In this thesis, I present a comprehensive study on spectrum allocation and decision-making methods for unmanned aerial vehicle (UAV) communications in cognitive radio networks. The rapid growth of UAV applications has created an urgent demand for efficient spectrum utilization, especially when multiple UAVs cooperate to support ground users. Cognitive radio technology offers a promising solution by enabling dynamic spectrum access. However, the inherent heterogeneity and dynamic nature of cognitive wireless networks introduce significant challenges, including mixed interference, high user density, and time-varying channel conditions. To address these challenges, I formulate three representative scenarios and propose respective optimization algorithms. My work focuses on maximizing the minimum user throughput and ensuring fairness while protecting primary network communications. The numerical simulations demonstrate that the proposed methods outperform conventional baselines in terms of throughput, fairness, and conflict reduction.
1. Introduction and Background
The deployment of unmanned aerial vehicle (UAV) communication systems has attracted considerable attention due to their flexibility, rapid deployment, and ability to establish line-of-sight links. However, a single UAV cannot always satisfy the coverage and capacity requirements of large-scale networks. Therefore, multi-UAV systems are utilized to improve coverage, computational capability, and mission adaptability. Yet, such systems require a large amount of spectrum resources, which are becoming increasingly scarce. According to the Federal Communications Commission, a large portion of licensed spectrum is significantly underutilized, and hence cognitive radio technology is considered a key enabler for dynamic spectrum sharing.
In a cognitive wireless network, primary users (PU) and secondary users (SU) coexist. The secondary network, which includes multiple UAV swarms, can access the licensed spectrum opportunistically or underlay with power constraints. Nevertheless, the spectrum allocation and decision-making for multi-UAV systems in cognitive radio networks are still challenging due to: (i) multiple interference forms, including co-channel and adjacent-channel interference, (ii) spectrum allocation unfairness when the number of ground users increases, and (iii) the difficulty of making accurate spectrum decisions under time-varying channel states without prior knowledge.
The main contributions of my work are summarized as follows. First, I propose a joint optimization method for channel assignment and power allocation in a mixed interference scenario. The method uses a heuristic algorithm to solve the subproblem of sub-channel allocation and the particle swarm optimizer to solve the power allocation, while the block coordinate descent technique is applied to iterate between them. This effectively enhances the minimum signal-to-interference-plus-noise ratio (SINR) among secondary users. Second, I develop a joint deployment and bandwidth allocation scheme for a high user density scenario where UAVs act as aerial relays. The K-Means clustering algorithm optimizes UAV positions and a linear precise method dynamically adjusts bandwidth to maximize the minimum user throughput while limiting interference to the primary network. Third, I adopt a multi-agent proximal policy optimization (MAPPO) framework with a temporary memory mechanism to implement dynamic spectrum access in a time-varying channel environment. The proposed algorithm enables UAVs to learn efficient channel access policies with limited sensing information, thereby reducing frequency conflicts and improving the sum rate.
In the following sections, the system models and problem formulations are described in detail, followed by the proposed solution methodologies, simulation setups, and conclusions.
2. Mixed Interference Scenario: Joint User Matching and Resource Allocation
2.1 System Model
I consider a cognitive radio network where multiple unmanned aerial vehicles are deployed as aerial base stations to serve ground users in a specified region. The primary network is active nearby, and the secondary network shares the licensed spectrum with an underlay mode. Therefore, the interference from the secondary users to the primary receiver must remain below a prescribed threshold. Meanwhile, the secondary users experience co-channel interference from the primary transmitter and adjacent-channel interference caused by the non-ideal receiver filters. The objective is to maximize the minimum SINR among all secondary users.
Let the set of UAVs be $\mathcal{M}=\{1,2,\dots,M\}$, the set of ground users be $\mathcal{K}=\{1,2,\dots,K\}$, and the set of sub-channels be $\mathcal{N}=\{1,2,\dots,N\}$. The UAV-user association variable is denoted as $w_{m,k}\in\{0,1\}$, where $w_{m,k}=1$ means UAV $m$ serves user $k$. Each user is associated with one UAV, which yields:
$$
\sum_{m=1}^{M} w_{m,k}=1, \quad \forall k \in \mathcal{K}.
