Intercell Interference Coordination for UAV Enabled URLLC With Perfect/Imperfect CSI Using Cognitive Radio

Ultra-reliable and low latency communications (URLLC) will be the backbone of the upcoming sixth-generation (6G) systems and will facilitate mission-critical scenarios. A design accounting for stringent reliability and latency requirements for URLLC systems poses a challenge for both industry and academia. Recently, unmanned aerial vehicles (UAV) have emerged as a potential candidate to support communications in futuristic wireless systems due to providing favourable channel gains thanks to Line-of-Sight (LoS) communications. However, usage of UAV in cellular infrastructure increases interference in aerial and terrestrial user equipment (UE) limiting the performance gain of UAV-assisted cellular systems. To resolve these issues, we propose low-complexity algorithms for intercell interference coordination (ICIC) using cognitive radio when single and multi-UAVs are deployed in a cellular environment to facilitate URLLC services. Moreover, we model BS-to-UAV (B2U) interference in downlink communication, whereas in uplink we model UAV-to-BS (U2B), UAV-to-UAV (U2U), and UE-to-UAV (UE2U) interference under perfect/imperfect channel state information (CSI). Results demonstrate that the proposed perfect ICIC accounts for fairness among UAV especially in downlink communications compared to conventional ICIC algorithms. Furthermore, in general, the proposed UAV-sensing assisted ICIC and perfect ICIC algorithms yield better performance when compared to conventional ICIC for both uplink and downlink for the single and multi-UAV frameworks.

the round-trip delay. Moreover, depending on the missioncritical application at hand, the ultra-high reliability target can drop to a lower value, i.e., 10 −9 , as such, 5G networks are not designed to cater to such stringent reliability targets for facilitating URLLC services [13]. Under such conditions, 6G networks come to the rescue and aim to enhance URLLC services by improving both fronts, namely reliability and latency targets by at least two magnitude folds. Presently, for radio access networks (RAN), 5G networks can realistically achieve the target of 1 ms transmission delay [14], [15], [16], [17]. Nonetheless, transmission delay amounts to a small portion of end-to-end (E2E) delay, whereas other delay types such as processing, queuing, round trip hybrid automatic repeat request (HARQ), and packet decoding delays also factor in and contribute significantly to the overall E2E delay, which in turn could deter URLLC targets by becoming bottlenecks in networks. Additionally, latency and reliability are present on the opposite end of the same spectrum, i.e., they are inversely proportional [18]. Therefore, these relevant issues about latency and reliability have not been properly investigated by present 5G networks. Moreover, URLLC systems generate short blocklength packets which follow the so-called finite blocklength capacity formula. As such, URLLC systems are not throughput-centric but are latency and reliability centric which signals a shift in paradigm from previous wireless communication technologies. Thus, for finite blocklength, URLLC packets, the capacity bounds, and channel coding rates were derived and reviewed in [19].
