UAV-assisted Computation Offloading Optimization with Service Caching

In recent years, the rapid advancement of mobile communication technologies has led to the emergence of numerous computation-intensive, latency-sensitive, and data-heavy applications, such as immersive virtual reality, autonomous driving, and tactile Internet. Mobile edge computing (MEC) has been regarded as a promising paradigm to bring computing, storage, and network resources closer to end users, thereby reducing service latency and improving quality of experience. However, conventional MEC servers attached to fixed ground base stations may not provide seamless coverage in remote or disaster-affected areas where communication infrastructures are absent or damaged. To overcome this limitation, the deployment of an unmanned aerial vehicle (UAV) as an aerial edge server has attracted extensive attention. The mobility and flexible deployment of an unmanned aerial vehicle enable on-demand computation offloading, effective coverage extension, and adaptive resource provisioning in dynamic environments. In this thesis, I focus on multi-UAV-assisted MEC systems and investigate the integration of service caching with computation offloading. Three different optimization frameworks are designed to minimize task completion delay, user energy consumption, and overall system energy consumption, respectively. Throughout this article, I emphasize the role of the unmanned aerial vehicle trajectory, resource allocation, service caching, and non-orthogonal multiple access (NOMA) in achieving efficient edge computing.

Background and Motivation

Mobile edge computing pushes cloud capabilities to the network edge, which is critical for supporting ultra-reliable and low-latency services. In a typical MEC system, a mobile device can offload its computational tasks to an edge server that is geographically close to the device. The edge server executes the tasks and returns the results, avoiding the long transmission delay of mobile cloud computing. Nonetheless, the coverage of a fixed edge server is limited by its location and wireless transmission range. When the number of users is large and their locations are geographically dispersed, a single ground base station cannot guarantee high-quality service to all devices. An unmanned aerial vehicle equipped with an MEC server can dynamically reposition itself to improve channel conditions and provide computation services directly to ground terminals. A multi-unmanned aerial vehicle system can further enhance scalability and reliability.

Another crucial aspect in MEC is service caching. Computing tasks often require specific service programs, such as libraries, executables, or virtual machine images, to be present on the edge server. If a requested service program is not cached, the server has to fetch it from the remote cloud, which introduces additional delay and communication overhead. Service caching, on the other hand, pre-stores popular service programs at the edge server, so that when a task arrives, the server can execute it immediately without a cold-start delay. However, the storage capacity of a UAV-mounted server is limited, and ground terminals also have limited cache capabilities. Therefore, it is essential to jointly decide which services to cache and how to offload tasks in order to fully exploit the benefits of service caching in UAV-assisted MEC networks.

The primary objective of my research is to design efficient algorithms that cooperatively optimize task offloading decisions, communication and computation resource allocation, UAV trajectories, user grouping, and service caching policies. Three optimization problems are formulated and solved using convex optimization, successive convex approximation, matching theory, and block coordinate descent methods. Simulation results verify the effectiveness of the proposed algorithms.

Chapter 3: Service-Caching-Based UAV-Assisted Task Offloading and Resource Allocation

3.1 System Model

I consider a multi-unmanned aerial vehicle assisted MEC system in a square area, where a set of \(K\) rotary-wing UAVs and \(M\) ground user terminals are deployed. Each UAV is equipped with an MEC server and is capable of caching a limited number of service programs. The ground users are assumed to be randomly located, and each user requests a computation task belonging to one of \(S\) service types. The users access the nearest UAV within their communication range. If the user does not have the requested service cached locally, it may download the service from the UAV or execute the task after fetching the program. If the UAV does not cache the requested service, it must download the service from the remote cloud through the wireless backhaul link.