$$
The channel allocation variable is $a_{k,n}\in\{0,1\}$, where $a_{k,n}=1$ indicates that user $k$ occupies channel $n$. Every user occupies exactly one channel, while for a given UAV, each channel is assigned to at most one user:
$$
\sum_{n=1}^{N} a_{k,n}=1, \quad \forall k,
$$
$$
\sum_{k=1}^{K} w_{m,k} a_{k,n} \le 1, \quad \forall m,n.
$$
The transmission power from UAV $m$ on channel $n$ is $p_{m,n}$, and the total power of each UAV is limited:
$$
\sum_{n=1}^{N} \sum_{k=1}^{K} w_{m,k} a_{k,n} p_{m,n} \le P_{m}^{\max}, \quad \forall m.
$$
The air-to-ground channel gain between UAV $m$ and user $k$ is modeled as:
$$
g_{m,k} = P_{LOS} d_{m,k}^{-\alpha} + \eta P_{NLOS} d_{m,k}^{-\alpha},
$$
where $P_{LOS}$ is the line-of-sight probability, $P_{NLOS}=1-P_{LOS}$, $\eta$ is the attenuation factor for non-line-of-sight links, $\alpha$ is the path-loss exponent, and $d_{m,k}$ is the distance between UAV $m$ and user $k$.
2.2 Problem Formulation
Let $\gamma_{k}$ be the SINR received by user $k$. It can be expressed as:
$$
\gamma_{k} = \frac{w_{m,k} a_{k,n} p_{m,n} g_{m,k}}{I_{CCI} + I_{ACI} + \sigma_{n}^{2}},
$$
where $\sigma_{n}^{2}$ is the additive white Gaussian noise power, $I_{CCI}$ is the co-channel interference from the primary transmitter:
$$
I_{CCI} = p_{p} g_{p},
$$
and $I_{ACI}$ is the adjacent-channel interference from other secondary users:
$$
I_{ACI} = \sum_{n’ \ne n} \sum_{j \ne k} a_{k,n} a_{j,n’} p_{m,n’} g_{m,j} \mu_{k,j}.
$$
The optimization problem is:
$$
\max_{\mathbf{W},\mathbf{A},\mathbf{P}} \min_{k} \gamma_{k}
$$
subject to the above constraints and the interference threshold to the primary network:
$$
\sum_{m=1}^{M} \sum_{n=1}^{N} p_{m,n} g_{m,k^{*}} \le \delta,
$$
where $k^{*}$ denotes the primary user location index.
2.3 Proposed Solution
Since the original problem is a mixed-integer nonlinear programming problem, I decompose it into three subproblems. First, the UAV-user matching is optimized by the K-Means clustering algorithm, which minimizes the average distance between each UAV and its assigned users. Second, for a fixed power allocation, the channel assignment is optimized by a genetic algorithm. Third, for a fixed channel assignment, the power allocation is optimized by the particle swarm optimization (PSO) algorithm. The block coordinate descent method alternates between the channel and power subproblems until convergence.
Algorithm 1 summarizes the joint optimization procedure.
Algorithm 1: Joint UAV placement, channel assignment, and power allocation
Input: number of UAVs $M$, users $K$, channels $N$, user coordinates.
Initialize: feasible $\mathbf{A}^{(0)}$, $\mathbf{P}^{(0)}$, $r=0$.
Repeat:
1. Apply K-Means to obtain the UAV deployment and user matching.
2. Fix $\mathbf{P}^{(r)}$, solve the channel assignment subproblem via genetic algorithm to obtain $\mathbf{A}^{(r+1)}$.
3. Fix $\mathbf{A}^{(r+1)}$, solve the power allocation subproblem via PSO to obtain $\mathbf{P}^{(r+1)}$.
4. $r=r+1$.
Until: maximum iterations reached.
Output: optimized $\mathbf{W}$, $\mathbf{A}$, $\mathbf{P}$.
In the genetic algorithm, the population size is set to 70, the crossover probability is 0.9, and the mutation probability is 0.1. The PSO uses an inertia weight of 0.8 and two learning factors equal to 1.5. I evaluated the convergence behavior with different population sizes and found that a larger population leads to faster convergence and a better minimum SINR because it provides a broader exploration of the search space.
Simulation results show that the minimum user SINR decreases as the number of users increases because more users compete for limited spectrum resources and the adjacent-channel interference becomes stronger. In contrast, increasing the number of UAVs significantly improves the minimum SINR because more power resources are available. The joint optimization method outperforms a random channel assignment baseline and a fixed channel baseline by a large margin. Furthermore, when the receiver filters are ideal (i.e., no adjacent-channel interference), the performance is always better, which confirms the need to consider adjacent-channel interference in practical systems.