In recent years, Unmanned Aerial Vehicles (UAV) have become more prominent and are used in public, military, and industrial applications. In this regard, a report from the Federal Aviation Administration (FAA) mentions that the fleet of drones will be almost doubled from 1.1 million units in 2017 to 2.4 million units by the year 2022 [20]. Moreover, UAV-enabled communications have appeared as a prevalent choice due to providing favourable channel gains thanks to line of sight (LoS) communications. Additionally, UAV can act as an aerial base station (ABS) and flying relays to facilitate communications as outlined in [21], [22], [23], [24], [25], [26], [27]. In [21], the authors studied a novel UAV trajectory design based on the so-called finite Fourier series (FFS) for multiple ground users. The authors aimed to optimize the max-min rate for all the users. Moreover, the authors demonstrated that their proposed approach yielded superior performance compared to the baseline scheme designed using discrete waypoints. In [22], the authors consider both the average packet error probability (APEP) as well as effective throughput (ET) of the control link in UAV communications, where the ground users are sending control signals to the UAV requiring URLLC. In this regard, the authors derived the closed-form expressions for both APEP and ET by making use of the so-called Gaussian-Chebyshev quadrature and then verified the analytical results by performing Monte-Carlo simulations. Similarly, in [23], the authors considered UAV as an ABS serving multiple ground users. The authors proposed a novel cyclical multiple access (CMA) to facilitate UAV-to-ground user communication scheduling. Furthermore, the authors sought to optimize the max-min rate between UAV and ground users. In [24], the authors proposed an energy-efficient UAV-relaying scheme to facilitate communication between Base Stations (BS) and ground users. The authors sought to optimize UAV and BS transmit power as well as speed, acceleration, and UAV trajectory to maximize energy efficiency. However, in [21], [22], [23], [24] the interference is overlooked. According to [28], studies have been conducted by the 3rd Generation Partnership Project (3GPP), which highlights the disadvantage of utilizing a UAV in a cellular infrastructure as it could lead to increased interference in a system containing both aerial and ground user equipment (UE) [29]. Thus, only relying on UAV Line of Sight (LoS) links without considering interference will not attain the visionary goal of large-scale UAV deployment in cellular infrastructure for futuristic wireless communication systems. A few works in open technical literature as outlined in [30], [31], have discussed intercell interference coordination (ICIC) techniques in cellular environments to mitigate the adverse effects of interference when UAV are deployed. In [30], the authors considered two user interference channels for the two UAV servicing them. The authors optimized the UAV positioning and transmit power to maximize the sum throughput. However, the authors considered LoS channels for only two users which give limited insights. Similarly, in [31], the authors studied ICIC in cellular-connected UAV using a cognitive radio approach. Nonetheless, the authors did not generalize to a multi-UAV framework concerning interference between UAV deployed in adjacent cells for uplink and downlink communications.
In this paper, we propose low-complexity algorithms for ICIC using cognitive radio when multi-UAVs are deployed in a cellular environment to facilitate URLLC services. Moreover, we model BS-to-UAV (B2U) interference in downlink communication, while in the uplink we model UAV-to-BS (U2B), UAV-to-UAV (U2U), and UE-to-UAV (UE2U) interference under perfect/imperfect channel state information (CSI). Our results show that our proposed perfect ICIC compared to conventional ICIC models fairness among single and multi-UAVs in downlink communications. Moreover, both UAV-sensing assisted ICIC and perfect ICIC algorithms perform better than conventional ICIC for uplink and downlink communications for the proposed single and multi-UAV framework.

