In my research, I study mission-driven spectrum allocation and collaboration for UAV drones operating in complex wireless environments. I focus on how swarms of UAV drones can share limited spectrum, satisfy heterogeneous mission requirements, and maintain reliable communication when malicious interference is present. The central idea of my work is that spectrum decisions should not be separated from mission semantics. A search-and-rescue mission, a reconnaissance mission, a relay mission, and a surveillance mission do not impose the same bandwidth, latency, reliability, or continuity requirements on UAV drones. Therefore, I treat mission demand as the primary driver of spectrum allocation, power control, channel switching, and multi-cluster collaboration.
The motivation is straightforward. UAV drones have become low-cost, flexible, and rapidly deployable platforms for military, public safety, environmental monitoring, logistics, and disaster response. A single UAV drone can perform simple tasks, but many practical missions require clusters of UAV drones to cover wide areas, collect heterogeneous data, and cooperate under dynamic conditions. Clusters of UAV drones improve coverage, fault tolerance, scalability, and mission completion speed. However, clusters of UAV drones also create severe spectrum pressure. Many UAV drones must communicate simultaneously over a finite set of channels. Cluster-head UAV drones must send aggregated data to ground control stations, while member UAV drones must send sensing data to cluster heads. Co-channel interference, spectrum scarcity, dynamic task changes, and adversarial jamming all degrade the performance of UAV drones.
To address these problems, I organize my work around three representative mission scenarios. In the first scenario, UAV drones execute homogeneous multi-task missions over multiple regions. The key issue is fair spectrum sharing when many UAV drones perform similar tasks but belong to different groups. In the second scenario, UAV drones execute heterogeneous multi-task missions, where task types, data volumes, and service requirements vary over time. The key issue is dynamic spectrum reallocation with channel switching costs. In the third scenario, UAV drones execute missions under malicious interference, and control links may be interrupted. The key issue is hierarchical spectrum collaboration between jammer behavior and cluster-head UAV drones. Across all three scenarios, I formulate optimization problems, decompose them into tractable subproblems, and design algorithms that improve throughput, reduce mission completion time, and mitigate spectrum conflicts.
I summarize the main challenges and my corresponding design responses in Table 1. The table emphasizes that spectrum allocation for UAV drones is not only a radio resource management problem, but also a mission coordination problem.
| Challenge for UAV drones | Observed consequence | My design response |
|---|---|---|
| Limited spectrum and many UAV drones | Severe co-channel interference and unfair group throughput | Group-based spectrum reuse, joint channel and power optimization, max-min fairness |
| Heterogeneous and time-varying missions | Static allocation becomes inefficient, mission completion is delayed | Dynamic spectrum allocation with channel switching cost and block coordinate descent |
| Malicious jamming and weak information exchange | Control links break, UAV drones cannot coordinate globally | Stackelberg game, Q-learning for jammer, multi-agent proximal policy optimization for cluster-head UAV drones |
| Different transmission roles of UAV drones | Cluster heads and member UAV drones interfere differently | Differentiated modeling of cluster-head UAV drones and member UAV drones |
| Mission continuity | Frequent channel switching wastes time and energy | Switching penalty in utility and dynamic allocation models |
Throughout my work, the term UAV drones refers to flying wireless nodes that may act as cluster heads, sensing members, relays, or data sources. I assume that UAV drones cooperate within groups, while groups may compete or interfere with each other. The spectrum is divided into channels. Cluster-head UAV drones send data to a ground control station, and member UAV drones send data to their cluster heads. Because the number of available channels is limited, I allow underlay spectrum sharing. This means that member UAV drones in one group may reuse channels used by cluster-head UAV drones in other groups, provided that interference remains manageable. This underlay approach improves spectral efficiency but requires careful interference modeling and power control.
