Wireless Mesh Networks (WMNs) have undergone significant development in recent years due to their low cost, ease of implementation, and dependable service coverage. However, the placement of nodes in these networks presents a crucial research problem for network operators and has a significant impact on the performance of WMNs. This issue is known to be NP-hard, and the use of approximate optimization algorithms, such as meta-heuristic techniques, is essential for its resolution with reasonable time execution. In this paper, three approaches were suggested to solve the Mesh Routers (MRs) placement problem with service priority. The first one is an adaptive version of Snake Optimizer (SO), called GSO, based on the incorporation of the Generalized Opposition Based-Learning (GOBL) strategy into the original SO to improve its exploration phase. The second one is a hybrid approach, called SOSA, based on the combination of SO with Simulated Annealing (SA). The last one is a hybrid approach, called ESOSA, based on the integration of GOBL mechanism and SA algorithm into the original SO, to improve both exploration and exploitation phases. A new problem formulation is suggested, where the main objective is to determine the optimal positions of the MRs in order to maximize the weighted coverage while satisfying the connectivity requirement, depending on a given set of Mesh Client (MCs) weights and locations. This is particularly relevant as most previous studies on MRs placement have treated all MCs equally, neglecting those with higher priority needs. The proposed approaches are validated in terms of fitness value while varying three key parameters: the number of MRs, the number of MCs, and the Transmission Range (TR) value. Simulation results obtained using MATLAB demonstrate that the proposed approaches outperform the standard SO and SA algorithms, as well as Particle Swarm Optimization (PSO), Grey Wolf Optimizer (GWO), Whale Optimization Algorithm (WOA), Firefly Algorithm (FA), and Harris Hawks Optimization (HHO). Furthermore, ESOSA achieves superior effectiveness in optimizing weighted coverage compared to all evaluated algorithms.
Energy-efficient task scheduling is a critical challenge in heterogeneous cloud data centers, where makespan, energy consumption, runtime, and resource feasibility must be optimized simultaneously. This paper proposes a constraint-aware hybrid metaheuristic framework for multi-objective cloud task scheduling by integrating Genetic Algorithm (GA), Grey Wolf Optimizer (GWO), threshold-based local refinement, and Colored Petri Net (CPN)-based event-driven simulation. The CPN module formally models task dispatching, parallel execution, CPU, RAM, and bandwidth contention, resource allocation, task completion, and waiting behavior, enabling feasibility-aware evaluation under concurrent execution. GA and GWO first explore the scheduling search space, and the best solutions are then refined through a threshold-guided local improvement mechanism without relying on penalty functions. Experimental results show that the proposed Hybrid method outperforms GA and GWO, achieving the lowest mean fitness of 674.20, makespan of 821.96, energy consumption of 452.56, and runtime of 10.20. It also records the smallest fitness standard deviation of 25.22, confirming higher stability. Statistical tests confirm significant improvements, with p-values of 1.024455e-07 against GA, 1.862645e-09 against GWO, and 2.066489e-10 for the Friedman test. Ablation, convergence, sensitivity, Pareto, ranking, energy, and scalability analyses further verify its robustness.
