Priority-aware CO–MRFO-based task scheduling for energy-efficient mobile ad hoc clouds

1 Introduction

The growth of mobile devices, wireless communication, cloud infrastructure, and online educational services has increased demand for computation-intensive applications on platforms with limited processing capacity, memory, and battery energy. Mobile Cloud Computing (MCC) extends mobile-device capabilities by allowing work to execute locally or on external computing resources (Armbrust et al., 2010; Dinh et al., 2013; Fernando et al., 2013). Mobile Edge Computing (MEC) places resources closer to users and can reduce wide-area communication delay (Mao et al., 2017), while computation offloading can conserve device resources when its communication cost does not outweigh the benefit of remote execution (Kumar and Lu, 2010). These capabilities are particularly relevant to mobile learning, where users may concurrently request live lectures, online assessments, artificial-intelligence inference, interactive content, downloads, and background synchronization (Shorfuzzaman et al., 2018).

The scheduling problem is difficult because tasks differ in computational demand, input size, memory use, deadline, bandwidth requirement, and priority, whereas candidate nodes differ in processing speed, available memory, queue length, residual energy, location, and connectivity. A scheduler must decide whether each task should run on its source device, a nearby cooperative mobile node, an edge resource, or a remote service-provider node. The decision effects response time, transmission overhead, execution energy, resource utilization, deadline compliance, and service continuity. Assigning every urgent task locally is ineffective when the source device is overloaded, while offloading every large task can increase latency and transmission energy. The best execution location must therefore be selected from current task, resource, and network conditions.

This study considers a mobile ad hoc cloud (MAC) formed by service-consumer nodes, cooperative smart-mobile nodes, and service-provider nodes. The terminology is used precisely: MCC is the broader mobile-cloud paradigm; MEC places computation near users; a mobile ad hoc network (MANET) supplies infrastructure-independent communication among nearby devices; and the MAC is the cooperative resource-sharing and scheduling environment evaluated here. Mobility makes the task assignment time-dependent because links can appear or disappear between scheduling intervals. A feasible method must consequently model connectivity, repair invalid assignments, and update queues, bandwidth, and residual energy as the topology changes.

Existing task-offloading approaches include mathematical optimization, deterministic heuristics, population-based metaheuristics, and learning-based methods. Deterministic rules are inexpensive but generally cannot optimize delay, energy, capacity, connectivity, and priority simultaneously. Population-based methods can explore large combinatorial spaces, but their performance depends on maintaining an effective balance between global exploration and local refinement. The Cheetah Optimizer (CO) provides search, sitting-and-waiting, attack, and return-to-home behaviours (Akbari et al., 2022), whereas Manta Ray Foraging Optimization (MRFO) provides chain, cyclone, and somersault foraging mechanisms (Zhao et al., 2020). Their complementary roles motivate a hybrid in which MRFO explores alternative task–resource assignments and CO refine promising schedules.

The principal contribution is a priority-aware CO–MRFO scheduler with five integrated features: (i) a single minimization objective; (ii) explicit one-hot task–node assignment with offloading derived from the selected node; (iii) node-dependent energy accounting for source devices, cooperative mobile nodes, and service-provider nodes; (iv) a linearly decreasing individual-level probability that coordinates MRFO exploration and CO refinement; and (v) a reproducible evaluation protocol with matched computational budgets, paired runs, statistical testing, ablation, scalability, sensitivity, and service-level analysis. CO refers specifically to the Cheetah Optimizer of Akbari et al. (2022), not Cuckoo Search; consequently, nests and Lévy flights are not part of the method.

The remainder of the paper is organized as follows. Section 2 reviews cloud-supported mobile learning, task offloading, and relevant optimization methods and states the research gap. Section 3 presents the system model and objective. Section 4 defines the hybrid scheduler, Section 5 gives the reproducible experimental protocol, and Section 6 reports the results and their boundaries. Sections 7 and 8 present limitations and conclusions.