3. High User Density Scenario: UAV Relay Deployment and Bandwidth Allocation
3.1 System Model
In this scenario, I consider a cognitive radio network in a dense urban environment where direct communication between a temporary relay node and ground users is severely obstructed by buildings. Therefore, multiple unmanned aerial vehicles are deployed as aerial relays to forward the signal from the relay node to the users. Let the set of UAV relays be $\mathcal{M}=\{1,\dots,M\}$ and the set of ground users be $\mathcal{K}=\{1,\dots,K\}$ with $K \gg M$. The primary network consists of a base station and primary users, and the secondary network shares the spectrum under an underlay mode. The interference from the UAV relays to the primary receiver must be kept below a threshold $\delta$.
Each user is assigned to one relay UAV, and the association variable $w_{m,k}$ is equal to 1 if user $k$ is served by relay $m$. The bandwidth of the temporary node is $W_{g}$, and the total bandwidth of all relays equals $W_{g}$:
$$
\sum_{m=1}^{M} W_{relay,m} = W_{g}.
$$
For each cluster $m$, the sum of user bandwidths equals the relay bandwidth:
$$
\sum_{k=1}^{K} w_{m,k} W_{k} = W_{relay,m}, \quad \forall m.
$$
In addition, the throughput of each relay cannot exceed the backhaul capacity provided by the temporary node, and the total throughput of all relays is limited by $\tau$:
$$
\sum_{k=1}^{K} w_{m,k} T_{k} \le T_{m}, \quad \sum_{m=1}^{M} T_{m} \le \tau.
$$
3.2 Problem Formulation
The objective is to maximize the minimum user throughput $T_{k}$ while ensuring the interference constraint to the primary network:
$$
\max_{\mathbf{W},\mathbf{T}} \min_{k} T_{k}
$$
subject to the above constraints and $W_{relay,m} \ge W_{relay}^{\min}$, $W_{k} \ge W_{user}^{\min}$.
For a given user $k$ served by relay $m$, the SINR is:
$$
\gamma_{m,k} = \frac{w_{m,k} p_{m} g_{m,k}}{I_{CCI} + \sigma_{n}^{2}},
$$
where $p_{m}$ is the relay transmission power, $g_{m,k}$ is the channel gain, and $I_{CCI} = p_{G} g_{G,k}$ is the interference from the primary base station. The throughput of user $k$ is:
$$
T_{k} = W_{k} \log_{2}(1+\gamma_{m,k}).
$$
3.3 Proposed Solution
I use the K-Means clustering algorithm to simultaneously determine the UAV relay positions and user association. The cluster centers are iteratively refined to minimize the distance between relays and their assigned users. After the clustering converges, the bandwidth allocation is optimized by a linear precise algorithm. The key idea is to repeatedly find the user with the minimum throughput and increase its bandwidth as long as the total bandwidth constraint is satisfied and the backhaul bottleneck is not exceeded.
Algorithm 2: Linear precise bandwidth allocation
Input: user set $\mathcal{K}$, relay set $\mathcal{M}$, user positions, relay positions, backhaul capacity $\tau$.
Initialize: all relay bandwidths set to $W_{relay}^{\min}$, compute user throughputs.
Repeat:
1. Find the minimum throughput $T_{\min}$ and the second minimum throughput $T_{M}$.
2. Increase the bandwidth of the users with minimum throughput by an amount corresponding to $\beta (T_{M}-T_{\min})$, where $\beta$ is an adjustment factor.
3. Compute the required additional bandwidth using the Shannon formula and check if the total bandwidth of the temporary node is sufficient.
4. If the bandwidth limit is reached, stop; otherwise continue.
5. If the backhaul capacity is exceeded, reduce the throughput of the highest-throughput user until the total throughput falls below $\tau$.
Output: optimized user throughputs and relay bandwidths.
The adjustment factor $\beta$ is determined by:
$$
\beta = \min\left(1, \frac{W_{g} – \sum_{m} W_{m}}{A \cdot (T_{M}-T_{\min}) / \sum_{m} \log_{2}(1+\gamma_{g,m})}\right),
$$
where $A$ is the number of users with the minimum throughput and $\gamma_{g,m}$ is the SINR between the temporary node and relay $m$.
