and a single high definition stream per transmission tower. In the suburban scenario, there are up to 5 services per tower, leading to a higher density of spectrum usage. Table 1 lists the specifications of the DTMB, the DVB-T2 broadcasting transmitters, and the LTE-A BSs.

Link Budget and Path Loss Models.
The LTE-A EIRP corresponds with the maximum radiating power allowed by most regulatory domains for the operation of secondary cognitive radio devices in TV white spaces [13,22]. For Ghent LTE-A network planning, we consider an 8 MHz channel bandwidth based on the carrier aggregation mode. We include the aggregation of a 3 MHz and a 5 MHz channel. Hence, the OFDM specifications correspond with the equivalent for the aggregated carriers in this channel configuration. For the Havana LTE-A network, a channel of 5 MHz is considered. The OFDM carriers specifications for LTE-A are the ones defined in [34], for DTMB in [32], and for DVB-T2 in [33]. The frequencies of each LTE-A BS are dynamically assigned within the specified range by the network optimization algorithm.
The DTMB and DVB-T2 specifications correspond with the technological settings for the actual broadcasting transmissions in the areas under test. The current frequency of each broadcasting transmitter in the region is considered for the propagation and interference calculations. Noise figure and sensitivity correspond with the testbed hardware specifications for such settings (details on the testbed hardware are provided in Section 3.4.4). Notice that shadow margin and fade margin for DTMB and DVB-T2 are higher because we include the broadcasting transmitters in the surrounding of the areas under test with different propagation conditions. The protection margins are assumed according to the recommendations in TS 136 104 Release 12: LTE; Evolved Universal Terrestrial Radio Access (E-UTRA); base station (BS) radio transmission and reception [35]. For assessing the interference to and from the broadcasting transmitters, all the transmitters in the area and surrounding the LTE-A cognitive network are considered up to a distance of approximately 120 km. First, the path loss for accounting for the propagation losses experienced by the signal in the environment has to be quantified. Equation (1) defines a generic function for accounting for the path loss in any environment [36].
where PL (dB) is the path loss, f (MHz) is the transmission frequency, d (km) is the distance between the transmitter (tx) and the receiver (rx), and the transmitter and receiver heights are denoted by h. The function gð:Þ will depend on the propagation model that fits the propagation characteristics of the investigated scenario [36]. For defining gð:Þ, different propagation models have been proposed in the literature and further standardized. These models account for environmental factors, e.g., the terrain topology, the ground characteristics, and obstacles. For the rural scenario in the Havana outskirts, we consider the Okumura-Hata path loss model for rural environments [37], as this model showed a better fit than the ITU-R P.1546 [38] model according to a largescale measurement campaign in the rural area of Havana [39]. The ITU-R model predicts much lower propagation losses over near-flat terrain without significant obstacles compared to the actual measurements. Notice that most of the land use in our rural scenario consists of farm lands and small community houses. For the suburban scenario in Ghent, an empirical one-slope path loss model was determined based on a large-scale measurement campaign in Ghent [40]. This model was designed based on a large-scale measurement campaign in Ghent city for a reference frequency of 602 MHz. For reducing the error when accounting for other frequencies f (MHz), we apply a correction factor ε f as defined in

Wireless Communications and Mobile Computing
The Ghent model cannot be used in its surroundings from 10 km to 120 km. This is because the Ghent model is valid up to 10 km. Hence, in our application, within Ghent, the oneslope model is used, and for the broadcasting transmitters outside the Ghent area, the ITU-R P.1546 model is used.
Once the path loss is known, the signal level at any given location for all radiating devices can be quantified, taking into account the specified technological specifications of the transmitters and additional environmental factors. This allows for accounting for the interference margin among services using the same frequency, i.e., carrier signal to interference signal ratio C/I (dB). The coverage of a particular service (coverage radius) R (km) is defined by the maximum allowable path loss PL max (dB), the environmental propagation function, and a certain margin for accounting for the signal fading F d (dB) and shadowing S w [36]. Equation (3) defines the coverage radius as a function of the maximum allowable path loss.