For the communication model, I adopt orthogonal frequency division multiple access (OFDMA) on the access links. The total bandwidth \(B\) is equally divided among the users served by a UAV. Let \(M_k\) denote the number of users associated with UAV \(k\). The channel gain between user \(m\) and UAV \(k\) is given by \(h_{m,k} = \frac{h_0}{H^2 + \| \mathbf{q}_k – \mathbf{q}_m \|^2}\), where \(h_0\) is the channel gain at the reference distance, \(H\) is the UAV altitude, and \(\mathbf{q}_k, \mathbf{q}_m\) are the horizontal coordinates of UAV \(k\) and user \(m\), respectively. The uplink and downlink rates can be expressed as:

$$r_{m,k}^{\mathrm{up}} = \frac{B}{M_k} \log_2\left(1 + \frac{h_{m,k} p_m}{\frac{B}{M_k} N_0}\right),$$
$$r_{m,k}^{\mathrm{down}} = \frac{B}{M_k} \log_2\left(1 + \frac{h_{m,k} p_k}{\frac{B}{M_k} N_0}\right),$$

where \(p_m\) is the transmit power of user \(m\), \(p_k\) is the UAV transmission power, and \(N_0\) is the noise power spectral density.

The caching model is built upon a popularity matrix. Let \(p_{m,s}\) be the probability that user \(m\) requests service \(s\). The local popularity of service \(s\) at UAV \(k\) is computed as \(\bar{p}_{k,s} = \frac{1}{M_k} \sum_{m \in \mathcal{M}_k} p_{m,s}\). UAVs cache service programs according to the local popularity, while users cache a small number of services randomly.

For the computation model, each task is characterized by its input size \(l_m\) and the required CPU cycles \(C_m\). If the task is executed locally, the local execution time is \(T_m^l = C_m / f_m\), where \(f_m\) is the user’s CPU frequency. The local energy consumption is \(E_m^l = \kappa f_m^2 C_m\), with \(\kappa\) being the effective switched capacitance coefficient. If the task is offloaded to UAV \(k\), the transmission time is \(T_{m,k}^{\mathrm{u}} = l_m / r_{m,k}^{\mathrm{up}}\), and the uplink energy is \(E_{m,k}^{\mathrm{u}} = p_m T_{m,k}^{\mathrm{u}}\). The execution on the UAV takes \(T_{m,k}^{e} = C_m / f_{m,k}\), where \(f_{m,k}\) is the CPU frequency allocated by UAV \(k\) to user \(m\). If the required service is not cached at the UAV, an additional download time \(T_{s}^{\mathrm{back}} = c_s / R_b\) is incurred, where \(c_s\) is the size of the service program and \(R_b\) is the backhaul capacity.

I further introduce binary variables \(c_{m,k} \in \{0,1\}\) indicating whether user \(m\) offloads its task to UAV \(k\), and \(b_{m,k} \in \{0,1\}\) indicating the access association. The total task completion delay for user \(m\) is given by:

$$T_m = \sum_{k=1}^{K} \left(1 – c_{m,k}\right) T_m^l + \sum_{k=1}^{K} c_{m,k} \left( T_{m,k}^{\mathrm{u}} + T_{m,k}^{e} + (1-a_{k,s}) T_{s}^{\mathrm{back}} \right),$$

where \(a_{k,s}\) denotes whether UAV \(k\) has cached service \(s\). The corresponding user energy consumption is:

$$E_m = \sum_{k=1}^{K} (1-c_{m,k}) E_m^l + \sum_{k=1}^{K} c_{m,k} E_{m,k}^{\mathrm{u}}.$$

The UAV energy consumption consists of the computation energy for offloaded tasks and the energy for transmitting the service programs to users, which is constrained by a maximum energy budget \(E_k^{\max}\).