Simulations are conducted with $M=5$ relays and $K=100$ ground users distributed over a square area. The primary base station is located at (700,700), and the path-loss model follows the same A2G formula. The parameters are listed in Table 1.
| Parameter | Value |
|---|---|
| Number of UAV relays $M$ | 5 |
| Number of ground users $K$ | 100 |
| UAV height $H$ | 100 m |
| Path-loss exponent $\alpha$ | 3 |
| Primary base station position | (700,700) |
| Environment constants $\{a,b\}$ | {11.95, 0.136} |
| Noise power $N_0$ | -90 dBm |
| UAV transmission power | 1 W |
| Primary base station power | 5 W |
| Backhaul capacity $\tau$ | 50 Mbps |
Figure 3.2 illustrates the optimized UAV placement (red stars) and user clusters (colored points). The K-Means algorithm ensures that each relay is deployed close to its cluster members while maintaining the interference threshold to the primary receiver.
The convergence curves in Figure 3.3 compare the minimum user throughput for different total bandwidths. When the total bandwidth is 20 MHz, the algorithm converges to a higher minimum throughput than for 15 MHz, but the number of iterations required also increases. The curves also show that ignoring the interference threshold to the primary network would allow better relay positions, yielding higher throughput, but this would violate the primary protection constraint.
Figure 3.4 shows the throughput of each user before and after optimization. Users who were originally in the low-throughput region obtained significant gains, while some users with high initial throughput experienced a reduction. This behavior is expected because the algorithm redistributes resources to improve fairness and also enforces the backhaul bottleneck.
I also compare the proposed algorithm with a genetic algorithm and a particle swarm optimization method. The proposed method converges in about 50 iterations, which is comparable to the genetic algorithm, but the final minimum user throughput is higher. The particle swarm optimizer converges faster but reaches a lower objective value due to its global search lacking a targeted resource adjustment mechanism.
Finally, I investigate the impact of the number of UAV relays. Increasing $M$ from 3 to 7 improves the minimum user throughput because more relays allow a better spatial division and higher total bandwidth. However, the improvement saturates for large $M$ because the total bandwidth and backhaul capacity become the limiting factors.
4. Time-Varying Channel Scenario: Dynamic Spectrum Decision with MAPPO
4.1 System Model
In the third scenario, I focus on a cognitive radio network where the primary channel occupancy changes over time. Multiple unmanned aerial vehicles form a secondary network and need to select suitable channels to communicate with a ground control station. The primary network uses a round-robin pattern to switch its occupied channels. At each time slot, only a subset of channels is idle, and the channel state is unknown to the secondary UAVs unless they sense it. I adopt the overlay spectrum access mode, where the secondary UAVs can access the idle channels only when they avoid interfering with the primary users.
Let the set of UAVs be $\mathcal{M}=\{1,2,\dots,M\}$ and the set of channels be $\mathcal{N}=\{1,2,\dots,N\}$. The channel state of channel $n$ at time $t$ is:
$$
X_{n,t} =
\begin{cases}
+1, & \text{idle};\\
-1, & \text{busy}.
\end{cases}
$$
Each UAV can observe only the state of the channel it chooses to access. The observation of UAV $m$ at time $t$ is:
$$
O_{m,n,t}=
\begin{cases}
X_{n,t}, & \text{if UAV } m \text{ selects channel } n,\\
0, & \text{otherwise}.
\end{cases}
$$
This process is modeled as a partially observable Markov decision process. At each time slot, UAV $m$ chooses an action $a_m(t) \in \mathcal{N}$ representing the index of the selected channel. The reward shared by all UAVs is the sum rate:
$$
r_t = \sum_{m=1}^{M} R_m(t) = \sum_{m=1}^{M} B \log_{2}(1+\gamma_{m}(t)),
$$
where $\gamma_{m}(t)$ is the SINR of UAV $m$ at slot $t$.
The SINR at the ground control station receiving from UAV $m$ on channel $n$ is:
$$
\gamma_{m}(t) = \frac{p_m g_{m,g}}{\sum_{i \neq m} p_i g_{i,g} + \chi_{n,t} p_G g_{G,g} + \sigma^2},
$$
where $p_m$ is the UAV’s transmission power, $g_{m,g}$ is the channel gain between UAV $m$ and the ground station, $p_G$ is the primary base station power, $g_{G,g}$ is the channel gain from the primary base station to the ground control station, and $\chi_{n,t}$ indicates whether channel $n$ is occupied by the primary user at time $t$.