2 Literature review and research gap

2.1 Cloud-supported mobile learning

Cloud computing provides on-demand access to shared processing, storage, networking, software, and platform resources (Armbrust et al., 2010). In educational settings, these resources support content delivery, collaboration, analytics, mobile access, and elastic handling of peak demand. Cloud-based big-data analytics has also been examined as an enabling technology for mobile learning (Shorfuzzaman et al., 2018). The benefits of scalable access, however, do not remove the technical limitations of learner devices or wireless networks. Computation-intensive and delay-sensitive services still depend on effective placement of work across mobile, edge, and cloud resources.

Figure 1 inferences that educational requests are heterogeneous. Live lecture delivery and interactive services are sensitive to response delay; online assessment requests require timely and reliable processing; AI inference and AR/VR demand substantial computation; and downloads or background synchronization can often tolerate additional waiting time. This heterogeneity makes fixed local-only or cloud-only policies unsuitable. The present work uses synthetic educational-service classes to examine technical scheduling and offloading performance. It does not treat the synthetic workload as a measured institutional trace or infer gains in learning outcomes from system-level metrics.

2.2 Mobile-cloud scheduling and offloading

The delay–energy trade-off in mobile and edge computing is well established (Mao et al., 2017; Kumar and Lu, 2010; Liu et al., 2016). Local execution avoids radio transmission but can increase queueing delay and drain the source battery. Nearby cooperative or edge execution can reduce processing time while limiting wide-area delay, although it depends on connectivity and peer capacity. Remote execution offers greater resources but can incur routing latency, transmission energy, and variable network performance. Dynamic energy-aware cloudlet models illustrate the importance of jointly considering proximity, processing, and energy (Gai et al., 2016).

Earlier delay-optimal scheduling formulations established rigorous decision models for MEC (Liu et al., 2016). More recent work has considered multi-objective task offloading across edge–fog–cloud layers (Saif et al., 2024), heterogeneous CPU/GPU resources in industrial cloud–edge systems (Zhai et al., 2025), and joint offloading, resource allocation, and trajectory design for priority-aware multi-UAV MEC (Hao, 2024). These studies demonstrate the value of joint modeling, but their topologies and decision spaces differ from a mobility-aware mobile ad hoc cloud containing source devices, cooperative mobile nodes, and service-provider nodes. multi-objective controller placement in software-defined networks has also been studied in banking contexts (Abdulghani et al., 2026).

2.3 Metaheuristic foundation and hybridization

Task–node assignment is combinatorial, and the feasible set changes with resource availability and connectivity. Metaheuristics are therefore attractive when exhaustive enumeration is impractical. Genetic Algorithm, Particle Swarm Optimization, Whale Optimization Algorithm, and Grey Wolf Optimizer provide useful benchmark search mechanisms, but each can stagnate or converge inefficiently when exploration and exploitation are poorly balanced. Standalone comparisons are insufficient unless all methods receive matched candidate-evaluation budgets and identical task, topology, and initial-population conditions.

CO models distinct hunting phases that are well suited to local refinement (Akbari et al., 2022). MRFO combines chain, cyclone, and somersault movements that support broad exploration (Zhao et al., 2020). The proposed hybrid does not use a fixed handover between complete populations. Instead, each candidate independently selects the MRFO or CO branch at every iteration through a declining probability schedule. This individual-level mechanism preserves diverse exploration early in the search and increases refinement later without eliminating either branch.

2.4 Technical positioning and research gap table

The literature leaves four connected gaps. First, educational-cloud studies commonly emphasize adoption or applications rather than reproducible real-time task–node scheduling. Second, models developed for fixed edge, industrial, or UAV environments do not directly capture a cooperative mobile ad hoc resource pool with changing links. Third, priority is sometimes represented without coupling it to a unified deadline-sensitive objective. Fourth, aggregated energy terms can obscure whether energy is consumed at the source, a cooperative mobile node, or a service-provider node. The proposed CO–MRFO formulation addresses these gaps through explicit assignment and offloading coupling, node-dependent energy, priority-weighted deadline penalties, feasibility repair, mobility updates, and individual-level exploration–refinement switching. Table 1 summarizes the technical positioning.

TABLE 1. Technical literature comparison.

3 System model and problem formulation

3.1 Architecture and task model

Letdenote the set of computational tasks, and letdenote the set of service-consumer, smart-mobile, and service-provider nodes.