Deploying deep neural networks on microcontrollers requires balancing classification performance with strict memory constraints. This paper presents the selection, optimization, and deployment of a single-image Human Activity Recognition model on an ARM Cortex-M4 microcontroller. Four architectures—a Tiny CNN, MobileNetV2, EfficientNetB0, and MobileNetV1 with α = 0.25—were evaluated using a dataset of 600 RGB images distributed across four activity classes: running, sleeping, eating, and laughing. Although MobileNetV2 achieved the highest screening validation accuracy of 90.83
CPU scheduling governs which runnable task executes on which core and for how long, and therefore sits on the critical path of throughput, interactive latency, fairness, energy consumption, and real-time predictability in every computing system. The past two decades have transformed the problem far beyond its classical uniprocessor framing—through symmetric and asymmetric multicore processors, virtualised and datacentre-scale infrastructure, and, most recently, learning-based scheduling policies—yet existing surveys each cover only a single subfield. This article presents a systematic, cross-paradigm review of CPU scheduling. Following a structured search and screening protocol over six decades of literature through 2025, we synthesise 85 primary and foundational works spanning seven strands: classical uniprocessor heuristics, multilevel-feedback and proportional-share scheduling, production operating-system schedulers (Linux CFS/EEVDF, SCHED_DEADLINE, FreeBSD ULE), hard and soft real-time theory (Rate Monotonic, Earliest Deadline First, server-based reservation), multicore/NUMA/heterogeneous and energy-aware scheduling, datacentre and cloud resource management, and reinforcement-learning-based schedulers. The review is organised around four research questions concerning (i) the metric trade-off structure of the design space, (ii) how formal guarantees and adaptivity can be reconciled, (iii) how the multicore and datacentre transitions reshaped scheduler design, and (iv) what learning-based methods demonstrably add. Our comparative analysis—supported by a unified metric framework, corrected worked examples, complexity results, and cross-class comparison tables—confirms that no policy dominates across objectives: shortest-remaining-time minimises mean waiting time but starves long jobs; Round Robin bounds response time at quantifiable context-switch cost; EDF is utilisation-optimal on a uniprocessor yet fragile under overload; virtual-runtime fair schedulers won general-purpose computing by tracking a fluid ideal at O(log n) cost; and learned policies adapt to non-stationary workloads while forfeiting worst-case guarantees and interpretability. We distil these findings into a research roadmap and an evaluation checklist targeting the field’s principal open problems: standardised benchmarking and reproducibility, certifiable learning-based scheduling, portable contention models, and energy- and security-aware co-scheduling.
Atomic execution across heterogeneous blockchains is fundamentally constrained by their diverse finality guarantees. While fast-path protocols like MDCC excel at reducing latency in stable networks, they do not explicitly manage the rollback risks inherent in probabilistic finality. This paper introduces SACC, a cross-consensus protocol that confronts this challenge by integrating a risk-bounded planner with a certificate-based commit mechanism. SACC is the first to treat probabilistic finality as a quantifiable risk budget. It models finality as a rollback probability curve, allocates a global risk budget δ across chains, and derives the minimum required confirmations for each. Based on this risk plan, SACC then searches a rich strategy space—spanning fast/slow paths, observer usage, and committee configurations—to find an optimal execution plan under latency and fee constraints. This strategy is then executed with a deterministic timeout and a safe fallback. We implemented SACC in Go and evaluated it in a fault-injectable simulator across a comprehensive set of conditions, including partitions, adversarial omissions, and bursty faults. Our results show that SACC consistently outperforms traditional 2PC and modern fast-path protocols in challenging environments, significantly improving commit rates while strictly bounding tail latency.
Time series forecasting typically faces significant challenges, including non-stationarity, multi-scale coupling, and dynamic structural evolution. Under spectral drift and energy overlap, conventional decomposition methods based on fixed frequency partitions tend to cause trend-seasonal mixing, thereby weakening the modeling of long-term dependencies. To address these issues, we propose a phase-aware modeling framework with adaptive frequency decoupling, termed PAFD, for non-stationary time series forecasting. The proposed approach first performs adaptive frequency decoupling under time-varying spectral structures to effectively separate trend and seasonal components, and then incorporates phase information to improve the temporal alignment of periodic patterns. By jointly modeling frequency evolution and phase dynamics, PAFD provides a unified solution for capturing complex temporal structure variations. Extensive experiments demonstrate that PAFD achieves competitive performance across diverse forecasting settings, particularly for long-horizon forecasting and highly dynamic scenarios with spectral variation and phase drift.