4.2 Problem Formulation
The objective is to maximize the accumulated sum rate over the whole mission period by optimizing the channel selection for each UAV at each timeslot:
$$
\max_{\{a_m(t)\}} \sum_{t=1}^{T} \sum_{m=1}^{M} R_m(t),
$$
subject to the constraint that each UAV selects exactly one channel per slot.
4.3 Proposed MAPPO-Based Algorithm
I adopt the multi-agent proximal policy optimization (MAPPO) algorithm, which is an extension of the single-agent PPO to the multi-agent setting. The MAPPO framework uses a centralized critic and distributed actors. Each UAV has its own actor network that maps its local observation to a probability distribution over channels. A shared or separate critic network evaluates the global state to estimate the value function.
The actor network is updated by maximizing the clipped objective:
$$
L(\theta) = \frac{1}{DM} \sum_{i=1}^{D} \sum_{m=1}^{M} \min\left( \frac{\pi_{\theta_m}(a_m^i|o_m^i)}{\pi_{\theta_m^{old}}(a_m^i|o_m^i)} A_m^i, \; \text{clip}\left( \frac{\pi_{\theta_m}(a_m^i|o_m^i)}{\pi_{\theta_m^{old}}(a_m^i|o_m^i)}, 1-\epsilon, 1+\epsilon \right) A_m^i \right),
$$
where $A_m^i$ is the advantage function calculated by GAE:
$$
A_m^i = \sum_{l=0}^{\infty} (\gamma \lambda)^l (r_{i+l} + \gamma V_{\phi}(s_{i+l+1}) – V_{\phi}(s_{i+l})).
$$
The critic network is trained by minimizing the loss:
$$
L(\phi) = \frac{1}{DM} \sum_{i,m} \max\left( (V_{\phi}(s_i) – \hat{R}_i)^2, \; ( \text{clip}(V_{\phi}(s_i), V_{\phi_{old}}(s_i)-\varepsilon, V_{\phi_{old}}(s_i)+\varepsilon) – \hat{R}_i)^2 \right),
$$
where $\hat{R}_i$ is the discounted return.
In the original policy architecture, I introduce a temporary memory mechanism: the observation matrix from the previous $L$ time slots is concatenated with the current observations before being fed into the actor network. This enables the policy to exploit historical information to infer the channel state transition pattern. The memory length is set to $L=4$ in my simulations.
Algorithm 3: MAPPO-based dynamic channel selection
Initialize: actor and critic networks for each UAV.
For each episode:
1. Reset the environment and obtain the initial observation matrix.
2. At each slot, each UAV senses the selected channel and constructs the observation tensor.
3. The actor network outputs the action (channel index) according to the current policy.
4. The environment computes the reward and the next state.
5. Store the transition into the replay buffer.
6. After collecting a batch of data, compute advantages and update actor/critic networks.
The simulation parameters are summarized in Table 2. The area is 1000 m by 1000 m, the ground control station is at the origin, and the primary base station is located at (7000,3500). There are 16 channels, and the number of idle channels is set to 6 in the default setting. The channel switch probability is $p=0.2$.
| Parameter | Value |
|---|---|
| Ground control station position | (0,0) |
| Primary base station position | (7000,3500) |
| Number of channels $N$ | 16 |
| Channel bandwidth $B$ | 1 MHz |
| UAV power $p_m$ | 23 dBm |
| Primary base station power $p_G$ | 60 dBm |
| Path-loss environment constants | a=11.9, b=0.13 |
| NLOS attenuation factor | 20 dB |
| Noise power spectral density | -100 dBm/Hz |
| GAE parameter $\lambda$ | 0.95 |
| Discount factor $\gamma$ | 0.99 |
| Clip parameter $\epsilon$ | 0.2 |
| Learning rate | 5e-4 |
The simulation results in Figure 4.5 show the convergence of the average reward for different numbers of channels (24, 16, and 10) with 6 idle channels and 5 UAVs. In the early iterations, the curve with more channels obtains a higher reward because there is more opportunity to choose idle channels. As training proceeds, all curves converge to the same reward level, implying that the algorithm successfully learns to find the idle channels even in the scarce channel condition.