T = 1 , 2 , … , n

N = N S ∪ N M ∪ N P

Each task i is characterized by its computational requirement C i (cycles), input data size S i (bytes), memory requirement M i (MB), bandwidth requirement b i (bit/s), deadline D i (s), and priority weight

q i ∈ 1 , 2 , 3 .

Each node j is characterized by its processing capacity F j (cycles/s), maximum available memory M j max, available bandwidth B j max, queueing delay Q j t, and, for battery-powered nodes, residual energy E j r e s t. The structural organization of the proposed framework is depicted in Figure 1.

3.2 Assignment and offloading variables

The binary assignment variable is defined in Equation 1 as

y i j = 1 , if task i is assigned to node j , 0 , otherwise .

Each task must be assigned to exactly one node:

∑ j ∈ N y i j = 1 , ∀ i ∈ T (1)

Let s i denote the source node of task i. The offloading decision is not independent; it is derived from the assignment, as given in Equation 2:

o i = 1 − y i , s i (2)

Thus, o i = 0 indicates local execution at the source node, whereas o i = 1 indicates execution on an SMN or SPN. The priority value q i affects the cost associated with deadline violation but does not force a task to be assigned to a particular node.

3.3 Response-time model

For task i assigned to node j, the response time is defined asAccording to Equation 3, where R i j t is the effective data transmission rate between the source node and node j, and L i j t is the routing and protocol latency.

T i j = Q j t + C i F j + o i 8 S i R i j t + L i j t (3)

The deadline violation for task i is defined in Equation 4 as:

v i = max 0 , ∑ j ∈ N y i j T i j − D i (4)

Therefore, v i = 0 means that task i satisfies its deadline, while v i > 0 represents the amount of deadline violation.

3.4 Auditable node-dependent energy model

Energy consumption is calculated at the individual task level and subsequently aggregated according to the type of node executing the task.

For local execution on the source node, energy is given by Equation 5 as:where (Equation 5) P c p u s i is the CPU power consumption of the source node.

E i , s i l o c = P c p u s i C i F s i , j = s i , (5)

For execution on another mobile, edge, or cloud node, energy consumption is given by Equation 6 as:

E i j o f f = P t x s i 8 S i R i j t + P r x j 8 S i R i j t + P c p u j C i F j , j ≠ s i (6)

Here, P t x s i represents the transmission power of the source node, P r x j represents the receiving power of node j, and P c p u j represents the processing power of the node executing the task. Hence, the processing energy is explicitly node-dependent, and an SMN does not use a cloud-server power value.

The energy consumed by service-consumer nodes is given by Equation 7 as:

E S C N = ∑ i y i , s i E i , s i l o c (7)

For execution on smart-mobile and service-provider nodes, the energy is given by Equations 8a, b respectively as:

E S M N = ∑ i ∑ j ∈ N M y i j E i j o f f . (8a)

E S P N = ∑ i ∑ j ∈ N P y i j E i j o f f . (8b)

The total system energy consumption is therefore

E t o t = E S C N + E S M N + E S P N . (9)

Every reported total energy value must satisfy Equation 9.

For a battery-powered node, the available battery capacity in mAh is converted into joules usingwhere V is the declared nominal battery voltage.

E r e s = 3.6 V C m A h ,

3.5 Feasibility constraints

The assignment must satisfy the memory, computational, bandwidth, connectivity, and energy constraints of each node.

Per-node memory and CPU capacity constraints are given by Equations 10a, b as:

∑ i y i j M i ≤ M j max , ∀ j ∈ N . (10a)

∑ i y i j C i ≤ C j max t , ∀ j ∈ N . (10b)

Per-node bandwidth and connectivity constraints are given by Equations 11a, b as:

∑ i y i j b i ≤ B j max t , ∀ j ∈ N . (11a)

where

y i j ≤ a i j t (11b)

a i j t = 1 , if node j is reachable from the source and the link satisfies the minimum data ‐ rate requirement , 0 , otherwise .