I now present the unified notation that I use across scenarios. Table 2 defines the main symbols. I keep the notation consistent so that the three mission scenarios can be compared and generalized.
| Symbol | Meaning |
|---|---|
| \(\mathcal{G}=\{1,\dots,G\}\) | Set of UAV drone groups or clusters |
| \(\mathcal{N}=\{1,\dots,N\}\) | Set of member UAV drones in each group |
| \(\mathcal{M}=\{1,\dots,M\}\) | Set of available channels |
| \(\mathcal{H}=\{1,\dots,H\}\) | Set of heterogeneous mission types |
| \(\theta^{G}_{g,m}\) | Channel assignment indicator for cluster-head UAV drone \(g\) on channel \(m\) |
| \(\theta^{N,g}_{n,m}\) | Channel assignment indicator for member UAV drone \(n\) in group \(g\) on channel \(m\) |
| \(P^{G}_{g,m}\) | Transmit power of cluster-head UAV drone \(g\) on channel \(m\) |
| \(P^{N,g}_{n,m}\) | Transmit power of member UAV drone \(n\) in group \(g\) on channel \(m\) |
| \(B_0\) | Bandwidth of one channel |
| \(\gamma^{G}_{g,m}\), \(\gamma^{N,g}_{n,m}\) | Signal-to-interference-plus-noise ratio for cluster-head and member UAV drones |
| \(R^{G}_{g,m}\), \(R^{N,g}_{n,m}\) | Achievable rates of cluster-head and member UAV drones |
| \(R_g\) | Effective throughput of group \(g\) |
| \(\mu\) | Channel switching cost factor |
| \(D_{g,h}\) | Data volume required by group \(g\) for mission type \(h\) |
| \(T_{g,h}\) | Mission completion time of group \(g\) for mission type \(h\) |
| \(U_G\), \(U_J\) | Utility of cluster-head UAV drones and jammer |
In the first scenario, I consider homogeneous multi-task missions. UAV drones are divided into groups according to mission regions. Each group contains one cluster-head UAV drone and several member UAV drones. All cluster-head UAV drones use distinct channels to communicate with the ground control station, which avoids co-channel interference among cluster heads on the downlink/uplink to the station. However, member UAV drones may reuse the channels of other cluster heads in an underlay manner. This creates cross-group interference. The effective throughput of a group is limited by the weaker of two links: the member-to-cluster-head link and the cluster-head-to-ground-station link. Therefore, I model the group throughput as the minimum of the two aggregated rates.
The achievable rate of cluster-head UAV drone \(g\) on channel \(m\) is
$$
R^{G}_{g,m}=B_0\theta^{G}_{g,m}\log_2\left(1+\gamma^{G}_{g,m}\right).
\tag{1}
$$
The signal-to-interference-plus-noise ratio at the ground control station for cluster-head UAV drone \(g\) on channel \(m\) is
$$
\gamma^{G}_{g,m}
=
\frac{P^{G}_{g,m}d_{g,O}^{-\alpha}}
{\sigma_0+\sum_{j\in\mathcal{G},j\ne g}\sum_{n=1}^{N}
P^{N,j}_{n,m}d_{n,O}^{-\alpha}\theta^{N,j}_{n,m}}.
\tag{2}
$$
Here, \(d_{g,O}\) is the distance from cluster-head UAV drone \(g\) to the ground control station, \(d_{n,O}\) is the distance from member UAV drone \(n\) to the ground control station, \(\alpha\) is the path-loss exponent, and \(\sigma_0\) is the noise power. The interference term captures cross-group reuse by member UAV drones.
Similarly, the signal-to-interference-plus-noise ratio at cluster-head UAV drone \(g\) for member UAV drone \(n\) on channel \(m\) is
$$
\gamma^{N,g}_{n,m}
=
\frac{P^{N,g}_{n,m}d_{n,g}^{-\alpha}}
{\sigma_0+
\sum_{j\ne g}P^{G}_{j,m}d_{j,g}^{-\alpha}\theta^{G}_{j,m}
+
\sum_{j\ne g}\sum_{i\ne n}P^{N,j}_{i,m}d_{i,g}^{-\alpha}\theta^{N,j}_{i,m}}.
\tag{3}
$$
The rate of member UAV drone \(n\) in group \(g\) on channel \(m\) is
$$
R^{N,g}_{n,m}=B_0\theta^{N,g}_{n,m}\log_2\left(1+\gamma^{N,g}_{n,m}\right).
\tag{4}
$$
The effective throughput of group \(g\) is then
$$
R_g=
\min\left\{
\sum_{m=1}^{M}R^{G}_{g,m},
\sum_{n=1}^{N}\sum_{m=1}^{M}R^{N,g}_{n,m}
\right\}.
