…

The distinction between the two optimization engines at the heart of the hybrid scheduler deserves closer examination, because the success of any combined metaheuristic depends on whether its components genuinely occupy complementary roles rather than duplicating each other’s search behaviour. The Manta Ray Foraging Optimization component draws its inspiration from three documented feeding strategies of manta rays: chain foraging, in which individuals follow one another in a line toward plankton concentrations; cyclone foraging, which adds a spiral or barometric movement pattern that can carry candidates far from their current positions; and somersault foraging, which pulls solutions back toward a reference point that combines the best-known position with randomly weighted individual positions. These mechanisms give the algorithm a pronounced capacity for global exploration, meaning it can traverse large stretches of the combinatorial space of task-to-node assignments without becoming trapped in the vicinity of an early candidate solution. The Cheetah Optimizer, by contrast, models the hunting phases of cheetahs, including searching for prey, sitting and waiting for an opportune moment, attacking with rapid local movements, and returning to a home position when a hunt fails. These phases are inherently local and exploitative in character, making the algorithm well suited to polishing solutions that already lie within a promising region of the search space.

The coordination mechanism between these two engines is one of the more distinctive features of the reported work. Rather than switching the entire population from one algorithm to the other at a fixed iteration count, a common but brittle design choice, the scheduler assigns each candidate solution an individual probability of following the MRFO branch or the CO branch at every iteration. This probability decreases linearly as the optimization progresses, so that early iterations favour broad exploration while later iterations favour refinement. The individual-level granularity matters because populations of candidate schedules rarely advance uniformly; some candidates may already sit in high-quality regions where refinement is appropriate, while others remain in poorly explored regions where further exploration is beneficial. A population-wide switch would force all candidates into the same mode regardless of their individual circumstances, potentially discarding useful diversity or wasting refinement effort on weak solutions. The declining schedule itself follows a well-established principle in metaheuristic design, often described as shifting the exploration-exploitation balance over time, and embedding it at the per-candidate level is a natural refinement of that principle.

The encoding of candidate solutions also merits attention. Each candidate is decoded into a one-hot task-node assignment, meaning that for every task in the workload, exactly one execution node is selected from the set of available options, which includes the source device itself, nearby cooperative mobile nodes, edge servers, and remote cloud resources. This representation makes the offloading decision explicit rather than implicit: a task is considered offloaded precisely when the selected node differs from its source device, and the communication, queuing, and execution consequences follow directly from that selection. Because randomly generated or mutated candidates frequently violate resource or connectivity constraints, particularly in a mobile ad hoc environment where links appear and disappear between scheduling intervals, the framework includes a repair procedure that converts infeasible assignments into feasible ones before evaluation. Repair operators are a standard technique for handling constrained combinatorial problems with metaheuristics, and their presence here ensures that the optimizer never wastes evaluation budget on solutions that could not actually be executed by the system.

The objective function aggregates four distinct concerns into a single minimization target: normalized response time, node-dependent energy consumption, priority-weighted deadline violations, and feasibility penalties. Normalization is important when combining heterogeneous metrics, since raw response times measured in seconds and energy measured in joules occupy incomparable numerical ranges; without scaling, one term could dominate the objective and effectively silence the others. The node-dependent energy model is particularly relevant to the mobile ad hoc cloud setting because energy consumption is not uniform across node types. A source device executing a task locally draws on its own battery and incurs no transmission cost, whereas offloading to a cooperative mobile node or an edge server introduces radio transmission energy whose magnitude depends on data volume, distance, and link quality, while the executing node’s processing energy depends on its hardware characteristics. A scheduler that treated all nodes as energetically equivalent would systematically misjudge the true cost of its decisions, potentially draining the batteries of cooperative peers or underestimating the burden on constrained devices.

The priority-weighted treatment of deadline violations reflects the heterogeneity of the workload. Educational service scenarios, which motivated the study’s synthetic workload classes, include live lectures and interactive services that are highly sensitive to response delay, assessment requests that require timely and reliable processing, computationally intensive artificial intelligence inference and augmented or virtual reality workloads, and background synchronization or downloads that can tolerate additional waiting. A uniform penalty for missing any deadline would allow the scheduler to sacrifice urgent tasks for the sake of tolerant ones, or vice versa, without regard to service importance. Weighting violations by priority directs the optimizer to protect high-priority tasks first, which aligns the optimization objective with the service-level expectations that a real deployment would impose.

The reported experimental results should be interpreted with the structure of the evaluation protocol in mind. The comparison against the Grey Wolf Optimizer, a widely used population-based metaheuristic that has served as a frequent benchmark in scheduling literature, was conducted over thirty paired runs with one hundred tasks, and the reductions attributed to the hybrid are substantial: a 14.7 percent improvement in the composite objective, a 14.6 percent reduction in runtime, a 10.8 percent reduction in total energy, and a 78.1 percent reduction in the deadline-miss ratio. The deadline-miss figure is especially striking, and its magnitude is plausibly connected to the priority-weighted violation term, which gives the optimizer a strong and differentiated signal for avoiding late completions. The pairing of runs, in which each algorithm faces identical task sets, topology conditions, and initial populations, and the matching of computational budgets, address a common methodological weakness in metaheuristic comparisons, where differences in evaluation budget or problem instances can produce misleading conclusions about relative performance.