Notice that the maximum allowable path loss depends on the technology specifications, including the transmitters' and receivers' setup. The link budget defining the technologies' specifications and configuration for accounting for PL max is listed in Table 1.

Cognitive LTE-A Network Optimization in TV White
Spaces. For the cognitive wireless network dynamic optimization, we base ourselves on the heuristic capacity-based opti-mization algorithm GRAND (Green Radio and Access Network Design) [41] with the modifications for cognitive radio-based technologies described in [28]. The algorithm takes a set of users and traffic densities as input parameters and manages the user connections to the BSs, the radiated power, and the BSs' operational modes (idle/active) for optimizing a certain network parameter, e.g., network power consumption [41] [28]. However, the algorithm does not optimize the spectrum usage taking into account a dynamic interference assessment on a case-by-case basis. The KPIs considered here are network interference and spectrum usage.
In laboratory experiments for LTE-A BSs of three different service providers, we have observed that the LTE-A BSs can maintain the users' mean throughput degradation below 25% by dynamically managing the MCS and the PRBs (Physical Resource Blocks) among different users connected to the same cell. For both DTMB and DVB-T2 interference to the LTE-A network, the interference can be coped within a range of 3 to 5 dB. A throughput degradation of 25% corresponds with a Data Block Error Rate (DBER) lower than 10%, for which the experimental interference assessment in LTE networks is recommended in ITU-R BT.2215-7 Report [42]. The interference management depends mainly on the mobile network operator connection management and the signaling time (update time). By dynamically managing the interference in our optimization algorithm in a range of 5 dB from the maximum allowable interference, it is possible to achieve a better network spectrum usage and lower average network interference without significantly affecting the QoS of the wireless network. Algorithm 1 describes the dynamic optimization algorithm for the cognitive radio LTE-A network. This algorithm allows minimizing the interference between services and achieving a higher spectrum usage efficiency, without affecting the wireless network QoS. By reducing the interference to the broadcasting network, the TV service users' QoE is improved. The algorithm is heuristic and capacity based. The inputs are the user density, the traffic demand, the cognitive LTE-A, and the TV Broadcasting station specifications (see Table 1) and its geolocations (see Figure 2).
First, based on the input parameters and constraints, the algorithm generates a green-field network planning with the initial settings and locations of the BSs and TV stations. The users are randomly and uniformly distributed over the whole area (line 1 in Algorithm 1). This means each pixel in the map has the same probability of having a user requesting a connection to the network.
The algorithm runs for a total number of simulations (Max_Sim in line 2, Algorithm 1). The number of simulations is defined by the progressive average of the network KPIs, i.e., average network interference and spectrum usage.
Here, we consider that if the progressive average of any KPI has a standard deviation higher than 5%, the number of sim-ulations is not high enough. This is correlated with the proportion of the number of users and the area size and the random nature of the user allocations. The lower the number of users per square kilometer, the more diverse the possible distributions of those users, and the higher the number of possible solutions. As the algorithm is heuristic, the best solution is not found but a solution good enough for satisfying the defined constraints. Hence, for a higher degree of freedom and possible solutions, a higher number of simulations are required.
For each user u i in the area, the algorithm tries to find the connection to the cognitive LTE-A base station BS j that best suits the network KPI optimization goals. First, for BS j 's maximum coverage range, the algorithm seeks if the u i is located in its range. In such a case, the algorithm will analyze for each channel the interference from all other LTE-A BSs and TV broadcasting stations in the area. The interference signal level (ISL) threshold will be dynamically moved in a range from ISL min to ISL max with a resolution of 1 dB. The ISL max is defined in Table 1, Section 3.2, and ISL min is 5 dB lower. The spectrum allocation function will reduce the spectrum usage for each ISL constraint (lines 6 to 9 in Algorithm 1). The interference from all broadcasting stations and LTE-A devices at u i and BS j geolocation is calculated by the path loss models defined in Section 3.2. If the interference is Input: users, traffic, BSs, TV; 1: Generate (Green-Field_Network); 2: while Sim < Max_Sim 3: for u i < users 4: for BS j < BSs 5: Set_Radiated_Power (u i ,BS j ); 6: for ISL min : 1 dB : ISL max ); 7: Evaluate_Interference (u i ,BS j ); 8: Set_Spectrum_Allocation (u i ,BS j ); 9: end for; 10: if fit (u i ,BS j ) > current_fit 11: if Active (BS j ) && Bitrate (BS j ) 12: connect  7 Wireless Communications and Mobile Computing lower than the current ISL constraint, the connection at the channel being analyzed can be stablished. Otherwise, a new channel is analyzed. Lower frequencies in the spectrum are tried first because of their lower path loss attenuation, the radiated power can be decreased, and as a consequence, the interference and the network power consumption are reduced.