3.2 Problem Formulation

With the objective of minimizing the total task completion delay of all users, the optimization problem is formulated as:

$$\mathcal{P}_1: \min_{C,F,Q} \sum_{m=1}^{M} T_m$$
$$\text{s.t. } C_1: T_m \leq D_m^{\max}, \quad C_2: E_m \leq E_m^{\max}, \quad C_3: E_k \leq E_k^{\max},$$
$$C_4: \sum_{m=1}^{M} f_{m,k} \leq F_k^{\max}, \quad C_5: c_{m,k} \leq b_{m,k},$$
$$C_6: \sum_{k=1}^{K} b_{m,k} = 1, \quad C_7: \sum_{m=1}^{M} b_{m,k} \leq N_k^{\max},$$
$$C_8: \|\mathbf{q}_k – \mathbf{q}_l\| \leq d_{\max}, \quad C_9: b_{m,k} R_k^{\max} \geq \|\mathbf{q}_k – \mathbf{q}_m\|,$$
$$C_{10}: c_{m,k}, b_{m,k}, a_{k,s}, e_{m,s} \in \{0,1\}.$$

Here, \(D_m^{\max}\) is the maximum tolerable delay for user \(m\), \(E_m^{\max}\) is the maximum user energy budget, \(E_k^{\max}\) is the UAV energy budget, \(F_k^{\max}\) is the maximum CPU capacity of UAV \(k\), \(N_k^{\max}\) is the maximum number of users that UAV \(k\) can serve, and \(R_k^{\max}\) is the maximum coverage radius of UAV \(k\). This formulation is a mixed-integer nonlinear programming (MINLP) problem due to binary variables and the coupling among positions, offloading decisions, and resource allocations.

3.3 Proposed Solution

To handle the complexity, I decouple \(\mathcal{P}_1\) into three subproblems: offloading decision optimization, computation resource allocation, and UAV position optimization.

Offloading decision subproblem: Given the resource allocation and UAV positions, the offloading decision problem becomes a linear programming problem after relaxing \(c_{m,k}\) to continuous variables in \([0,1]\). I use the CVX toolbox to solve the relaxed problem and then reconstruct binary variables by sorting the fractional solutions and greedily setting them to one under feasibility constraints. The details are summarized in Algorithm 1.

Algorithm 1: Offloading decision recovery
1: Input: fractional solution \(\tilde{c}_{m,k}\)
2: For each UAV \(k\), sort \(\tilde{c}_{m,k}\) in descending order
3: Initialize \(i=1\)
4: Repeat: set \(c_{i,k}=1\), others to zero until constraints C1–C4 are violated
5: Output: binary offloading decision \(c_{m,k}\)

Computation resource allocation subproblem: With fixed offloading decisions and UAV positions, I optimize \(f_{m,k}\) to minimize the total delay. This problem is convex, and I use the Lagrange multiplier method. The Lagrangian is:

$$\mathcal{L} = \sum_{m=1}^{M} T_m + \sum_{m=1}^{M} \lambda_m (E_m – E_m^{\max}) + \sum_{k=1}^{K} \nu_k \left( \sum_{m=1}^{M} f_{m,k} – F_k^{\max} \right).$$

By applying KKT conditions, the optimal CPU frequency allocation is obtained as:

$$f_{m,k}^{\star} = \sqrt{\frac{C_m \tau}{\kappa \nu_k}},$$

where \(\nu_k\) is the Lagrange multiplier associated with the resource constraint and \(\tau\) is a constant related to the delay expression.

UAV position subproblem: For fixed offloading and resource allocation, the objective is non-convex with respect to the UAV positions \(\mathbf{q}_k\). The transmission rate \(r_{m,k}^{\mathrm{up}}\) is a convex function of \(\|\mathbf{q}_k – \mathbf{q}_m\|^2\), so I apply successive convex approximation (SCA). At each iteration, the rate is lower-bounded by its first-order Taylor expansion:

$$r_{m,k}^{\mathrm{up}} \geq \log_2\left(1 + \frac{h_0 p_m}{\frac{B}{M_k} N_0 (H^2 + \|\mathbf{q}_k – \mathbf{q}_m\|^2)}\right) \geq r_{m,k}^{\mathrm{lb}},$$

where \(r_{m,k}^{\mathrm{lb}}\) is a concave approximation around the current point. The transformed position subproblem becomes convex and can be solved efficiently using CVX. The overall joint optimization is realized by iterating among the three subproblems until the decrease of the objective is below a threshold \(\epsilon\).