For battery-powered nodes, the residual-energy constraint is given by Equation 12 as:

∑ i y i j E i j ≤ E j r e s t . (12)

SPNs are assumed to be mains-powered. Therefore, they are constrained by processing capacity, memory, queueing, bandwidth, and connectivity requirements rather than battery capacity.

3.6 Unified minimization objective

The proposed scheduler minimizes a unified objective function that jointly considers response time, energy consumption, priority-weighted deadline violation, and constraint violations. The scheduler minimizes the unified objective given in Equation 13:

J = w T T ^ + w E E ^ + w P P ^ + μ V (13)

The three normalized objective components - response time, total energy, and priority-weighted deadline violations - are given by Equations 14a, b, c respectively as: and

T ^ = ∑ i ∑ j y i j T i j T r e f (14a)

E ^ = E t o t E r e f (14b)

P ^ = ∑ i q i v i D i ∑ i q i (14c)

Here, T r e f and E r e f are reference values used for normalization.

The term V represents the normalized aggregate of memory, computational, battery, bandwidth, assignment, and connectivity violations. The fixed objective weights arewith the penalty coefficient

w T = 0.5 , w E = 0.3 , w P = 0.2 ,

μ = 1000 .

For sensitivity analysis, the combined time-priority emphasis is defined aswhile the energy emphasis is

α = w T + w P ,

β = w E .

Therefore,

α + β = 1 .

A lower value of J represents a better scheduling solution.

4 Proposed CO–MRFO scheduler

4.1 Continuous representation and decoding

The proposed CO–MRFO scheduler represents each candidate solution using a continuous vector

x ∈ − , 66 n

For task i, let F i t denote its ordered set of currently feasible nodes, and letdenote the number of feasible nodes.

m i = ∣ F i t ∣

The continuous variable x i is transformed into a discrete node index using the sigmoid function:

σ x = 1 1 + e − x

The decoded index is computed using Equation 15 as:

k i = min m i , − 1 ⌊ m i σ x i ⌋ (15)

The selected node is the k i-th element of F i t, and the corresponding assignment variable y i j is set to one.

The initial population is generated using independent uniform random samples within

− , 66 .

4.2 MRFO exploration

The MRFO branch is based on the Manta Ray Foraging Optimization strategy and employs three source-consistent operators: chain foraging, cyclone foraging, and somersault foraging (Zhao et al., 2020).

4.2.1 Chain foraging

Chain foraging updates a candidate from its predecessor and the current best, as given in Equation 16:where u ∼ U 0 , 1, x * t represents the current best solution, and

x r t + 1 = x r t + u x r − 1 t − x r t + α x * t − x r t (16)

α = 2 u ∣ ln ⁡ u ∣ .

4.2.2 Cyclone Foraging

Cyclone foraging uses a spiral coefficient beta, as given in Equation 17:where the spiral coefficient iswith u 1 ∼ U 0 , 1.

x r t + 1 = x r t + u x r − 1 t − x r t + β x * t − x r t (17)

β = 2 ⁡ exp u 1 I max − t + 1 I max sin ⁡ 2 π u 1 ,

4.2.3 Somersault Foraging

Somersault foraging is given by Equation 18 as:

x r t + 1 = x r t + S u 2 x * t − u 3 x r t (18)

where

S = 2 ,

and u 2 , u 3 ∼ U 0 , 1.

Within the MRFO branch, the three operators are selected according to the following probabilities.

• Chain foraging: 0.40
• Cyclone foraging: 0.40
• Somersault foraging: 0.20

4.3 CO refinement

The CO branch follows the Cheetah Optimizer proposed by Akbari et al. (2022). The search operation is given by Equation 19 as:where r ^ r , d − 1 is a standard normal random variable and α t is a bounded local step.

x r , d t + 1 = x r , d t + r ^ r , d − 1 α t (19)

Sitting-and-waiting retains the current coordinate, as given in Equation 20:

x r , d t + 1 = x r , d t (20)

The attack operation is given by Equation 21 as:

x r , d t + 1 = x r , d t + r ˇ r , d β r , d t (21)

wherefor a randomly selected neighbouring candidate k ≠ r, and

β r , d t = x k , d t − x r , d t ,

r ˇ r , d = ∣ z ∣ exp z 2 sin ⁡ 2 π z , z ∼ N 0 , 1

The return-to-home operation replaces a stagnated leader with its last successful best position and reinitializes one randomly selected candidate.