\tag{5}
$$
For homogeneous multi-task missions, I aim to maximize the minimum group throughput. This ensures fairness among groups of UAV drones. The optimization problem is
$$
\begin{aligned}
\max_{\theta^{G},P^{G},\theta^{N},P^{N}} \quad & \min_{g\in\mathcal{G}} R_g \\
\text{s.t.} \quad
& \theta^{G}_{g,m},\theta^{N,g}_{n,m}\in\{0,1\}, \quad \forall g,m,n, \\
& \theta^{G}_{g,m}+\theta^{N,g}_{n,m}\le 1, \quad \forall g,m,n, \\
& \sum_{g=1}^{G}\theta^{G}_{g,m}\le 1, \quad \forall m, \\
& \sum_{m=1}^{M}\theta^{G}_{g,m}=1, \quad \forall g, \\
& \sum_{m=1}^{M}\theta^{N,g}_{n,m}=1, \quad \forall g,n, \\
& 0\le P^{G}_{g,m}\le P^{G}_{\max}, \quad \forall g,m, \\
& 0\le P^{N,g}_{n,m}\le P^{N}_{\max}, \quad \forall g,n,m, \\
& R_g\ge \eta, \quad \forall g.
\end{aligned}
\tag{6}
$$
Problem (6) is a mixed-integer nonlinear programming problem. It contains binary channel assignment variables and continuous power variables. I decompose it into channel allocation and power allocation subproblems. For channel allocation, I design an improved genetic algorithm. For power allocation, I use geometric programming and convex transformation. I summarize the solution architecture in Table 3.
| Subproblem | Variables | Method | Purpose |
|---|---|---|---|
| Channel allocation | \(\theta^{G}\), \(\theta^{N}\) | Improved genetic algorithm with elite retention and constraint repair | Assign channels to cluster-head and member UAV drones simultaneously |
| Cluster-head power | \(P^{G}\) | Log transformation and geometric programming | Optimize transmit power of cluster-head UAV drones |
| Member power | \(P^{N}\) | Log transformation and geometric programming | Optimize transmit power of member UAV drones |
| Joint iteration | All variables | Alternating optimization | Obtain a near-optimal joint spectrum and power strategy |
The improved genetic algorithm uses binary encoding. A chromosome stores the channel assignment of all cluster-head UAV drones and all member UAV drones. The fitness function is the max-min group throughput. I use roulette-wheel selection, single-point crossover, mutation, and elite retention. Constraint repair ensures that each cluster head uses exactly one channel, member UAV drones use one channel, and no intra-group conflict occurs. This algorithm is important because it searches the discrete channel space while preserving feasibility.
For power allocation, I introduce logarithmic variables
$$
\tilde P^{G}_{g,m}=\log_2\left(P^{G}_{g,m}\right), \quad
\tilde P^{N,g}_{n,m}=\log_2\left(P^{N,g}_{n,m}\right).
\tag{7}
$$
After transformation, the non-convex rate constraints become convex or concave forms suitable for standard solvers. I alternate between channel allocation and power allocation until convergence. Simulation parameters are listed in Table 4.
| Parameter | Value |
|---|---|
| Mission area | \(2000\times 2000\ \text{m}^2\) |
| Number of UAV drone groups \(G\) | \(\{4,5,6,7,8\}\) |
| Member UAV drones per group \(N\) | \(\{2,3,4\}\) |
| Available channels \(M\) | \(\{4,6,8,10\}\) |
| Channel bandwidth \(B_0\) | \(1\ \text{MHz}\) |
| UAV drone altitude \(H\) | \(100\ \text{m}\) |
| Noise power \(\sigma_0\) | \(-90\ \text{dBm}\) |
| Path-loss exponent \(\alpha\) | \(2\) |
| Maximum cluster-head power \(P^{G}_{\max}\) | \(4\ \text{W}\) |
| Maximum member power \(P^{N}_{\max}\) | \(2\ \text{W}\) |
The convergence behavior shows that the max-min group throughput increases monotonically and stabilizes after several iterations. When the number of groups, member UAV drones, or channels increases, the algorithm requires more iterations because the search space grows. The results also show that increasing available channels reduces co-channel interference and improves group throughput. When the number of available channels is large enough, adding more member UAV drones can increase group throughput because the interference penalty is smaller. When the number of groups is large relative to the number of channels, adding member UAV drones may reduce the minimum group throughput because cross-group interference becomes dominant.