3.4 Simulation Results

I set up a simulation scenario with a \(1 \times 1\) km area, 30 user terminals, and 2 UAVs hovering at an altitude of 200 m. The total bandwidth is 20 MHz, the carrier frequency is 2 GHz, and the maximum UAV CPU frequency is 10 GHz. The number of service types is 100, and each user can cache 5 services. The detailed simulation parameters are listed in the following table.

Parameter Value
Channel gain \(h_0\) −50 dB
Bandwidth \(B\) 20 MHz
Noise power \(N_0\) −114 dBm
User transmit power \(p_m\) 0.01 W
Maximum UAV energy \(E_k^{\max}\) 40 J
Backhaul rate \(R_b\) 5 Mbps
Switching coefficient \(\kappa\) \(10^{-27}\)

Figure 3.3 in the original thesis demonstrates that the total task completion delay grows with the number of users. Compared to the no-cache algorithm and the benchmark in [32], my proposed algorithm significantly reduces the delay. The advantage stems from the fact that service caching avoids repeated downloads from the remote cloud, and the joint optimization of offloading and UAV placement improves channel conditions. Figure 3.4 shows the delay as a function of the number of UAVs. When more UAVs are deployed, more devices can offload their tasks, which reduces the delay. My algorithm outperforms the benchmarks, especially when the number of UAVs is moderate. Figures 3.5 and 3.6 reveal that increasing the UAV energy budget or computation capability can reduce the delay, but the marginal benefit diminishes due to other constraints. Figure 3.7 illustrates that caching more services at user terminals lowers the delay, because local execution becomes possible without service downloading.

Chapter 4: UAV-Assisted Edge Computing Offloading for User Energy Minimization

4.1 System Model

The second research problem addresses the energy consumption of ground users in a UAV-assisted MEC system. I consider a scenario where ground users have partial offloading capability: each user can split its task into a portion executed locally and the remaining portion offloaded to one of the UAVs. Different from Chapter 3, the system here assumes that UAVs cache all service programs, while each ground user randomly caches a limited number of services. The time horizon \(T\) is discretized into \(N\) time slots of equal duration \(\tau\).

Let \(\mathbf{q}_k[n] = (x_k[n], y_k[n], H)\) denote the horizontal position of UAV \(k\) at slot \(n\). The distance between user \(m\) and UAV \(k\) at slot \(n\) is \(d_{m,k}[n] = \sqrt{\|\mathbf{q}_k[n] – \mathbf{q}_m[n]\|^2 + H^2}\). Instead of the free-space path loss model, I adopt a probabilistic air-to-ground channel model. The LoS probability is expressed as

$$P_{m,k}^{\mathrm{LoS}}[n] = \frac{1}{1 + a \exp\left(-b \left( \sin^{-1}(H/d_{m,k}[n]) \frac{180}{\pi} – a \right) \right)},$$

where \(a\) and \(b\) are environment-dependent constants. The NLoS probability is \(1 – P_{m,k}^{\mathrm{LoS}}[n]\). The average path loss is

$$L_{m,k}[n] = P_{m,k}^{\mathrm{LoS}}[n] \eta_{\mathrm{LoS}} + P_{m,k}^{\mathrm{NLoS}}[n] \eta_{\mathrm{NLoS}} + 20\log_{10}\left( \frac{4\pi f_c d_{m,k}[n]}{c} \right),$$

where \(\eta_{\mathrm{LoS}}\) and \(\eta_{\mathrm{NLoS}}\) are excessive path loss factors, \(f_c\) is the carrier frequency, and \(c\) is the speed of light. The channel gain \(h_{m,k}[n]\) is the reciprocal of the average path loss.