Edge computing platforms function as geo-distributed micro-clusters, where isolate-based execution units process requests at globally distributed points of presence. Understanding how architectural decisions—isolate lifecycle management, request routing policies, and execution locality—impact performance is essential for distributed systems design. This paper presents an empirical study of two production edge platforms using latency-sensitive cryptocurrency analytics workloads, conducted from a local Southeast-Asian client and three supplemental AWS cloud clients (EU-West-1, US-East-1, AP-SE-1). Under tested conditions, isolate pre-warming reduces cold-start overhead by 1.4× – 30× depending on client-edge geography (largest from our Vietnam primary baseline, smallest from US-East-1) compared to on-demand provisioning. Multi-region testing reveals that platform-level latency advantages are dominated by client-edge proximity rather than systemic runtime differences: from Vietnam, Cloudflare Workers exhibited approximately 2× lower mean latency than Vercel Edge Functions; from EU-West-1, US-East-1, and AP-SE-1, Vercel led by factors of 1.9-−3.7 × . The 2× differential originally observed is a regional effect, not a systemic architectural advantage. Burst traffic reveals distinct scaling behaviors attributable to resource provisioning strategies. Response-header telemetry (198K requests, CF-Ray + x-vercel-id) further reveals architecturally distinct routing patterns: Cloudflare routes 100
This study investigates the adoption of sustainable practices in software development through a mixed-method approach that integrates quantitative and qualitative analyses. Grounded in the best practices proposed by the Green Software Foundation (GSF), a research instrument was developed to assess professionals’ perceptions and the extent of sustainability adoption in different software development contexts. The quantitative analysis suggested a moderate level of maturity, with greater adherence to practices related to data reduction and code optimization, but limited attention to energy monitoring and adaptive mechanisms. Complementary interviews reinforced the influence of cultural, organizational, and governance-related factors that hinder systemic adoption. As an illustrative complement, a small set of five blockchain-based software projects was examined to observe how energy measurement mechanisms, resource optimization strategies, and adaptability to diverse infrastructures appear at the artifact level. This component is exploratory and is not intended as a primary evidentiary pillar. Consistent with the survey and interviews, it suggests that energy efficiency tends to remain a secondary concern, often addressed indirectly through performance optimization rather than as a core sustainability goal. Overall, the research contributes to the understanding of sustainable software engineering by linking energy efficiency, governance, and organizational practices within an integrated analytical perspective.
Wireless sensor networks (WSNs) are crucial for sensing and communication applications, but they are also highly susceptible to routing attacks, energy depletion, and trust mismanagement. Blockchain technology has recently been proposed as a decentralized mechanism to enhance confidence, integrity, and traceability in WSN routing. Nevertheless, blockchain-enabled routing remains plagued by key problems, including excessive energy consumption, scalability constraints, and the efficient selection of cluster heads (CHs). To overcome these challenges, this paper proposes a new hybrid metaheuristic, Enhanced Grey Wolf Optimization with Cuckoo Search (EGWO-CS), for trusted routing in blockchain-based WSNs. The proposed EGWO-CS combines the hierarchical hunting behavior of Grey Wolf Optimization (GWO) with modified exploration–exploitation control and the global search capability of Cuckoo Search (CS) via Lévy flights. Within this framework, EGWO-CS is applied to optimize CH selection and next-hop routing by considering multiple objectives, including residual energy, link reliability, latency, and trust index. Blockchain is used for trustworthy, secure, immutable, and tamper-proof routing and trust data management. Simulation results suggest that the proposed EGWO-CS-based approach offers significant improvements in packet delivery ratio, energy efficiency, trust precision, and network lifetime compared to conventional GWO, GWO-CS, and other standard schemes.