I compare the proposed joint optimization with three baselines: random allocation, channel-only optimization, and power-only optimization. The joint method achieves the best max-min throughput and provides better fairness among UAV drone groups. The gain over baselines is at least about sixteen percent in the tested scenarios. This confirms that channel allocation and power control must be optimized together for UAV drones in homogeneous multi-task missions.
In the second scenario, I study heterogeneous multi-task missions. UAV drones execute different mission types over time. For example, one mission may require a small amount of telemetry data, while another may require high-volume image or video transmission. The data volume, service priority, and continuity requirement may change across mission phases. I divide time into slots. In each slot, the task type and spectrum allocation can be adjusted. I model the data volume for group \(g\) and mission type \(h\) as \(D_{g,h}\). The completion time for that mission type is
$$
T_{g,h}=\frac{D_{g,h}}{R_g}.
\tag{8}
$$
Because all groups must complete their missions, the overall mission completion time is determined by the slowest group:
$$
T_{\max}=\max_{g\in\mathcal{G}}\sum_{h=1}^{H}T_{g,h}.
\tag{9}
$$
My objective is to minimize \(T_{\max}\) by jointly optimizing channel assignment and transmit power for UAV drones. I also introduce a channel switching cost. When a cluster-head UAV drone or a member UAV drone changes its channel, it loses part of the slot for link establishment and synchronization. This is important for UAV drones because frequent switching wastes energy and reduces effective transmission time. The switching loss for a cluster-head UAV drone on channel \(m\) in slot \(t\) is
$$
R^{\text{loss},G}_{g,m}(t)=\mu R^{G}_{g,m}(t)\theta^{\text{switch},G}_{g,m}(t),
\tag{10}
$$
where \(\mu\) is the switching cost factor and \(\theta^{\text{switch},G}_{g,m}(t)\) indicates whether a switch occurs. Similarly, for a member UAV drone,
$$
R^{\text{loss},N,g}_{n,m}(t)=\mu R^{N,g}_{n,m}(t)\theta^{\text{switch},N,g}_{n,m}(t).
\tag{11}
$$
The actual throughput of a cluster-head UAV drone and a member UAV drone after switching loss is
$$
R^{\text{actual},G}_{g,m}(t)=R^{G}_{g,m}(t)-R^{\text{loss},G}_{g,m}(t),
\tag{12}
$$
$$
R^{\text{actual},N,g}_{n,m}(t)=R^{N,g}_{n,m}(t)-R^{\text{loss},N,g}_{n,m}(t).
\tag{13}
$$
The effective group throughput in a time interval is
$$
R^{\text{actual}}_g=
\min\left\{
\sum_{t=1}^{T}\sum_{m=1}^{M}R^{\text{actual},G}_{g,m}(t),
\sum_{t=1}^{T}\sum_{n=1}^{N}\sum_{m=1}^{M}R^{\text{actual},N,g}_{n,m}(t)
\right\}.
\tag{14}
$$
The optimization problem for heterogeneous multi-task UAV drones is
$$
\begin{aligned}
\min_{\theta^{G},P^{G},\theta^{N},P^{N}} \quad & T_{\max} \\
\text{s.t.} \quad
& \theta^{G}_{g,m}(t),\theta^{N,g}_{n,m}(t)\in\{0,1\}, \\
& \theta^{G}_{g,m}(t)+\theta^{N,g}_{n,m}(t)\le 1, \\
& \sum_{g=1}^{G}\theta^{G}_{g,m}(t)\le 1, \\
& \sum_{m=1}^{M}\theta^{G}_{g,m}(t)=1, \\
& \sum_{m=1}^{M}\theta^{N,g}_{n,m}(t)=1, \\
& 0\le P^{G}_{g,m}(t)\le P^{G}_{\max}, \\
& 0\le P^{N,g}_{n,m}(t)\le P^{N}_{\max}, \\
& T_{g,h}=\frac{D_{g,h}}{R^{\text{actual}}_g}, \\
& T_{\max}\ge \sum_{h=1}^{H}T_{g,h}, \quad \forall g.