4.2 Offloading and Computation Model

Each user \(m\) has \(D_m\) bits to process within the time horizon. Let \(f_m^{\mathrm{user}}[n]\) be the local CPU frequency, and \(f_{m,k}^{\mathrm{uav}}[n]\) be the CPU frequency allocated by UAV \(k\) to user \(m\) at slot \(n\). The number of bits processed locally is

$$d_m^{\mathrm{user}}[n] = \frac{\tau f_m^{\mathrm{user}}[n]}{C_m^{\mathrm{user}}},$$

and the number of bits processed at the UAV is

$$d_{m,k}^{\mathrm{uav}}[n] = \frac{\tau f_{m,k}^{\mathrm{uav}}[n]}{C_k^{\mathrm{uav}}},$$

where \(C_m^{\mathrm{user}}\) and \(C_k^{\mathrm{uav}}\) are the required CPU cycles per bit for local and UAV execution, respectively. A binary offloading variable \(c_{m,k}[n]\) indicates whether user \(m\) offloads to UAV \(k\) at slot \(n\). The offloaded data rate is given by

$$R_{m,k}[n] = \frac{B}{M_k} \log_2\left(1 + \frac{h_{m,k}[n] p_{m,k}[n]}{\frac{B}{M_k} N_0}\right),$$

where \(p_{m,k}[n]\) is the user’s transmission power. Note that \(M_k\) is the number of users associated with UAV \(k\), which is determined by the nearest-distance association.

4.3 Energy Consumption and Problem

The local energy consumption of user \(m\) at slot \(n\) is \(e_m^{\mathrm{user}}[n] = \kappa (f_m^{\mathrm{user}}[n])^3 \tau\). The energy for transmitting tasks is \(e_{m,k}^{\mathrm{off}}[n] = p_{m,k}[n] \tau\). The UAV computation energy for user \(m\) is \(e_{m,k}^{\mathrm{uav}}[n] = \kappa (f_{m,k}^{\mathrm{uav}}[n])^3 \tau\). The total user energy includes the local computation energy and the transmission energy:

$$E^{\mathrm{user}} = \sum_{m=1}^{M} \sum_{n=1}^{N} \kappa (f_m^{\mathrm{user}}[n])^3 \tau + \sum_{m=1}^{M}\sum_{k=1}^{K}\sum_{n=1}^{N} c_{m,k}[n] p_{m,k}[n] \tau.$$

The UAV energy consists of computation and propulsion energy. The propulsion energy of a fixed-wing UAV is approximated by the kinetic energy term \(E^{\mathrm{move}} = \sum_{k=1}^{K}\sum_{n=1}^{N} \frac{W_k}{2} \|\mathbf{v}_k[n]\|^2 \tau\), where \(W_k\) is the UAV weight and \(\mathbf{v}_k[n]\) is its velocity.

The optimization problem to minimize the total user energy under delay and resource constraints is formulated as:

$$\mathcal{P}_2: \min_{f^{\mathrm{user}}, f^{\mathrm{uav}}, Q, C} E^{\mathrm{user}}$$
$$\text{s.t. } C_1: 0 \leq f_m^{\mathrm{user}}[n] \leq f_m^{\max}, \quad C_2: \sum_{m=1}^{M} f_{m,k}^{\mathrm{uav}}[n] \leq f_k^{\mathrm{uav,max}},$$
$$C_3: \sum_{t=1}^{n} c_{m,k}[t] R_{m,k}[t] \tau \geq \sum_{t=1}^{n} d_{m,k}^{\mathrm{uav}}[t],$$
$$C_4: \sum_{n=1}^{N} \left( d_m^{\mathrm{user}}[n] + \sum_{k=1}^{K} d_{m,k}^{\mathrm{uav}}[n] \right) \geq D_m,$$
$$C_5: E^{\mathrm{uav}} \leq E^{\mathrm{uav,max}}, \quad C_6: \|\mathbf{q}_k[n] – \mathbf{q}_k[n-1]\| \leq V_{\max}\tau,$$
$$C_7: \mathbf{q}_k[0] = \mathbf{q}_I, \mathbf{q}_k[N] = \mathbf{q}_F, \quad C_8: c_{m,k}[n] \in \{0,1\}, \quad C_9: \sum_{k=1}^{K} c_{m,k}[n] \leq 1.$$

4.4 Proposed Algorithm

This problem is MINLP and non-convex. I decompose it into three subproblems: offloading decision, resource allocation, and UAV trajectory optimization.