A. ORGANIZATION AND NOMENCLATURE
The rest of this paper is organized as follows. In Section II, we present the system model and formulate the problem in the context of UAV-assisted cellular systems. Then, in Section III, we propose our algorithms for single UAV and multi-UAV interference cases, respectively. In Section IV, we present numerical results to verify the proposed algorithm's effectiveness. Finally, we present the conclusion in Section V. The symbols and notations used in this paper are summarized in Table 1.

II. SYSTEM MODEL AND PROBLEM FORMULATION
We model both downlink and uplink communications facilitated by UAV deployed in hexagonal cells as shown in Fig. 1. Accordingly, a simplistic assumption is made that outside interference is negligible. In our model, BS are primarily placed at the center of each cell with circumradius given by r c . Additionally, terrestrial UE and UAV are positioned randomly in the given cell environment. As such, each UE and UAV is considered to be assigned to the BS in the cell it is located in or above. Furthermore, each UE gets random resource blocks (RB) assigned such that upon each assignment every previous RB allocation is considered in the following way: there is no resource sharing allowed such that no UE or BS can utilize the same requested RBs in neighbouring cells of nearest q tiers. Equivalently, we can say if q is unity, then no two neighbouring cells can have UE or BS using the same RB. Similarly, if q is two, then we look at two neighbours, i.e., no neighbours of any neighbours of the considered cell may be using the requested RB. Resultantly, strong inter-cell interference (ICI) is mitigated. Additionally, the RB allocation for UAV is performed following the aforementioned neighbour principle relative to its serving BS. It is worth mentioning that downlink communication is established between BS and UAV, and conversely, uplink communication is established between UAV and BS. According to [32], to calculate the data rate R dl, ul (n) 1 of each UAV k at each RB n in the uplink and downlink communications for short URLLC packets, we use the data rate equation where is the inverse Q-function, B is the bandwidth, and V(γ dl, ul (k, n)) = 1−(1+γ dl, ul (k, n)) −2 is the channel dispersion, which refers to the channel variability. For the sake of generality, we did not approximate channel dispersion equal to unity as only for high Signalto-noise ratio (SNR), such an assumption holds, whereas for low SNR it becomes invalid. The second term in (1), is the finite blocklength penalty term. Thereafter, for the downlink communication between BS and UAV, we define interference I dl (k, n) and SINR γ dl (k, n) as where P j (n) is the transmission power of BS j in RB n, F j (k, n) is the downlink channel power gain from BS j to 1. For uplink and downlink communications, Shannon's capacity for a user is given by R dl,ul (k, n) = B log 2 (1 + γ dl, ul (k, n)). Comparably, the finite blocklength capacity equation is given by . Furthermore, the additional terms are the penalty terms incurred due to short blocklengths. Therefore, for the large blocklengths, i.e., approaching infinity, the penalty terms cease to exist, and the finite blocklength capacity simply reduces to Shannon's capacity.
UAV k in RB n, j k denotes the BS serving UAV k, and σ 2 is the received Gaussian noise power at any UAV. Moreover, I dl (k, n) is the total Gaussian terrestrial ICI power at UAV k in RB n, J (n) is the set of BSs using RB n. Now, for a multi-UAV environment, we define J (n), which requires subtraction of sets J (n) \ {j k } when used as a set of indices. Now, for the uplink communication between the UAV and the BS, we model SINR γ ul (k, n) as where p ul (k, n) is the transmission power of UAV k in RB n, G j (k, n) is the uplink channel power gain from UAV k to BS j in RB n, and σ 2 is the Gaussian noise power at the receiver of any BS. Moreover, I ul (j k , n) denotes the total uplink interference at UAV k's serving BS j k from all other non-k UAVs in RB n. Additionally, another important interference is the maximum interference UAV k causes to any other non-j k BS using the same RBs, which is given as where N ul (k) denotes the RBs set for the uplink that is allocated to UAV k by its serving BS. Conversely, N dl (k) denotes the RBs set for the downlink, respectively. In both the conventional and UAV-sensing-assisted ICIC for the downlink, the power control is simple: BS always use peak available power P dl to maximize the UAV's achievable rate in the downlink. Similarly, in the conventional ICIC in the uplink peak power P ul is always used. However, in the uplink in the UAV-sensing-assisted model outlined in [31], the power control is more complicated. Here, we generalize it to the k th UAV as where U is a defined threshold of the BS receiver noise power σ 2 , E ul (k, n) is the sensed uplink transmission at UAV k in RB n, α L denotes the LoS path-loss exponent, and ρ k has a complex expression for distance as outlined in (8) in [31]. Moreover, the main idea behind E ul (k, n) as described in detail in [31] is that we do not have the knowledge of the CSI, hence we have to predict it from sensed UE uplink transmissions. Thus, the fundamental difference here is that the sensed transmissions will also include transmissions from other UAV. This will result in a higher E ul (k, n) and hence a lower p ul (k, n) than in [31]. The formulation is given as where E j (k, n) is the channel gain between UAV k and the UE or UAV served by BS j in RB n in the uplink. Moreover, if we know the CSI, the power allocation would simply use the known G j (k, n) as Finally, achievable rates R dl (k, n) or R ul (k, n) for both uplink or downlink communications are given by replacing Signalto-interference-plus-noise ratio (SINR) given by (2) or (4) in (1). Therefore, the total sum achievable rates over the used RB are given by According to [31], for all channel gains the probabilistic 3GPP Urban Macro path loss model is the suitable choice. Additionally, with our given simulation parameters, a UAV altitude of at least 200 metre (m) gives us a pure LoS model, to which we add small-scale Ricean fading denoted by X f , thus we have (11) where f GHz c is the carrier frequency expressed in GHz. To obtain gain, we negate PL and convert it to linear.

III. PROPOSED ALGORITHMS
In this section, we outline our proposed algorithms for single and multi-UAV interference scenarios.