\end{aligned}
\tag{15}
$$
Problem (15) is also a mixed-integer nonlinear programming problem. I solve it using block coordinate descent. At each block, I optimize one set of variables while fixing the others. First, I optimize channel assignment using an improved genetic algorithm. Then, I optimize cluster-head power using geometric programming and gradient projection. Next, I optimize member power using the same approach. I iterate these steps until the mission completion time converges. Table 5 summarizes the algorithmic blocks.
| Block | Fixed variables | Optimized variables | Method |
|---|---|---|---|
| Channel block | \(P^{G}\), \(P^{N}\) | \(\theta^{G}\), \(\theta^{N}\) | Improved genetic algorithm |
| Cluster-head power block | \(\theta^{G}\), \(\theta^{N}\), \(P^{N}\) | \(P^{G}\) | Geometric programming and gradient projection |
| Member power block | \(\theta^{G}\), \(\theta^{N}\), \(P^{G}\) | \(P^{N}\) | Geometric programming and gradient projection |
| Global iteration | None | All variables | Block coordinate descent |
The simulation parameters for heterogeneous missions are shown in Table 6. I consider two mission types. The first type has data volume randomly distributed in \([100,200]\ \text{MB}\). The second type has data volume randomly distributed in \([1024,2024]\ \text{MB}\). This captures the difference between low-volume and high-volume missions of UAV drones.
| Parameter | Value |
|---|---|
| Mission area | \(2000\times 2000\ \text{m}^2\) |
| Number of UAV drone groups \(G\) | \(\{4,5,6,7,8\}\) |
| Member UAV drones per group \(N\) | \(\{2,3,4\}\) |
| Available channels \(M\) | \([4,12]\) |
| Channel bandwidth \(B_0\) | \(1\ \text{MHz}\) |
| UAV drone altitude \(H\) | \(100\ \text{m}\) |
| Noise power \(\sigma_0\) | \(-90\ \text{dBm}\) |
| Path-loss exponent \(\alpha\) | \(2\) |
| Channel switching cost \(\mu\) | \(0.2\) |
| Maximum cluster-head power \(P^{G}_{\max}\) | \(4\ \text{W}\) |
| Maximum member power \(P^{N}_{\max}\) | \(2\ \text{W}\) |
The results show that the joint optimization method converges in about five iterations, while power-only optimization requires more iterations and channel-only optimization converges quickly but to a worse value. The joint method reduces mission completion time by about fourteen percent compared with power-only optimization and about sixteen percent compared with channel-only optimization. When compared with random allocation, the improvement can reach forty percent. I also observe that mission completion time increases with the number of groups because more UAV drones compete for the same spectrum. Increasing the number of channels reduces completion time because co-channel interference decreases. The channel switching cost is beneficial when it prevents unnecessary switching; with switching cost, the UAV drones adopt more stable channel strategies and reduce spectrum collaboration overhead.
In the third scenario, I study mission execution under malicious interference. Cluster-head UAV drones must send data to a ground control station, but a jammer attempts to disrupt their links. The control link may be interrupted, so centralized coordination is not reliable. I model the interaction as a Stackelberg game. The jammer is the leader and chooses which channel to jam. The cluster-head UAV drones are followers and choose their transmission channels after observing the jamming strategy or inferring it from local measurements. This hierarchical structure captures the asymmetric information and adversarial nature of the scenario. The equilibrium condition is that neither the jammer nor any cluster-head UAV drone can improve its utility by unilaterally changing its strategy.
I model the air-to-ground channel using line-of-sight and non-line-of-sight components. The average channel loss for cluster-head UAV drone \(g\) in slot \(t\) is
$$
h_g(t)=P^{LoS}_g(t)h^{LoS}_g(t)+P^{NLoS}_g(t)h^{NLoS}_g(t).
\tag{16}
$$
The line-of-sight probability is
$$
P^{LoS}_g(t)=\frac{1}{1+b_1\exp\left(-b_2\left(\xi_g(t)-b_1\right)\right)},
\tag{17}
$$
where \(\xi_g(t)\) is the elevation angle from cluster-head UAV drone \(g\) to the ground control station. The non-line-of-sight probability is
$$
P^{NLoS}_g(t)=1-P^{LoS}_g(t).