Offloading decision subproblem: With fixed resource allocations and positions, the problem becomes linear after relaxing the binary variables. I use the same procedure as in Chapter 3.

Resource allocation subproblem: For fixed offloading and trajectories, the objective function is convex because the energy expressions are polynomial functions with positive coefficients. The constraints are linear. Hence the subproblem can be solved by CVX.

Trajectory subproblem: The rate expression \(R_{m,k}[n]\) is non-concave with respect to \(\mathbf{q}_k[n]\). I apply SCA by finding a concave lower bound at the current trajectory. The bound is derived from the convexity of \(\|\mathbf{q}_k – \mathbf{q}_m\|^2\), leading to:

$$R_{m,k}[n] \geq \frac{B}{M_k} \log_2\left(1 + \frac{h_0 p_{m,k}[n]}{\frac{B}{M_k} N_0 (H^2 + \|\mathbf{q}_k[n] – \mathbf{q}_m[n]\|^2)}\right) \geq R_{m,k}^{\mathrm{lb}}[n].$$

Then the trajectory subproblem becomes convex and can be solved with CVX. The overall algorithm iterates among the three subproblems until convergence.

4.5 Simulation Results

I simulate a network with 8 ground users, 2 UAVs, and a time horizon of 100 seconds. The users are randomly distributed in a 1 km by 1 km area. Each user must process a task of 120 Mb. The UAV altitude is 50 m, the initial and final positions of UAV 1 are (50,50,50) and (0,1000,50), and for UAV 2 they are (50,50,50) and (1000,0,50). The simulation parameters are summarized in the following table.

Parameter Value
Maximum user CPU frequency \(f_m^{\max}\) 1 GHz
Maximum UAV CPU frequency \(f_k^{\mathrm{uav,max}}\) 10 GHz
Bandwidth \(B\) 4 MHz
Noise power spectral density \(N_0\) −114 dBm
Maximum UAV speed \(V_{\max}\) 15 m/s
UAV weight \(W_k\) 10 kg
Carrier frequency \(f_c\) 2 GHz
Switching coefficient \(\kappa\) \(10^{-28}\)

Figure 4.2 compares the proposed algorithm with two benchmarks: equal UAV resource allocation and half-offload strategy. The proposed algorithm achieves significantly lower user energy consumption. Figures 4.3 and 4.4 show the impact of user and UAV computation capabilities. A higher user computation capability reduces user energy, while a higher UAV capability encourages more offloading, which may slightly increase user energy due to more transmission. Figure 4.5 indicates that increasing bandwidth lowers user energy because the transmission time and thus the transmission energy are reduced.

Chapter 5: NOMA-Based Weighted Total Energy Consumption Optimization

5.1 System Model and NOMA

In the third research problem, I extend the UAV-assisted MEC system to a non-orthogonal multiple access (NOMA) framework. Multiple ground users can share the same sub-channel when offloading tasks to the same UAV. The UAV employs successive interference cancellation (SIC) to decode the superimposed signals. I assume the users are partitioned into groups, and each group is served by one UAV via NOMA. Each UAV operates on an orthogonal frequency band, so there is no inter-UAV interference. The system contains \(K\) UAVs and \(M\) users. Every user has a task of \(D_m\) bits to complete within the time horizon.