3.3.2. KPI Optimization within the Algorithm. The final connection of u i to BS j among all j possibilities is determined by a fitness function that balances both network KPIs: interference and spectrum usage (line 10 in Algorithm 1). The network fitness function is defined by where ISL A is the interference caused by the LTE-A BSs and ISL TV is the interference caused by the TV broadcasting stations. These values are normalized to the maximum interference for the TV broadcasting transmitters (ISL TVmax ) and the LTE-A BSs (ISL Amax ). The spectrum usage S U is the number of channels being required in the area of the u i ; BS j connection is settled and is normalized by the maximum number of channels in the band S max . The allowed interference has an inverse relationship with the spectrum usage. For a lower interference threshold, a higher number of channels will be used by the network. Hence, the weights w 1 and w 2 are used for balancing the relevance of both KPIs for the final network optimization solution. Here, we analyze three cases of weight combinations, i.e., ðw 1 ; w 2 Þ = fð1 ; 0Þ, ð0:5 ; 0:5Þ, ð0 ; 1Þg. Each normalized parameter in the fitness function responds to the average network performance. Hence, the user u i will be connected to the BS j and in the channel that best fits the network KPI optimization goals rather than maximizing the user KPI or QoS itself. Figure 3 shows an example of a user connection (u x ) to the best solution from the entire network's point-of-view, instead of the user's point-of-view. In this example, the connection of u x to BS 1 (operating frequency f 1 ) maximizes its effective throughput by using 64-QAM. However, in terms of interference to the broadcasting service, this is not the best network solution. The algorithm might choose the BS 2 (operating frequency f 2 ) for connecting u x , depending on the interference weight in the fitness function. Hence, it could happen that a user is connected to a channel that is not the best for this particular user, but by doing so, the overall network interference is reduced, or other channels can be reused by other devices, leading to higher overall network performance. The established connection must satisfy the user traffic demand and a minimum QoS threshold.
For satisfying the QoS requirement, the allowed interference is limited. The interference is low enough for allowing the users to satisfy their traffic demand with some of the available MCS (e.g., QPSK in the example of Figure 3). In addition, a minimum number of LTE-A BSs are activated depending on the network load and coverage requirements (lines 11 and 12 in Algorithm 1) [28]. In this way, other relevant network parameters like power consumption can be minimized. Nevertheless, if one of the active BSs cannot serve the user traffic demand, because of capacity, coverage, or interference constraints, a new BS is activated. The algorithm will seek the BS with the highest fitness and will try to balance the network load by moving users to different BSs if this allows the current user connection, and the network fitness is not decreased (lines 15 to 19 in Algorithm 1).
A further reduction in the radiated power will lead to a lower interference, as long as the connection is maintained (line 22 to 26 in Algorithm 1), i.e., the experienced path loss to the BS is lower than the maximum allowable path loss (PL < PL max , see Section 3.2). Also, the reduction of the radiated power and number of active BSs causes a reduction in network power consumption.
Finally, an optimal network solution for each simulation is generated. For each simulation, the software retrieves the spectrum usage distribution and an interference profile, including the achieved optimization gain G O (dB) for this KPI. The optimization gain is calculated by means of the following equation: Here, ISL 50 is the mean interference signal level in the whole area for a certain frequency. For the analysis of the results and for performing the experimental emulation (details in the following subsections), we consider the channels with the highest interference (worst-case).