The four CO operations are selected using the following probabilities.

• Attack: 0.60
• Sit-and-wait: 0.20
• Search: 0.15
• Return-to-home: 0.05

4.4 Exploration–exploitation switching

To balance global exploration and local exploitation, the scheduler dynamically changes the probability of selecting the MRFO branch during optimization.

For each candidate and iteration, the switching probability is computed using Equation 22 as:

p t = p max − p max − p min t − 1 I max − 1 (22)

where

p max = 0.70 , p min = 0.30

For each candidate, ifwhere u ∼ U 0 , 1, the MRFO branch is selected. Otherwise, the CO branch is applied.

u < p t ,

Therefore, exploration is more frequent during the early iterations, whereas CO-based local refinement becomes increasingly dominant toward the end of the optimization process.

For sensitivity analysis, the midpoint probability is defined aswhile the probability span remains fixed at

p m i d = p max + p min 2 ,

p max − p min = 0.40

4.5 Boundary handling and repair

Updated coordinates are clipped to [−6,6], decoded by (Jain et al., 1984), and repaired in descending priority order. For an infeasible task, all currently connected nodes satisfying memory, compute, bandwidth, and battery constraints are evaluated. The task is reassigned to the node with the smallest incremental J. If no feasible node exists, it remains pending for the next interval and contributes to the drop and violation metrics. A candidate replaces its predecessor only when it has a lower J.

4.6 Mobility and termination

Because the system contains mobile nodes, node positions and the connectivity matrixare updated at every scheduling interval according to the Random Waypoint mobility model.

a i j t

If a selected communication link disappears before task execution, the task is returned to the pending queue and repaired using the current network topology.

The optimization process terminates when either of the following conditions is satisfied:orconsecutive iterations occur without improvement in the best objective value.

t = I max = 100 ,

H s t a l l = 20

Thus, the algorithm terminates after a maximum of 100 iterations or earlier when the optimization process stagnates for 20 consecutive iterations.

4.7 Computational complexity

When feasible-node lists are maintained, candidate decoding and objective evaluation requireoperations for n tasks.

O n

The repair mechanism evaluates multiple feasible nodes for infeasible tasks and has a worst-case complexity of

O n m ,

whereis the total number of nodes.

m = ∣ N ∣

With population size N p and maximum iteration count I max, the worst-case computational complexity is given by Equation 23 as:and the memory requirement is

O I max N p n m (23)

O N p n + n m .

Hence, the proposed scheduler maintains a population-based continuous optimization process while decoding each solution into a feasible task-to-node assignment and explicitly accounting for mobility, resource constraints, execution time, energy consumption, and deadline violations. Algorithm 1 flow can be traced back to Figure 2.

Algorithm 1 CO–MRFO scheduling.

5 Reproducible experimental protocol

5.1 Educational-service workload

The workload is synthetic and is not presented as a real educational trace. It contains the service classes in Table 2 so that priority and QoE can be evaluated explicitly.

TABLE 2. Task classes and simulation ranges.

Task attributes are sampled independently from uniform distributions over the stated ranges unless a trace is supplied. Arrivals follow a Poisson process; the default rate is selected to produce 25, 50, 100, 250, and 500-task windows. Table 3 summarize the system configuration in detail for further implementation of experimental protocol.

TABLE 3. System configuration.

5.2 Nodes, mobility, network, and power

The area is 1000 × 1000 m. Random Waypoint speed is 0–10 m/s with a 0–10 s pause. Transmission range is 100 m. Link bandwidth is 10–100 Mbps, and routing/protocol latency is 10–50 m. All ranges are sampled uniformly and recorded in the configuration file for each run. Table 4 explains the above discussed parameters in detail.

TABLE 4. Baselines and parameter settings.

5.3 Algorithms and matched budgets

All stochastic methods use a population of 50 and 100 iterations, corresponding to 5000 candidate evaluations before repair re-evaluations. The same task set, topology sequence, and initial candidate population are used in each paired run.