\tag{18}
$$
The signal-to-interference-plus-noise ratio of cluster-head UAV drone \(g\) on channel \(m\) in slot \(t\) is
$$
\gamma_{g,m}(t)=
\frac{
P_G 10^{-h_g(t)/10}
}{
\sum_{i\ne g}P_G\theta_{i,m}(t)10^{-h_i(t)/10}
+
P_J\theta_{J,m}(t)10^{-h_J(t)/10}
+
\sigma
}.
\tag{19}
$$
The achievable rate is
$$
R_{g,m}(t)=B_0\log_2\left(1+\gamma_{g,m}(t)\right).
\tag{20}
$$
I also include channel switching loss for cluster-head UAV drones. The actual rate is
$$
R^{\text{actual}}_{g,m}(t)=R_{g,m}(t)-R^{\text{loss}}_{g,m}(t),
\tag{21}
$$
where
$$
R^{\text{loss}}_{g,m}(t)=\frac{\mu}{2}\left(
\left|\theta_{g,m}(t)-\theta_{g,m}(t-1)\right|
+
\left|\theta_{g,m}(t)-\theta_{g,m}(t-1)\right|
\right)R_{g,m}(t).
\tag{22}
$$
The utility of the cluster-head UAV drones is the sum of actual rates:
$$
U_G=\sum_{g=1}^{G}\sum_{t=1}^{T}\sum_{m=1}^{M}R^{\text{actual}}_{g,m}(t).
\tag{23}
$$
The utility of the jammer is the negative of the total UAV drone rate:
$$
U_J=-\sum_{g=1}^{G}\sum_{t=1}^{T}\sum_{m=1}^{M}R_{g,m}(t).
\tag{24}
$$
The Stackelberg equilibrium satisfies
$$
U_J\left(C_J^*,C_G^*\right)\ge U_J\left(C_J,C_G^*\right),
\tag{25}
$$
$$
U_G\left(C_J^*,C_G^*\right)\ge U_G\left(C_J^*,C_G\right).
\tag{26}
$$
For the leader subproblem, the jammer uses Q-learning to update its channel selection probability. The Q-value update is
$$
Q_J(\tau+1)=(1-\alpha_\tau)Q_J(\tau)+\alpha_\tau U_J(\tau).
\tag{27}
$$
The channel selection probability follows a softmax rule:
$$
p_{J,i}(\tau+1)=
\frac{
\exp\left(Q_{J,i}(\tau+1)/\Xi\right)
}{
\sum_{j\in\mathcal M}\exp\left(Q_{J,j}(\tau+1)/\Xi\right)
}.
\tag{28}
$$
For the follower subproblem, I model the cluster-head UAV drones as agents in a partially observable Markov game. The tuple is
$$
\left\langle \mathcal S,\mathcal A,\mathcal P,\mathcal R,\mathcal O \right\rangle.
\tag{29}
$$
Each cluster-head UAV drone observes the received signal strength indicator of each channel. The RSSI observation for UAV drone \(g\) on channel \(m\) in slot \(t\) is
$$
I_{g,m}(t)=
P_G\theta_{g,m}(t-1)h_g(t-1)
+
\sum_{j\ne g}P_G\theta_{j,m}(t-1)h_j(t-1)
+
P_J\theta_{J,m}(t-1)h_J(t-1)
+
\sigma.
\tag{30}
$$
The reward of agent \(g\) is its actual sum rate:
$$
r_g^t=\sum_{m=1}^{M}R^{\text{actual}}_{g,m}(t).
\tag{31}
$$
The return is
$$
Z_g^t=\sum_{l=0}^{\infty}\gamma^l r_g^{t+l}.
\tag{32}
$$
I use multi-agent proximal policy optimization for distributed spectrum collaboration. The actor loss is
$$
L_{\theta}^{Actor}=
\frac{1}{DG}\sum_{i=1}^{D}\sum_{g=1}^{G}
\left[
\min\left(
\varsigma_{i,g}A_i^g,
\operatorname{clip}\left(\varsigma_{i,g},1-\epsilon,1+\epsilon\right)A_i^g
\right)
+
\varkappa S\left(\pi_{\theta_g}\right)
\right],
\tag{33}
$$
where the probability ratio is
$$
\varsigma_{i,g}=
\frac{
\pi_{\theta_g}\left(a_i^g\mid o_i^g\right)
}{
\pi_{\theta_g^{old}}\left(a_i^g\mid o_i^g\right)
}.