Let \(\mathcal{M}_k\) be the set of users assigned to UAV \(k\). The users within this group are ordered by their channel gains, i.e., \(h_{r_1,k}[n] \le h_{r_2,k}[n] \le \cdots \le h_{r_{M_k},k}[n]\). When UAV \(k\) decodes the signal of the \(m\)-th user, it treats the signals of users with larger indices as interference. Thus, the achievable data rate of user \(m\) in the \(m\)-th order of SIC is:

$$R_{m,k}[n] = \frac{B}{M_k} \log_2\left(1 + \frac{p_{m,k}[n] h_{m,k}[n]}{\sum_{l=m+1}^{M_k} p_{l,k}[n] h_{l,k}[n] + \frac{B}{M_k} N_0}\right).$$

The service caching model is similar to Chapter 3: each UAV caches popular services based on the request probabilities of its associated users. The binary variable \(a_{k,s}\) indicates whether UAV \(k\) caches service \(s\). If a service is not cached, the UAV needs to download it from the remote cloud, causing extra backhaul delay and energy consumption.

5.2 Energy Consumption and Problem Formulation

The total energy consumption comprises user local computation energy, user transmission energy, UAV computation energy, and UAV propulsion energy. The weighted objective is defined as

$$E = w_{\mathrm{user}} \left( E_{\mathrm{comp}}^{\mathrm{user}} + E_{\mathrm{off}} \right) + w_{\mathrm{uav}} \left( E_{\mathrm{comp}}^{\mathrm{uav}} + \alpha E_{\mathrm{move}} \right),$$

where \(w_{\mathrm{user}} + w_{\mathrm{uav}} = 1\), and \(\alpha\) is a scaling factor for the propulsion energy. The optimization problem is:

$$\mathcal{P}_3: \min_{f^{\mathrm{user}}, f^{\mathrm{uav}}, Q, P} E$$
$$\text{s.t. } C_1: 0 \le f_m^{\mathrm{user}}[n] \le f_m^{\max}, \quad C_2: \sum_{m=1}^{M} f_{m,k}^{\mathrm{uav}}[n] \le f_k^{\mathrm{uav,max}},$$
$$C_3: \sum_{t=1}^{n} R_{m,k}[t] \tau \ge \sum_{t=1}^{n} d_{m,k}^{\mathrm{uav}}[t],$$
$$C_4: \sum_{n=1}^{N} \left( d_m^{\mathrm{user}}[n] + d_{m,k}^{\mathrm{uav}}[n] \right) \ge D_m,$$
$$C_5: 0 \le p_{m,k}[n] \le p_m^{\max}, \quad C_6: \|\mathbf{q}_k[n] – \mathbf{q}_k[n-1]\| \le V_{\max}\tau,$$
$$C_7: \mathbf{q}_k[0] = \mathbf{q}_I, \mathbf{q}_k[N] = \mathbf{q}_F.$$

5.3 Joint Optimization Algorithm

The problem is highly non-convex due to the coupling of user grouping, NOMA interference, power allocation, and UAV trajectories. I decompose it into three subproblems: user grouping, resource and power allocation, and trajectory optimization.

User grouping subproblem: I formulate user association as a many-to-one matching game. Each user has a preference list of UAVs based on the weighted energy consumption if it joins a UAV. Similarly, each UAV has a preference list of users. The matching algorithm starts with an initial feasible association and iteratively swaps paired users or changes assignments if the total utility improves. The algorithm is summarized in Algorithm 2.

Algorithm 2: Many-to-one matching for user grouping
1: Initialize: user set \(\mathcal{M}\), UAV set \(\mathcal{K}\), maximum quota \(a\)
2: Each user ranks UAVs by distance and sends access requests
3: Accept if the UAV quota is not full; otherwise race to the next UAV
4: Compute utilities \(U_M(m,k)\) and \(U_K(m,k)\)
5: For each user \(m\) and each UAV \(k\), determine swaps that minimize total utility
6: Repeat until no beneficial swap exists
7: Output final matching \(\mu\)

Resource and power allocation subproblem: With fixed user grouping and UAV trajectories, the allocation problem is still challenging due to the power variables in the NOMA rate expression. By rearranging the rate formula, the sum of transmit power of all users in a group can be expressed in a recursive form. After some algebraic manipulation, the total transmission power becomes a convex function of the accumulated rate variables. Consequently, the subproblem can be transformed into a convex problem and solved by CVX. The details involve substituting \(p_{m,k}[n]\) with an exponential expression of the sum of rates, and then showing that the objective is convex because each term is an exponential with positive coefficient.