5.4 Runs, software, seeds, and statistical analysis

Each method is evaluated in 30 paired runs. NumPy SeedSequence(20260730).spawn(30) generates the run seeds; the child seed and complete JSON configuration are saved for every run. The reference implementation uses Python 3.11.6, NumPy 1.24.3, NetworkX 3.1, SciPy, and Matplotlib on Ubuntu 22.04, Intel Core i7-12700K, and 32 GB RAM.

Report mean, standard deviation, median, and 95% confidence interval. For paired comparisons with CO–MRFO, apply a paired t-test when the paired differences are approximately normal and the Wilcoxon signed-rank test otherwise. Correct the family of baseline comparisons by Holm’s method (Holm, 1979). Report the adjusted p-value and paired effect size. Convergence must use archived per-iteration values rather than curves reconstructed from final means.

5.5 Reproducibility record and material availability

Independent numerical reproduction requires the implementation source, environment file, the master seed and 30 child-seed records, run-specific JSON configurations, task and topology files, task-level energy records, per-iteration objective histories, and table/figure generation scripts. These materials are not embedded in this Word manuscript. Accordingly, the manuscript is self-contained at the mathematical and protocol-specification level, but its numerical results cannot be independently reproduced from the DOCX alone. The complete record should accompany any resubmission as supplementary files or a public repository deposit.

5.6 Result-record and consistency controls

Each paired run is defined by a single record containing the task set, topology sequence, initial candidate population, terminal objective J, optimizer runtime, SCN/SMN/SPN energy, deadline violations, completed and dropped tasks, makespan, and the per-iteration best-objective history. A record is admissible only when every task has one assignment or an explicit pending status, all reported units match the definitions in Sections 3 and 5, and the unrounded energy identity E_tot = E_SCN + E_SMN + E_SPN holds within numerical tolerance.

All objective comparisons use the same minimization direction: a smaller J is better. Summary statistics must be calculated from the same 30 paired terminal records, while percentage improvement over comparator b is 100(J_b−J_CO–MRFO)/J_b for minimized metrics. Optimizer runtime is wall-clock search time; throughput is N_completed/T_makespan and is never calculated from optimizer runtime. Values displayed after rounding may differ from component sums only by the stated rounding precision.

6 Results and discussion

All values in this section are reported from 30 paired runs under the protocol in Section 5. Lower objective, runtime, energy, and deadline-miss values indicate better performance; higher throughput, fairness, and QoE indicate better service quality.

6.1 Main scheduling results

The node-dependent energy decomposition for the 100-task workload is shown in Figure 3. CO–MRFO achieved the lowest mean objective (12.84 ± 0.31), runtime (8.74 s), and total energy (412.6 J). Compared with GWO, the strongest conventional metaheuristic in Table 5 it reduced the objective by 14.7%, runtime by 14.6%, energy by 10.8%, and deadline misses by 78.1%, while increasing throughput by 11.8%.

TABLE 5. Main results for the 100-task workload (30 paired runs).

Optimizer runtime is wall-clock scheduling time. System throughput is computed independently as Ncompleted/Tmakespan, where Tmakespan is the simulated service-completion interval; therefore, throughput is not calculated as the number of tasks divided by optimizer runtime.

The final 30-run objective distribution is summarized in Figure 4. Iteration-wise convergence trajectories must be generated directly from the archived per-iteration logs and must not be reconstructed from terminal means.

No claim about convergence speed is made from Figure 4. A convergence figure is valid only when the mean trajectory and its variability band are calculated directly from archived per-iteration histories for all 30 paired runs; terminal summary values must not be used to reconstruct such trajectories.

6.2 Statistical comparison

The prespecified paired comparisons were significant after Holm correction (adjusted p < 0.001). The objective improvement remained significant against both standalone components and GA, while energy and deadline-miss improvements were significant against the recent MFA, baseline and GWO, respectively. The large paired effect sizes indicate practical as well as statistical importance. All the above metrics comparitively listed in Table 6 which includes statistical metrics as discussed in this section.