A. PROPOSED ALGORITHMS FOR SINGLE-UAV INTERFERENCE
We propose three algorithms UAV-enabled common (CMN)-ICIC, UAV-enabled cognitive-sensing based superior (CSS)-ICIC, and UAV-enabled cognitive-sensing based (CSB)-ICIC. Algorithm 1 outlines UAV-enabled CSS-ICIC, which is repeated over random initializations L in times, then all the results are averaged over L avg iterations. Furthermore, for Algorithm 1, the UAV-enabled CMN-ICIC would omit step 9 and determine N dl, ul by simply assigning the UAV random RB with neighbor criterion in mind, and by setting p ul (n) = P ul ∀n. Additionally, M d denotes the number of RB candidates in the downlink, and UAV-enabled CSB-ICIC sets M d to the number of available RBs after UE RB allocation. In other words, UAV-enabled CSB-ICIC is equivalent to UAV-enabled CMN-ICIC if we set M d or M u equal to the number of available RBs after UE RB allocation. Hence, Algorithm 1 represents all the proposed algorithms with a few changes. Additionally, for the proposed schemes, including UAV-enabled CMN-ICIC, CSS-ICIC, and CSB-ICIC for the downlink, the power control is simple: the BSs always use peak available power P dl to maximize the UAV's achievable rate. Similarly, for the UAV-enabled CMN-ICIC, in the uplink, peak power P ul is always used. Conversely, UAV-enabled CSS-ICIC and CSB-ICIC both utilize (6) for power control in uplink communications to maximize the UE's data rate. To determine the sum complexity, we note Calculate: F i (n) or G i (n) ∀i ∈ J (n) ∪ {1} using (11).

13:
Calculate: p ul (n) using (7), and γ ul (n) using (4), for the uplink. 14: Calculate: R dl, ul (n) using (1). 15: end for 16: Calculate: N dl, ul = {Highest N d or N u RB ranked by value of R dl, ul (n)}. 17: Return: R T dl, ul using (10) or I ul using (6)   Since the UAV-enabled CSB-ICIC is equivalent to the UAVenabled CMN-ICIC with maximum M d or M u , hence it yields a similar time complexity. However, since a higher number of candidates means more iterations in the for loop, this translates to more calculation time and energy expenditure by the UAV. In this regard, we could optimize the time complexity of the UAV-enabled CSB-ICIC using vectorization of loops, which moderately increases the memory resources. Nevertheless, this is not an issue for modern UAVs equipped with a powerful system of chips (SoC). Return: If in uplink, calculate power control by (7). 9: end for

B. PROPOSED ALGORITHMS FOR MULTI-UAV INTERFERENCE CASE
For the conventional ICIC, a requested number of RB is simply assigned randomly. However, in the UAV-sensing-assisted ICIC proposed in [31], a more complex algorithm is at play wherein a given number of candidates are first randomly picked out complying with the neighbor requirement, of which those with the lowest sensed interference are chosen to be assigned to the UAV. Now, we outline the UAV-sensing assisted ICIC scheme for multi-UAV framework, whereas the conventional ICIC would eliminate the entire candidates part and determine N dl, ul (k) by simply assigning the UAV random RB with neighborhood criterion, and by setting p ul (k, n) = P ul , ∀k∀n. Further, M d or M u denotes the number of RB candidates in the downlink/uplink, and a perfect ICIC takes M d or M u equal to the total number of RBs in the downlink/uplink. Moreover, N d or N u is the number of RBs the UAV requests. Additionally, UAVs become operational one after another, and RB allocation for each UAV is to be done only upon it becomes operational in the service area. This avoids an endlessly iterative process where UAVs would constantly be changing their allocation to account for changes in the environment. Here, Algorithm 2, outlines the program executed in each UAV, whereas Algorithm 3, outlines the scheme we use to simulate the whole environment and extract real data rates and interferences. Now we determine the time complexity of the ICIC Algorithm 2, implemented in each UAV. The for loop deterministically cycles through all M d or M u candidates and performs spectrum sensing. All other steps are constant time. Therefore, the total time complexity for the UAV-sensing The conventional ICIC avoids a loop and simply randomly assigns RBs, hence it is constant time.

IV. NUMERICAL RESULTS
For all calculations, we use the setup defined in Table 2. Run: Algorithm 2. For sensed interferences use (3) or (8).