\tag{34}
$$
The advantage function is
$$
A_i^g=
\sum_{l=0}^{\infty}(\gamma\lambda)^l
\left[
r_{i+l}^g+\gamma V_{\phi_g}\left(s_{i+l+1}\right)-V_{\phi_g}\left(s_{i+l}\right)
\right].
\tag{35}
$$
The critic loss is
$$
L_{\phi}^{Critic}=
\frac{1}{DG}\sum_{i=1}^{D}\sum_{g=1}^{G}
\max\left[
\left(
V_{\phi_g}(s_i)-\hat r_i^g
\right)^2,
\left(
\operatorname{clip}\left(V_{\phi_g}(s_i),V_{\phi_g^{old}}(s_i)-\epsilon,V_{\phi_g^{old}}(s_i)+\epsilon\right)-\hat r_i^g
\right)^2
\right].
\tag{36}
$$
The hierarchical adversarial algorithm alternates between jammer learning and UAV drone learning. The jammer updates its jamming policy to maximize interference. The cluster-head UAV drones update their channel selection policies to maximize actual throughput while reducing switching. The process continues until the UAV drone utility stabilizes or the maximum number of iterations is reached. Table 7 summarizes the jamming-scenario parameters.
| Parameter | Value |
|---|---|
| Number of UAV drone groups \(G\) | \(5\) |
| Carrier frequency \(f_c\) | \(2\ \text{GHz}\) |
| Channel bandwidth \(B_0\) | \(1\ \text{MHz}\) |
| UAV drone altitude \(H\) | \(100\ \text{m}\) |
| Noise power \(\sigma\) | \(-114\ \text{dBm}\) |
| GAE parameter \(\lambda\) | \(0.95\) |
| Channel switching cost \(\mu\) | \(0.1\) |
| Cluster-head power \(P_G\) | \(23\ \text{dBm}\) |
| Jammer power \(P_J\) | \(60\ \text{dBm}\) |
| Environment constants \(b_1,b_2\) | \(0.13,11.9\) |
| Non-line-of-sight excess loss | \(20\ \text{dB}\) |
| Discount factor \(\gamma\) | \(0.99\) |
| Clip parameter \(\epsilon\) | \(0.2\) |
| Actor learning rate | \(5\times 10^{-4}\) |
| Critic learning rate | \(5\times 10^{-4}\) |
The simulation results show that when more channels are available, the average throughput reward of UAV drones increases and converges to a higher value. This is because more channels reduce mutual interference among cluster-head UAV drones and provide more opportunities to avoid the jammer. When fewer channels are available, the policy converges faster because the action space is smaller. The number of spectrum conflicts also decreases rapidly with training. With more channels, the conflict count is slightly lower, but with fewer channels, the UAV drones learn to coordinate quickly and reach a stable low-conflict state.
I also study the effect of the channel switching cost factor. When the switching cost is zero, cluster-head UAV drones change channels frequently to chase momentary rate improvements. When the switching cost is positive, the number of channel switches decreases significantly. This demonstrates that including switching cost in the utility function improves spectrum collaboration stability and reduces unnecessary control overhead. Table 8 summarizes the performance outcomes across the three scenarios.
| Scenario | Main objective | Key algorithm | Observed benefit |
|---|---|---|---|
| Homogeneous multi-task UAV drones | Maximize minimum group throughput | Improved genetic algorithm and convex power optimization | Fairness improved and throughput improved by at least about sixteen percent over baselines |
| Heterogeneous multi-task UAV drones | Minimize maximum mission completion time | Block coordinate descent, genetic algorithm, gradient projection | Completion time reduced by about sixteen to forty percent depending on baseline |
| Jammed mission UAV drones | Maximize UAV drone utility under jamming | Stackelberg game, Q-learning, MAPPO | Higher reward, fewer spectrum conflicts, stable channel selection |
From a mission-driven perspective, the three scenarios share a common principle. Spectrum allocation for UAV drones should adapt to the mission state. In homogeneous missions, fairness is the dominant concern because groups perform similar tasks and no group should be starved. In heterogeneous missions, mission completion time is the dominant concern because different tasks have different data volumes and continuity requirements. In jammed missions, robustness and decentralized collaboration are the dominant concerns because the jammer and link interruptions make centralized control unreliable. My methods reflect these differences by using different objectives, constraints, and learning architectures.