Broadcasting Service Experimental Interference
Assessment. Figure 4 describes the interference conditions that we want to emulate in a real area with certain controlled variables, i.e., guarantee the desired broadcasting signal level at least 99% of the time. Here, our goal is to investigate the critical conditions for which a certain frequency f 1 can be reused, the caused interference and its impact on the  Wireless Communications and Mobile Computing broadcasting users QoE. We also investigated the impact of the G O for reducing the harmful interference to the broadcasting network receivers. This model is based on the average emission footprints of the local broadcasting stations, i.e., average height, radiated power, and propagation conditions. Here, our goal is to emulate the critical conditions (worst-case) for the broadcasting network being interfered with by the cognitive LTE-A network. Hence, the subjective test is performed for the same conditions at the coverage edge of the broadcasting station (see the solid line and marker subjective QoE test in Figure 4). The protection area of the broadcasting transmitter will increase depending on the achieved optimization gain for the interference KPI (see Equation (5) in Section 3.3.2). The worst-case corresponding with 0 dB optimization (at ISL = −95 dBm) is denoted with a red dashed line in Figure 4). The higher the optimization gain for the interference KPI, the higher the protection area radio and the higher the path loss from the LTE-A device to the edge of the broadcasting transmitter (see dotted lines in Figure 4). Hence, a lower interference from the LTE-A devices to the broadcasting network is achieved. In Sections 3.4.2 and 3.4.3, we provide details about the specifications for the suburban and rural scenario models, respectively.
The goal of the subjective QoE experiments is to quantify the perceived quality of the broadcasting service interfered with by an LTE-A cochannel signal. We evaluate the perceived quality of the broadcasting service interfered with by the actual LTE-A signal. We compare the broadcasting video degradation caused by the actual LTE-A on-the-air-signal with a channel power equivalent to the network operating with a fixed ISL threshold versus optimized with the dynamic interference optimization (interference optimization gain, see the algorithm output in Section 3.3).

QoE Subjective Indicators.
Two leading subjective indicators have been proposed in the literature for quantifying the threshold for which the objective technical parameters are degraded enough to affect the Quality of Experience of broadcasting service users. Report ITU-R BT.2035-2/2008 defines the Threshold of Visibility (TOV) as part of the guidelines and techniques for the evaluation of digital terrestrial television broadcasting systems in laboratory environments [43]. Hence, the TOV is the threshold where a trained observer notices any kind of pixelation in the image in a time-lapse of 1 minute [43]. Recommendation ITU-R BT.1368-13/2017 defines the planning criteria, including protection ratios, for digital terrestrial television services in the VHF/UHF bands [44]. This recommendation defines the Subjective Failure Point (SFP) as a method for quantifying and defines protection ratio measurements in a unified way. Then, SFP is defined as the quality level, where no more than one error is visible in the picture for an average observation time of 20 seconds [44].
By accounting on how many timestamps the TOV and SFP are exceeded at the broadcasting coverage edge (solid line in Figure 4), we will compare the performance of the developed algorithm for spectrum optimization only, interference optimization only, and an equally balanced strategy for both KPIs considering a flexible range of interference allowance. In addition, we will compare these results with the case of spectrum management by a traditional cognitive radio BS, considering the highest allowable interference (-95 dBm).
Notice that both TOV and SFP have been defined as a criterion to find a limit for a just error-free picture at a TV screen for protection ratio measurements. In the context of our experiment, our goal is not to find the protection ratio limits, but how many times these limits are exceeded. If exceeding these limits causes such BER that more than one pixelation is noticed within the same timestamp, a single assertion is recorded.

Rural Scenario Specifications.