11:
Update: J (n) ← J (n) ∪ {j k } ∀n ∈ N dl, ul (k) 12: Update (uplink only): K(n) ← K(n) ∪ {k} ∀n ∈ N ul (k) 13: end for  Additionally, there are 3, 10, or 15 UAV with an altitude 2 between 200 and 500 m. In downlink and uplink, the UAV requests 1 and 5 RB, respectively. Furthermore, 2. It is also worth mentioning that according to 3GPP release 15, for UAVs flying at a high altitude, i.e., above 100 m, there is a 100% probability of achieving the LoS. Moreover, LoS path-loss will dominate over the other NLoS components. Thus, UAV will associate itself with a BS with the largest LoS channel gain.  we run 1000 simulations with random channel and location initializations, over which all results are averaged.

A. RESULTS FOR SINGLE UAV INTERFERENCE CASE
In Fig. 7, we demonstrate how the UAV sum uplink rate and maximum interference vary when changing the thresholds for the UAV-enabled CSS-ICIC and UAV-enabled CSB-ICIC. Since there is no threshold for UAV-enabled CMN-ICIC, there is only one point for it in the plot. It is to be noted that on the plot, the slopes on the left side are negative. To further elaborate on this effect, we again plot the UAV sum uplink rate and maximum interference with variable blocklengths M and decoding error probabilities ε. Now, we look at Fig. 8, the left tails of the curves are sloping down. Due to monotonicity, there is a clear trade-off between interference and data rate in the Shannonian equation with  short blocklength data rate due to non-monotonicity; such a trade-off does not necessarily exist. Consequently, we do not have to sacrifice one for the other if we happen to also want to constrain ourselves below a certain interference level, e.g., if we are required to keep interference below −140 dBm which is achieved by appropriately adjusting the threshold U , then we can simply choose the leftmost point on the curve and be assured that we will both decrease the interference and maximize the data rate under our constraint I ul ≤ −140 dBm. This is a significant advantage that mitigates the disadvantage implied by the penalty term. In Fig. 9, we present the effect of the blocklength M and decoding error probability ε on the data rate R dl . Here, we fix M = 200 symbols (resp. ε = 10 −9 ) and vary ε (resp. M). Moreover, we set P dl = 10 dBm and we observe the behavior of Fig. 9, based on Eq. (1), where blocklength M is present in the denominator of the penalty term of the data rate and Q −1 (ε) function which is monotonically decreasing; hence higher values of M and ε tend to yield higher data rates. In Fig. 10, we compare the UAV downlink sum  achievable data rate R dl to the peak downlink power P dl . Here, we observe that UAV-enabled CSB-ICIC yields the best performance while UAV-enabled CMN-ICIC shows the worst performance. Moreover, we also observe that UAVenabled CSS-ICIC yields a mediocre performance. However, as M d increases, its performance becomes superior to that of UAV-enabled CMN-ICIC. Finally, in Fig. 11, we study how the peak power P dl of the BS is adjusted according to the blocklength and decoding error probability. For reference, we use P dl = 30 dBm. The graphs in Fig. 11 tell us that to recover from the lower data rates at lower M and ε, we must increase the power. Resultantly, UAV-enabled CMN-ICIC should adjust its power much more since the slope of its graph in Fig. 10 is much shallower than that of UAV-enabled CSB-ICIC. As evidenced by Fig. 11, the performance of UAV-enabled CSB-ICIC is better than UAV-enabled CMN-ICIC, especially when smaller blocklengths are considered and a lower decoding error probability is required, which is the case for URLLC. Therefore, we recommend the UAVenabled CSB-ICIC for URLLC systems as it yields the best  performance, which is noticeable for blocklengths of around 200 symbols and a decoding error probability of 10 −9 . Thus, the recommended scheme, which provides the best results for URLLC systems, requires moderately higher memory resources at the UAV as described earlier.