I also emphasize the role of differentiated treatment of cluster-head UAV drones and member UAV drones. Cluster-head UAV drones aggregate data and communicate with the ground control station. Member UAV drones collect data and communicate with cluster heads. Their interference patterns, power budgets, and service requirements differ. Treating them uniformly leads to inefficient spectrum use. In my models, I assign separate channel matrices, power variables, and utility terms to cluster-head and member UAV drones. This differentiated design is one of the key reasons for performance improvement.
The use of channel switching cost is another important element. For UAV drones, switching channels is not free. It consumes time, energy, and control signaling. In dynamic missions, frequent switching can erase the gains from spectrum adaptation. By including switching cost, I encourage stable allocation while still allowing adaptation when the mission changes. This is especially important for heterogeneous multi-task missions and jamming scenarios. The switching cost acts as a regularizer that balances responsiveness and stability.
For the jamming scenario, the Stackelberg formulation captures the adversarial hierarchy. The jammer moves first or maintains a dominant strategy, and the UAV drones respond. Q-learning is suitable for the jammer because the jammer can observe the aggregate effect of its channel choices. MAPPO is suitable for cluster-head UAV drones because they operate in a partially observable environment, cannot rely on a central controller, and must learn cooperative channel selection from local RSSI observations. The centralized training and decentralized execution structure of MAPPO allows UAV drones to learn team coordination while executing policies independently.
The simulation results support several practical conclusions. First, joint optimization of channel and power is superior to optimizing either resource alone. Second, dynamic spectrum allocation with switching cost reduces mission completion time and prevents unnecessary reconfiguration. Third, hierarchical game learning enables UAV drones to mitigate jamming and reduce mutual conflicts even when control links are unreliable. Fourth, the number of available channels strongly affects performance; more channels generally improve throughput and reduce conflicts, but the marginal benefit may diminish because of power limits and interference.
I now summarize the main algorithmic components in Table 9. This table clarifies how each component addresses a specific difficulty in spectrum allocation and collaboration for UAV drones.
| Algorithmic component | Used for | Why it is suitable |
|---|---|---|
| Improved genetic algorithm | Discrete channel assignment | Handles binary variables, multiple constraints, and non-convex search space |
| Geometric programming | Non-convex power optimization | Transforms rate expressions into convex forms under log variables |
| Gradient projection | Continuous power updates | Projects updates onto power constraints and improves convergence |
| Block coordinate descent | Joint channel and power optimization | Decomposes coupled variables into tractable blocks |
| Stackelberg game | Jammer-UAV drone interaction | Captures leader-follower adversarial behavior |
| Q-learning | Jammer channel selection | Learns optimal jamming policy in discrete channel space |
| MAPPO | Distributed UAV drone spectrum collaboration | Supports partial observability, cooperative learning, and decentralized execution |
For future work, I see several extensions. First, UAV drones may operate in highly dynamic three-dimensional environments where altitude, speed, and mobility affect channel quality. I plan to incorporate mobility prediction and spatial-temporal spectrum maps. Second, heterogeneous missions may have priorities, deadlines, and random arrivals. I plan to model mission priority and use online learning to adapt quickly to new tasks. Third, jamming techniques may become more intelligent and adaptive. I plan to integrate anti-jamming techniques such as spread spectrum, encrypted control channels, and robust reinforcement learning. Fourth, UAV drones may collaborate with ground base stations, satellites, or other aerial platforms. I plan to extend the hierarchical game and multi-agent learning framework to mixed aerial-terrestrial networks. Fifth, energy constraints are critical for UAV drones. I plan to include energy consumption, battery state, and flight control in the spectrum allocation objective.
In conclusion, my research develops mission-driven spectrum allocation and collaboration methods for UAV drones. I formulate max-min fairness for homogeneous multi-task missions, dynamic completion-time minimization for heterogeneous multi-task missions, and Stackelberg game-based spectrum collaboration for jammed missions. I design joint channel and power optimization algorithms, incorporate channel switching costs, and use hierarchical learning for adversarial scenarios. The results show that mission-aware spectrum decisions improve throughput, reduce completion time, and enhance robustness for UAV drones. I believe these methods provide a useful foundation for future autonomous UAV drone networks that must operate under severe spectrum scarcity and complex mission dynamics.