Beyond the headline comparison, the paper reports statistical testing, ablation studies, scalability analysis, sensitivity analysis, and service-level analysis. Ablation studies are particularly informative for hybrid methods because they can isolate the contribution of each design element: removing the hybrid switching mechanism, the priority weighting, or the node-dependent energy model and observing how the objective degrades reveals which components carry the performance. The abstract indicates that all three elements received support from these analyses, which strengthens the claim that the hybrid design, rather than any single fortunate choice, accounts for the observed gains. Scalability analysis addresses whether the method remains tractable as task counts grow, an important question for population-based methods whose per-iteration cost scales with both population size and the cost of evaluating each candidate schedule. Sensitivity analysis probes how parameter choices affect outcomes, and service-level analysis connects the optimization metrics back to the quality-of-service guarantees that users of a mobile learning system would experience.

Several contextual considerations frame the interpretation of these findings. The evaluation was conducted in a simulated environment with synthetic educational-service workload classes, and the authors are explicit that the synthetic workload should not be treated as a measured institutional trace, nor should system-level metrics be read as evidence of improved learning outcomes. Simulation is the standard and appropriate methodology at this stage of research, since it permits controlled comparison, reproducibility, and systematic variation of conditions that would be impossible in a live deployment, but it also means that the reported percentages describe performance under the modeled conditions rather than guarantees for production systems. Real deployments would introduce workload distributions, mobility patterns, hardware heterogeneity, and network dynamics that no simulation captures perfectly, and the transition from simulated to deployed performance is a well-known gap in the mobile computing literature.

The terminology used throughout the work is worth noting for readers navigating the broader literature. Mobile cloud computing refers to the general paradigm of extending mobile device capabilities with external cloud resources. Mobile edge computing places those resources physically closer to users, reducing wide-area communication delay at the cost of more limited capacity. A mobile ad hoc network supplies infrastructure-independent communication among nearby devices, allowing connectivity without fixed access points. The mobile ad hoc cloud studied here combines these elements into a cooperative resource-sharing environment in which service-consumer nodes, cooperative smart-mobile nodes, and service-provider nodes all participate in executing tasks. This three-tier node structure is what distinguishes the setting from conventional edge-cloud offloading studies, because cooperative mobile nodes are themselves battery-constrained and mobile, so their availability and capacity change over time in ways that fixed infrastructure does not.

Time-dependence is a fundamental complication that the scheduler must confront. In a mobile ad hoc environment, the set of feasible task-node assignments is not static: links between devices appear and disappear as nodes move, queue lengths evolve as tasks complete, bandwidth fluctuates with interference and distance, and residual energy depletes with every computation and transmission. A schedule that is optimal at the moment of assignment may become partially infeasible moments later. The reported approach models connectivity explicitly, repairs infeasible assignments, and updates queues, bandwidth, and residual energy as the topology changes, which means the optimizer operates on a moving target. This dynamic character helps explain why the runtime of the optimizer, 8.74 seconds for the reported configuration, is itself a reported metric rather than an afterthought: in a time-varying environment, a scheduler that takes too long to converge may deliver solutions that are already stale by the time they are applied.

The broader significance of the work lies in its demonstration that hybrid metaheuristics with principled coordination mechanisms can outperform single-algorithm baselines on multi-objective scheduling problems in dynamic wireless environments. The combination of exploration-heavy and refinement-heavy search engines, coordinated through a declining per-candidate probability, offers a template that other scheduling studies could adapt. At the same time, the emphasis on reproducibility, matched budgets, paired statistical testing, and explicit ablation sets a methodological standard that addresses persistent criticism of metaheuristic research, where novel algorithms are sometimes compared under unequal conditions. For the mobile learning context that motivated the study, the practical implication is that priority-aware, energy-conscious scheduling across device, peer, edge, and cloud tiers can materially reduce deadline violations and energy consumption relative to simpler heuristic alternatives, provided the scheduler has accurate models of node capabilities, connectivity, and task requirements at the moment of assignment.

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

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

Article References: Tamilmani, A., & Subramanian, M. (2026). Priority-aware CO–MRFO-based task scheduling for energy-efficient mobile ad hoc clouds. Frontiers in Communications and Networks, 7, Article 1903388. https://doi.org/10.3389/frcmn.2026.1903388

Image Credits: AI Generated

DOI: 10.3389/frcmn.2026.1903388

Keywords: Priority-aware, MRFO-based, task, scheduling, energy-efficient, mobile, clouds, scientific research