B. RESULTS FOR MULTI-UAV INTERFERENCE CASE
In Fig. 10, we consider sum achievable downlink data rates averaged over deployed UAVs, we computeR dl = 1 N UAV k R T dl (k), and then averaged it over all simulations, and compare it to peak downlink power P dl . Here, we observe that the conventional ICIC gives the worst results and the perfect UAV-assisted ICIC gives the best results. Furthermore, we see that the average data rate among deployed UAVs gets worse as the number of UAV N UAV grows, which is intuitively expected, as the UAV have to share more of the same RB and hence the probability of interference is greater. In Fig. 4, we again consider BS peak downlink transmit power versus UAV sum achievable downlink rate, but we make a different comparison. First, one of the lines depending on N UAV and type of ICIC, from Fig. 10 is placed on each of the 4 subplots. This is always the middle line in each subplot which is colored orange. Moreover, around it, we average and maximize the maximums and similarly minimize the minimums in the following way. In each simulation, besides the average data rate, we also record the deviations of the minimum and maximum inter-UAV rates: min k R T dl (k) −R dl and max k R T dl (k) −R dl . Then, we average these deviations and inter-UAV averages denotedR dl over all simulations, subsequently, we add the simulation-averaged deviations back into the simulation-averaged inter-UAV averages. The extra steps help dampen the effect of variance between simulations due to the random nature of their initialization and ensure that deviation comparisons happen on a per-simulation basis. The resulting curves are adjacent to the middle curve in each subplot which is shown by red and purple lines. Additionally, aside from averaging the deviations over simulations, we also find the minimum of the down-deviations and maximum over the up-deviations over the simulations, adding them back into simulation-averagedR dl . This results in the outermost curves on each subplot shown by blue and green lines. Moreover, this graph helps us assess the average dispersion between the highest-rate UAV and the lowest-rate UAV. As we see, in the conventional ICIC we inevitably end up with some UAV having a 0 data rate, while with the perfect ICIC this was always avoided. This was done at the cost of reducing data rates of the top-performing UAV, as the top line is closer to the middle line in the perfect ICIC. Thus, we guarantee fairness among UAV for downlink communications. In the graphs that follow we explore the uplink. In Fig. 5, we observe the inter-simulation average of the inter-UAV average maximum interference power over the RBs assigned to the UAVs relative to threshold U . Here, we have set P ul = 10 dBm and M u = 10. Moreover, the interference for predicted CSI is computed with power allocation from (7), while interference for perfect CSI is computed with power allocation from (9). Now, we see that the perfect CSI does not perfectly follow a linear curve, which is not surprising. As the UAVs become operational one after the other, so each new UAV mitigates the interference it provides to BS without knowing which new RBs, the BSs will be assigned in the uplink for the UAVs that become operational afterward. Therefore, the dashed lines are for the most part above the identity line up to the plateau caused by the fact that power control is limited from above by peak power P ul . Furthermore, the interference with predicted CSI denoted by solid lines is as expected, lower for each corresponding N UAV . Also, despite the uneven nature of the lines, they are still, as should be, monotonously increasing up to the plateau. This is because a higher threshold allows for higher uplink power, which increases interference at the BSs. The effect from the variability of simulations is smoothed over by averaging over them, as is done in all graphs. Lastly, in Fig. 7, we plot how the sum uplink rate and maximum interference vary with each other by varying thresholds to obtain these values. Since for the conventional ICIC, there are no thresholds to vary, there is only one point for it on the plot for each scenario. As with downlink, a higher number of UAVs generally worsens the data rate for a fixed interference, as does a lower number of RB candidates.

V. CONCLUSION
In this paper, we studied ICIC for the single and multi-UAV framework to facilitate URLLC services under perfect/imperfect CSI using cognitive radio. The goal was to mitigate strong ICI in such single and multi-UAV-assisted cellular networks. To achieve this goal, we modelled BSto-UAV (B2U) interference in downlink communication, whereas in the uplink we modelled UAV-to-BS (U2B), UAVto-UAV (U2U) and UE-to-UAV (UE2U), respectively. In this regard, we proposed low-complexity algorithms namely perfect ICIC, UAV-sensing assisted ICIC and perfect ICIC. Results showed that the proposed perfect ICIC provided fairness among UAVs especially in downlink communications compared to conventional ICIC algorithms. Lastly, we demonstrated in general, that the proposed UAV-sensing assisted ICIC and perfect ICIC algorithms gave better performance than conventional ICIC for both uplink and downlink for the single and multi-UAV framework.