In the DTMB rural measurements, the mean broadcasting tower height above average terrain (height in Figure 4) is approximately 111 m. The average EIRP of the broadcasting transmitters (in Figure 4) is 65.4 dBm. These parameters are based on the current settings of the operating transmitters and were retrieved from the local service providers. The receiver's theoretical sensitivity for 6 MHz channel bandwidth is -84.5 dBm (this defines the solid line in Figure 4). In a large-scale laboratory measurement, the average sensitivity throughout the UHF band was -84 dBm, with a standard deviation of approximately 2 dB [23]. The edge of the broadcasting station is located around 52 km from the transmitter (solid line in Figure 4). This is the radius for which under the considered propagation conditions, it is guaranteed a signal level higher than -84 dBm in 90% of the locations 99% of the time (Figure 4, solid line). For the worst-case interference conditions, the protection contour of the broadcasting station is located 60 km from the coverage edge (dashed line in Figure 4). This value is obtained considering the highest allowable interference, i.e., the mean broadcasting signal level is -95 dBm (dashed line in Figure 4). The worst interference case corresponds with an LTE BS or CPE radiating the maximum allowable EIRP in the edge of the broadcasting station protection contour for G O = 0 dB (located in the dashed line in Figure 4). Hence, accounting for the proper path loss, for the worst-case, there is an LTE-A mean interference signal level of -103.7 dBm at the coverage edge of the broadcasting station with a fading variation of approximately 8 dB (see marker LTE-A mean interference level over the solid line in Figure 4).

Suburban Scenario Specifications.
For Ghent, we also considered the average footprint of the regional DVB-T2 broadcasting transmitters in the area. The average height of the transmitters considered is 112 m and average EIRP = 74 dBm. These parameters are also based on the current settings of the operating transmitters in the region. Here, we considered only the transmitters with an EIRP < 75 dBm and height above the average terrain below 200 m. For the transmitters with higher EIRP and/or higher height above the average terrain, the mean ISL from the LTE-A devices is lower than -129 dBm (LTE-A interference over the solid line in Figure 4). As a consequence, a critical harmful interference condition is never reached. For the DVB-T2 9 Wireless Communications and Mobile Computing modulation and coding scheme defined in the link budget, the minimum broadcasting signal level is -82.5 dBm. Here, we consider an additional margin of 0.5 dB, being the broadcasting signal level for the experiments -83 dBm (minimum broadcasting signal level, the solid line in Figure 4). The coverage radius of the broadcasting service is then just 17.9 km, and the protected contour is located another 12.5 km away from the coverage edge (for the interference worst-case, being the optimization gain equal to 0 dB). It is important to mention that the difference in coverage and the protected areas between the rural and the suburban scenario is not because of the technologies link budget differences but because of the propagation environment (different propagation models). Then, for the DVB-T2 worst-case (G O = 0 dB), the mean interference from the LTE-A cognitive network in the coverage edge of the broadcasting transmitter is -105.7 dBm for the critical conditions described for this scenario (LTE-A interference marker over the solid line in Figure 4).

Testbed
Hardware. We designed testbed hardware for emulating as close as possible to the reality the scenario described in Figure 4. Figure 5 shows a diagram of the designed testbed hardware for the experiments.
An on-the-air LTE-A signal is filtered and translated in frequency to the UHF band. A UHF band filter attenuates (>30 dB) the signals generated out of the band for avoiding causing interference to local radio services. The return loss is higher than 18 dB. A step attenuator is configured at a fixed attenuation of 19 dB or 34 dB for having an interference signal at the receiver with a mean level (50-percentile) of -104 dBm (mean LTE-A interference in Figure 4 and also see spectrum analyzer in Figure 5) for the rural scenario and -106 dBm for the suburban scenario. The signal channel power is measured and averaged over a period of 1 minute with a sampling rate higher than 26 Hz, as recommended in [45].
A field-programmable gate array (FPGA) (see Figure 5) is loaded with the attenuation values relative to the difference between the optimized interference solutions generated by GRAND at each time stamp and the maximum allowable interference in a traditional cognitive radio network (optimization gain). The FPGA transcodes this value to the equivalent attenuation settings in the testbed hardware. The maximum allowable interference at the edge of the protection contour of the TV broadcasting signal is -95 dBm (see dashed line in Figure 4). For instance, if the optimization algorithm reduces the interference on average 3.1 dB for a certain simulation, the FPGA will configure the attenuator for having a reduction of the channel power by 3 dB. The attenuator in our hardware has a resolution of 0.5 dB. Hence, any value will be approximated to the closest fraction of 0.5 dB. Then, in the rural scenario, the LTE-A mean interference signal level will be reduced from -104 dBm to -107 dBm. However, these are mean values. In the experimental testbed, the signal will be affected by the real fading during the measurements. Therefore, the equivalent interference signal level will be −107 dBm ± 8 dB for this specific optimization output.