I also count the frequency conflicts during training. A conflict occurs when two or more UAVs select the same channel. As presented in Figure 4.6, the number of conflicts first increases and then decreases. The peak conflict count is larger when the total number of channels is smaller. After about 600 episodes, the conflict count converges to a low value for all cases, demonstrating that the learned policies can coordinate with each other even without explicit information exchange.
In Figure 4.7, I compare the converged reward and the average conflict count under different numbers of idle channels. While the final reward values are similar, the average conflict count drops significantly when the idle channel count increases from 4 to 8. This is intuitive because more idle channels reduce the probability of multiple UAVs picking the same channel.
Figure 4.8 shows the total system throughput versus the number of UAVs. When the number of idle channels is 5, the total throughput initially increases with the number of UAVs but then drops when the number of UAVs exceeds 6. The reason is that when more UAVs than idle channels exist, two UAVs may have to share the same channel, causing co-channel interference and reducing the signal-to-interference-plus-noise ratio. Consequently, the aggregate throughput declines. This observation highlights the importance of matching the number of UAVs to the available spectrum resources.
5. Comparison of Methods and Overall Discussion
The three proposed methods target different challenges in cognitive radio networks with unmanned aerial vehicles. To give a clear summary, I list the main features of each method in Table 3.
| Scenario | Challenge | Algorithm | Key Variables | Performance Metric |
|---|---|---|---|---|
| Mixed interference | Co-channel and adjacent-channel interference, unfairness | K-Means + Genetic + PSO + BCD | User association, channel assignment, power | Minimum user SINR |
| High user density | Dense obstacles, limited bandwidth, unfairness | K-Means + Linear precise allocation | UAV positions, user bandwidth | Minimum user throughput |
| Time-varying channel | Unknown channel states, dynamic occupancy | MAPPO with temporary memory | Channel selection action | Sum rate, conflict count |
The first method is deterministic and relies on solving non-convex optimization. It is suitable for offline planning or quasi-static environments. The second method is also deterministic but more lightweight, making it suitable for real-time deployment in dense urban scenarios. The third method uses reinforcement learning and is able to adapt to fast time-varying environments without explicit modeling of the primary network dynamics. In my simulations, all three methods improve the fairness metric compared with random or equal-allocation baselines.
One common finding is that the position of the UAVs plays a critical role in spectrum allocation. In the first two scenarios, optimizing the UAV deployment reduces path loss and decreases the interference to the primary network. In the third scenario, although I did not optimize the UAV trajectories, the channel selection learning implicitly benefits from the fixed favorable positions. In future work, I plan to jointly optimize the 3D trajectories of the UAVs together with the spectrum decisions.
Another important observation is that the primary user protection constraint can significantly affect the achievable performance of the secondary network. For example, in the high-density scenario, if the interference threshold to the primary network is relaxed, the minimum user throughput increases. In the time-varying scenario, the overlay mode allows the secondary network to use only idle channels, and the constraint is automatically satisfied if the sensing is accurate. However, in practice, spectrum sensing may suffer from false alarms and missed detections, which I did not model in this thesis. This is a valuable direction for future research.
6. Conclusion
In this thesis, I investigated spectrum allocation and decision-making methods for multi-UAV communication systems in cognitive radio networks. I addressed three challenging scenarios: mixed interference, high user density, and time-varying channel states. For each scenario, I formulated a mathematical optimization problem and designed an efficient algorithm. The block coordinate descent-based joint optimization significantly improves the minimum user SINR under both co-channel and adjacent-channel interference. The K-Means clustering plus linear precise bandwidth allocation achieves fairness in dense urban environments while protecting the primary network. The MAPPO-based dynamic channel selection with a temporary memory mechanism learns effective access policies under partial observability and reduces frequency conflicts. Simulation results confirmed the effectiveness of the proposed methods in terms of throughput, fairness, and adaptability. These findings provide practical guidelines for the deployment and spectrum management of unmanned aerial vehicles in future cognitive wireless networks.
There are several directions for future work. First, the energy consumption and mission cost of the unmanned aerial vehicle swarm should be incorporated into the objective function to balance performance and resource expenditure. Second, joint optimization of UAV trajectories and spectrum decisions is necessary for mobile scenarios. Third, the resilience against malicious jamming should be studied because UAV communications are increasingly threatened by intelligent jammers. I hope my work contributes to the understanding of spectrum management in UAV-assisted cognitive networks and inspires further research in this important area.