Trajectory subproblem: For fixed grouping and power allocation, the objective contains the path loss in the transmission power and the propulsion energy. I approximate the path loss \(L_{m,k}[n]\) as a quadratic function of the distance, e.g., \(L_{m,k}[n] \approx p_1 d_{m,k}^2[n] + p_2 d_{m,k}[n] + p_3\), where \(p_1, p_2, p_3\) are fitting coefficients. Thus, the transmission power becomes a quadratic function of the UAV position, and the propulsion term is a convex quadratic function of the position differences. The whole trajectory subproblem becomes convex and can be solved efficiently. The overall algorithm iterates among the three subproblems until convergence, as summarized in Algorithm 3.

Algorithm 3: Joint optimization for NOMA-UAV-MEC
1: Initialize: user grouping \(\mu^{(0)}\), resource variables \(f^{(0)}\), positions \(Q^{(0)}\), power \(P^{(0)}\)
2: Set \(i=0\)
3: Repeat
4: Given \(Q^{(i)}, f^{(i)}, P^{(i)}\), solve user grouping to obtain \(\mu^{(i+1)}\)
5: Given \(\mu^{(i+1)}, Q^{(i)}\), solve convex resource and power allocation to obtain \(f^{(i+1)}, P^{(i+1)}\)
6: Given \(\mu^{(i+1)}, f^{(i+1)}, P^{(i+1)}\), solve convex trajectory to obtain \(Q^{(i+1)}\)
7: Update \(i = i+1\)
8: Until the decrease in the objective is less than \(\epsilon\)

5.4 Simulation Results

I consider 8 users, 2 UAVs, and a 1 km by 1 km area. Each user has a 100 Mb task that must be completed within 100 seconds. The UAVs fly at an altitude of 50 m with initial and final positions same as in Chapter 4. The weight factors are \(w_{\mathrm{uav}}=0.2, w_{\mathrm{user}}=0.8\). The simulation results are presented in the following figures. Figure 5.2 compares the proposed algorithm with three benchmarks: fixed transmission power, equal UAV resource allocation, and half-offload strategy. The proposed algorithm achieves the smallest weighted total energy. Figure 5.3 shows that as the UAV computation capability increases, the weighted energy decreases because more tasks can be executed at the UAV, but the marginal gain becomes smaller. Figure 5.4 depicts the effect of the number of cached services at the user terminal. More local caches reduce the necessity of downloading service programs, thus lowering the total energy. Figure 5.5 demonstrates that the weighted energy grows with the number of users because more users require offloading and transmission. Figure 5.6 compares NOMA with OMA. The proposed NOMA-based system utilizes the spectrum more efficiently, leading to lower energy consumption for the same task completion.

Conclusion

In this thesis, I have investigated three key optimization problems in UAV-assisted mobile edge computing systems with service caching. The first problem targeted total task completion delay minimization through joint offloading, resource allocation, and UAV positioning. The second problem focused on minimizing user terminal energy consumption by optimizing partial offloading, resource allocation, and UAV trajectories. The third problem extended the study to NOMA-based multi-UAV MEC networks and minimized the weighted total energy consumption through user grouping, power allocation, resource allocation, and trajectory design. All three problems are non-convex MINLP problems, and I proposed efficient decomposition algorithms combined with convex optimization, successive convex approximation, and matching theory. Simulation results confirmed the superiority of the proposed algorithms over several benchmark schemes. The integration of an unmanned aerial vehicle with MEC and service caching proves to be an effective approach for providing low-latency and energy-efficient computation services in dynamic environments. Future work may consider dynamic service caching, delay-tolerant task scheduling, and decentralized learning-based decision making. The role of an unmanned aerial vehicle as a mobile edge node will remain an important research direction in the evolution of future wireless networks.

Scroll to Top