The LTE-A interference signal is added by a directional combiner to a controlled digital television signal (DTMB or DVB-T2), generated by the R&S Broadcast Test Center (in Figure 5). Here, the broadcasting video signal is locally  generated to guarantee the desired signal level more than 99% of the time. This matches the defined coverage contour of the broadcasting transmitter (solid line in Figure 4). The R&S Broadcast Test Center configuration is listed in Table 2. For both scenarios, the broadcasting signal channel power is +0.5 dB above the threshold defined for the coverage edge of the transmitter (solid line in Figure 4). The LTE-A signal mean channel power is monitored in time windows of 10 minutes and averaged throughout 1 minute to avoid and detect any significant change in the propagation conditions or more substantial fading variations. This is performed to guarantee that the initial system conditions are maintained throughout the test.
For assessing the variations on QoE by means of the TOV, a trained observer will assert 1 if at least one pixelation or distortion is noticed in any timestamp of 1 minute. The asserted timestamps indicate the percentage of timestamps the TOV is exceeded. In a similar way for the SFP metric, a trained observer will assert 1 if at least one pixelation or distortion is noticed in any timestamp of 20 seconds. The asserted timestamps indicate the percentage of timestamps the SFP is exceeded. The observer is located at a recommended distance of 1.2 m to 1.8 m from the TV receiver (in Figure 5). The total observation time is 40 minutes divided into 4 sets of 10 minutes for avoiding visual fatigue, minimizing the probability of missing the occurrence of a noticeable picture distortion. Figure 6 shows the optimization results for the required spectrum for the broadcasting and LTE-A networks in each scenario and the equivalent average interference signal levels. These results correspond to the KPI weights defined in Section 3.3.

Dynamic Network Optimization Results.
The highest ISL from the broadcasting network is reached when the algorithm optimizes SU without taking into account the interference metric for deciding LTE-A user connections (markers 2a and 4a, in Figure 6). In the rural scenario, the highest average ISL from the broadcasting DTMB signal is -98.5 dBm (marker 2a in Figure 6). For the suburban scenario, the ISL of the broadcasting DVB-T2 signal is approximately 1.1 dB higher (marker 4a in Figure 6). This difference is due to the fact that in the rural scenario with a lower traffic per square kilometer, more spectrum can be reused despite the fact that a better propagation might increase the interference from the surrounding broadcasting towers. Indeed, for this optimization case, the spectrum usage in the rural scenario is, on average, approximately 1.7 channels lower than in the suburban scenario (markers 1a and 2a compared with markers 3a and 4a in Figure 6).
When the interference KPI weight is maximum (w 2 = 1), the algorithm leads to an average ISL from the broadcasting network of -100 dBm for both scenarios (markers 2c and 4c in Figure 6). However, due to the strictest interference constraint, the spectrum usage increases 2.3 channels on average in the rural scenario (see markers 2c and 2a in Figure 6) and 6.2 in the suburban scenario (see markers 4c and 4a in Figure 6). This considerably higher difference between both scenarios is due to a higher traffic density per square kilometer in the suburban region and a higher number of broadcasting services using different frequencies in the same region. This reduces the probability of reusing the same channels by the LTE-A BSs.
For an equal weight of both KPIs (w 1 = 0:5; w 2 = 0:5) in the optimization algorithm, the interference from the broadcasting network compared with the best optimization case for the ISL KPI is increased by less than 0.2 dB (marker 2b versus 2c in Figure 6) and 0.4 dB (marker 4b versus 4c in Figure 6), for the rural and suburban scenarios, respectively. However, the spectrum usage is reduced by 1.5 and 2.7 channels, respectively (average for the whole area). Similar relative results are obtained for the LTE-A network self-interference (markers 1 and 3 in Figure 6).