TABLE 6. Paired statistical comparison with CO–MRFO.

6.3 Ablation study

Removing any proposed component degraded all three outcomes. The largest deterioration followed removal of node-dependent energy accounting, increasing the objective by 10.4%, energy by 32.7 J, and deadline misses by 0.96 percentage points. Standalone MRFO and CO were also inferior, confirming that the gain is not attributable to either parent optimizer alone. Table 7 exhaustively deals with ablation study in this regard.

TABLE 7. Ablation study.

6.4 Scalability and sensitivity

Scalability of CO-MRFO and GWO across five workload sizes is shown in Figure 5. Runtime and energy increased with workload size, but CO–MRFO, retained lower objective, energy, and deadline-miss values than GWO, at every scale. At 500 tasks, it reduced runtime by 16.4%, objective by 15.5%, energy by 11.0%, and deadline misses by 63.2%.

The selected configuration used α = 0.7 and β = 0.3, population 50, 100 iterations, and switching midpoint p_mid = 0.50, corresponding to (p_max, p_min) = (0.70,0.30). Larger populations or longer runs produced only marginal objective gains at substantially higher runtime. the qualitative study of scalability and sensitivity showcased in Tables 8, 9.

TABLE 8. Scalability results.

TABLE 9. Parameter sensitivity.

6.5 Priority and educational-service performance

As we move on to analysis of service level QoE the Table 10 demonstrates executionable priorities of performance metrics. Service-class response time and QoE are shown in Figure 6. Across service classes, mean response time ranged from 0.46 to 1.24 s, deadline misses remained at or below 0.84%, and Jain fairness remained above 0.978. Priority-3 assessment, inference, and AR/VR tasks maintained low miss and drop rates without forcing a fixed execution location. Lower-priority background work experienced the largest delay, consistent with the priority-aware penalty.

TABLE 10. Service-level and QoE results by task class.

Jain fairness is computed from completed-task rates z u of U users using Equation 24:

F J a i n = ∑ u = 1 U z u 2 U ∑ u = 1 U z u 2 (24)

where.

• Uis the total number of users.
• z uis the completed-task rate of user u ^.
• F J a i nis the Jain fairness index.

The class-level QoE score is fixed before analysis and computed as follows, where Rk and Wk are in seconds and Mk and Gk are deadline-miss and drop percentages expressed in percentage points:

The QoE score for service class k is as given in Equation 25:where clip 1 , 5 · restricts the resulting QoE value to the range 1–5.

Q o E k = clip 1 , 5 5 − 0.15 R k 1 s − 0.10 W k 1 s − 0.05 M k − 0.05 G k (25)

6.6 Discussion boundaries

The evidence supports improved simulated scheduling efficiency, not improved learning outcomes or deployment readiness. The results show lower objective, energy, and deadline misses under the declared workload and network model. Generalization to real educational platforms requires trace-driven and prototype evaluation.

6.7 Limitations

The workload is synthetic, the wireless channel model is simplified, and security and privacy are outside the objective. Although 30 paired endpoint runs are reported, any iteration-wise convergence figure must be regenerated from the archived logs rather than inferred from final summaries. Real educational traces, prototype deployment, richer channel models, and privacy-aware scheduling remain future work.

7 Conclusion

This study developed a reproducible CO–MRFO scheduler for priority-aware task assignment in mobile ad hoc clouds. CO is unambiguously the Cheetah Optimizer; MRFO performs exploration and CO performs refinement under the explicit switching schedule. The model couple’s assignment and offloading, accounts separately for SCN, SMN, and SPN energy, and includes feasibility, mobility, repair, complexity, and termination rules. In the 100-task experiment, CO–MRFO achieved the lowest objective, runtime, total energy, and deadline-miss ratio among the evaluated methods. Ablation and sensitivity analyses confirm the contributions of adaptive switching, priority weighting, and node-dependent energy accounting. These conclusions are limited to the simulated environment.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Mauro Femminella, University of Perugia, Italy

Reviewed by: Nafees Akhter Farooqui, Integral University, India

Abdulrahman M. Abdulghani, Putra Malaysia University, Malaysia