We considered, for all cases, an equal weight when assessing the interference from the broadcasting network and the   11 Wireless Communications and Mobile Computing interference from the LTE-A devices, for the spectrum assignment in the LTE-A wireless network. For applications where the protection to/from a specific service has a higher relevance, a different ratio can be assigned for the optimization algorithm. The difference in absolute values is due to the different protection ranges for each technology (see Table 1 in Section 3.2). Figure 7 shows the average interference reduction with respect to the maximum allowable interference for each scenario and weight combination at each simulation stamp. This is the GRAND algorithm optimization output. These datasets are transmitted to the FPGA for the experimental emulation (see Section 3.3.1).
In both scenarios, there is a variation between simulations depending on the users' distribution for a particular simulation. Depending on the user location relative to the BS and the EIRP, the interference will vary. The highest variation is 2 dB corresponding to the rural scenario. The standard deviation has an impact on the TOV and SFP emulation results. When the interference restriction becomes stricter (when w 2 = 1), the difference between simulations (different users' geolocations) is reduced: the connection decisions are based only on this KPI and the algorithm will lead to the lower limit of the interference threshold. This means the algorithm would likely connect more users at different frequencies for reducing the interference up to the minimum interference in the boundaries considered for the network optimization.

Experimental
Assessment of the Interference to the Primary Licensed Service. Figure 8 shows the QoE metrics from the emulation measurement results as a function of the spectrum usage modelled by the optimization algorithm for the DTMB rural scenario and the DVB-T2 suburban scenario.
In our experiments, we do not compare the performance between DTMB and DVB-T2 but the performance of the optimization algorithm in two different scenarios. In terms of sensitivity, there is just 1 dB difference between the chosen configuration for DTMB and DVB-T2. The difference in the used data rate is lower than 7.2%. Any difference in terms of interference is caused by the propagation conditions and broadcasting transmitters' radiation footprints, its relative geolocation, and the LTE-A traffic density per square kilometer in each scenario.
For the DTMB rural scenario (darker lines in Figure 8), the degradation of the perceived video quality at the edge of coverage of the broadcasting transmitter is similar for both metrics TOV (marker t# in Figure 8) and SFP (marker s# in Figure 8). The maximum recorded difference between both metrics for the rural scenario was 4.2% (see markers s1 and t1 in Figure 8). In the DVB-T2 suburban scenario, the maximum difference between SFP and TOV was~10% (see markers t6 and s6 in Figure 8). This was caused by more spaced burst errors. As the SFP stamps are shorter than the TOV, the number of timestamps affected is reduced. This effect is caused by a higher level of congestion in the rural LTE network under test with nearly constant load traffic higher than 95%, compared with an average 80% to 90% (peak) in the suburban scenario.
For the optimization of spectrum usage without a dynamic range of the interference threshold and proper MCS management for minimizing interference, the percentage of timestamps with perceptible degradation is 65% and 69.2% in the rural measurements for the TOV and SFP indicators, respectively (markers t1 and s1 in Figure 8). In the suburban measurements, the TOV is exceeded on 62.5% of timestamps and 70% for the SFP indicator (markers t5 and s5 in Figure 8). The inclusion of interference optimization based on dynamic interference management (with a dynamic selection of the interference threshold in a caseby-case-based decision) has a significant impact on the video degradation minimization (as a QoE metric). For the worst optimization result, i.e., the software making decisions based on the spectrum result only (w 1 = 1; w 2 = 0), the reduction in the video degradation is 50% to 58% in the rural measurements (see marker t1 versus t2 and marker s1 versus s2 in Figure 8) and 27.5% to 30% in the suburban region (see marker t5 versus t6 and marker s5 versus s6 in Figure 8. The drawback is an 8% higher spectrum usage for both scenarios. By assigning a certain channel, the LTE BS based on traditional cognitive radio technology will measure and assess the interference in a certain channel. For optimizing the spectrum usage, the BS will minimize the usage of channels, if the interference is below the defined interference threshold. This means the interference is not considered by the software as a KPI. For the channels of nearby transmitters, the mean interference will lead to the fixed interference threshold defined. With the introduction of interference as a KPI, even for the case of the highest weight for spectrum optimization, a lower interference is expected. This is because     (5). These values are further used for setting the interference to the broadcasting coverage edge by the LTE-A network. 13 Wireless Communications and Mobile Computing a flexible range of interference allowance causes a higher degree of freedom for assigning the spectrum while keeping interference low. Hence, even for the case when the algorithm makes the user connection decisions based on the spectrum assignment results only, a slightly lower interference is achieved by the algorithm.
In the suburban scenario, the usage of frequencies by the primary broadcasting service is higher due to a higher density of traffic per square kilometer. As a consequence, the algorithm probability for reducing the interference is lower, taking into account that this KPI does not define the network optimization final result for the case of w 1 = 1 and w 2 = 0.
When the spectrum usage and interference metrics are equally weighted by the algorithm for deciding the user's connections, the SFP and TOV are exceeded in only 4.2% and 5% of the time for the rural measurements (markers s3 and t3 in Figure 8), and just 1.7% and 2.5% for the suburban measurements (markers s7 and t7 in Figure 8, SFP and TOV, respectively). As a drawback, spectrum usage is increased by 6% and 20% for the rural and suburban scenarios, respectively.
When interference minimization is the only KPI considered by the algorithm for deciding the user connections, there is no noticeable distortion in the transmitted video (markers 4 and 8 in Figure 8). Hence, in such a case, at the edge of the coverage of the broadcasting transmitter, there is no any disturbance in the users' QoE. The drawback is a higher spectrum usage by 10% compared with the one in marker s3|t3 and 22% higher compared with marker s1|t1 for the rural scenario. In the suburban scenario, the spectrum (marker 8) is increased by 13% compared with the one in marker s7|t7 and 36% higher compared with marker s5|t5.

Experiment Result Interpretation.
In a real LTE-A scenario based on the traditional distributed cognitive radio architecture, the protection margins recommended in the literature are not enough for protecting the primary broadcasting service from harmful interference. Our experiments revealed that for an LTE-A BS load higher than 80%, the broadcasting video is degraded up to 70% of the time (at the broadcasting transmitter coverage edge). This is because such margins were developed for typical broadcasting scenario cases, not taking into account the variation of the transmitter's configurations and propagation environments for different scenarios. The variation in the broadcasting transmitters' height, MCS, EIRP, and environmental propagation conditions cause significant variability in the radiation footprints and the radius of the protection contour.
A simple reduction in the maximum allowable interference (e.g., from -95 dBm to -97 dBm) might solve the interference problems, but at the cost of higher spectrum usage, that might be totally inefficient. For instance, reducing the maximum allowable interference threshold in the suburban area by 2 dB, the spectrum usage increases by 17.1%. Hence, a flexible interference range and dynamic interference management on a case-by-case basis must be applied, for guaranteeing a better trade-off between 14 Wireless Communications and Mobile Computing interference and spectrum usage. The designed algorithm for a centralized access management in the cognitive radio network allows reducing interference by at least 27% with just an 8% increase in spectrum usage.

Conclusions
In this paper, we presented an optimization algorithm for dynamic interference management in TVWS and for the minimization of spectrum occupancy. By emulating the critical conditions for the primary licensed service in an experimental testbed, we validated the algorithm performance for reducing harmful interference. Our experiments and measurements reveal that the protection margins proposed in the literature for DTMB and DVB-T2 cause severe interference when the LTE-A BS load is higher than 80%. Because of the variability of the broadcasting station configurations, environmental and scenario characteristics, a flexible range for managing interference is more efficient and effective than defining a fixed constraint. The proposed algorithm for a centralized spectrum management architecture for cognitive radio technologies allows reducing the interference at least by 50% and 27.5% in the conducted rural and suburban subjective measurements, respectively, with just an 8% increase in spectrum usage. Future work will consist of the implementation of our algorithm using a deep machine learning system for automated real-time interference and spectrum usage optimization.

Data Availability
Data underlying the findings of this study are openly available through Research Gate.

Conflicts of Interest
The authors declare that there is no conflict of interest regarding the publication of this paper.