Remote sensing small object detection aims to identify objects in images taken at high altitude by high-resolution satellites. Currently, most approaches address this task using data augmentation or multi-scale feature fusion. However, the above methods are deficient in dealing with the detailed features of the image and lack effective means to deal with key information such as edges and textures in the image, which are crucial for small target detection. Therefore, we propose a novel network combining Frequency-Domain Guided Enhancement (FDGNet) for remote sensing small object detection. Firstly, we propose a new Frequency-Domain Feature Refinement (FDFR) module, which is capable of generating multi-dimensional features to effectively extract and enhance different detailed features. Secondly, in order to better integrate the interaction and alignment between frequency information and feature extraction results, we propose a new Perceptual Alignment Compression (PAC) module. This module can effectively fuse the frequency information with the image feature information to improve the interaction and alignment between them. Finally, we design a Multi-Scale Efficiency-Enhancing Mechanisms (MSEEM), by which features of different scales can be fused efficiently. This integration strategy plays an important role in preserving and enhancing the critical detail information of small targets, thus significantly improving the detection accuracy and robustness. Extensive experiments have demonstrated that FDGNet outperforms the state-of-the-art method in remote sensing dataset. On the RSOD and HRRSD dataset, our method achieved a 4.29
Communication networks in today’s world must support heterogeneous services with strict delay and loss constraints under highly bursty traffic conditions. Traditional queue management and scheduling mechanisms are often unable to maintain stable performance when traffic exhibits strong short-term correlations and heavy-tailed arrivals. This paper proposes a hybrid reinforcement learning (RL) framework that jointly controls queue management and bandwidth allocation in bursty multi-service networks. A Deep Q-Network (DQN) dynamically regulates queue thresholds to suppress congestion and stabilize latency, while a Proximal Policy Optimization (PPO) agent selects from a small set of interpretable orthogonal bandwidth allocation templates to adapt resource sharing. Traffic arrivals are modeled using generalized exponential (GE) burst processes with heterogeneous per-class deadlines, enabling realistic evaluation of delay, jitter, and application-visible packet loss. Extensive simulation results demonstrate that the proposed approach consistently reduces delay, tail latency, jitter, and packet loss compared with widely used heuristic and RL scheduling schemes, while maintaining high utilization. The results highlight the effectiveness of coordinated learning-based control for stable and QoS-aware operation in bursty networked systems.
Top-k aggregation querying poses challenges across various applications such as e-healthcare, log management, and edge computing. These applications typically generate and store distributed data, which is logically interrelated but physically dispersed in distributed networks. However, existing algorithms that aim to improve the querying process still allow a large number of irrelevant items to be collected, leading to excessively high network bandwidth consumption and execution costs, particularly when operating on large datasets. To tackle the challenges, this paper presents iterative exponential sampling (IES). The novel top-k aggregation algorithm effectively prevents low-quality items from being collected by pruning them using an exponential random distributed-based filter. Moreover, it incorporates an Iterative Threshold scheme to optimize input variables and the validation step, thereby reducing the overall execution cost. Theoretical analysis and experimental findings reveal that IES exhibits excellent scalability with the number of items and consistently reduces bandwidth consumption by up to more than 90% compared to existing algorithms. It is also approximately twice as fast as existing algorithms when tested on large datasets, all while maintaining guaranteed output accuracy.
Large language models (LLMs) increasingly operate as components inside larger software systems. In that setting, correctness is not all that matters: operators also need to see the evidence behind an answer, to know when the model is unsure, and to keep latency and token cost predictable. We ask whether a small, transparent control layer can steer how a fixed base model responds, without retraining it or wrapping it in a heavyweight agent. Inspired by active inference, the selector reads a compact task state and picks one of three response styles (direct, grounded, or bounded) by scoring each against an explicit objective over helpfulness, truthfulness, groundedness, calibration, hallucination risk, and verbosity. Every decision it makes is logged. On 2,000 questions spanning HotpotQA, MuSiQue, StrategyQA, and TruthfulQA, the controller makes response style a settable property: support exposure and uncertainty handling rise substantially, and the controlled system is preferred in most pairwise comparisons on evidence-rich tasks, while it loses ground where a concise commitment is rewarded. Task accuracy decreases in aggregate (from 0.318 to 0.233); this is the cost of the shift, not its objective, and the gains are strongly dataset-dependent rather than a uniform improvement in dependable behavior. We quantify the token and latency cost and trace the effect dataset by dataset. In a supplementary study with thinking mode enabled, the control-layer effect replicates on a second model family (DeepSeek-V4-Pro), and the state-conditioned selector outperforms a static always-grounded prompt, which collapses on low-evidence tasks; the policy choice is also robust to the scoring weights. The study concerns controllability rather than task performance: response style can be steered and audited at low cost, and the conditions under which this helps or hurts are reported per dataset.
Metaheuristic algorithms are effective for complex optimization, yet their performance is frequently hindered by reliance on manual parameter tuning and rigid exploration-exploitation transitions, often leading to premature convergence. To address these challenges, we propose a hybrid framework named the Reinforcement Learning-Driven Differential-Moss Growth Optimizer (RLDMGO). This method employs a Q-learning controller to adaptively schedule a Differential Evolution (DE) mechanism for global exploration and a Moss Growth Optimization (MGO)-based operator for local exploitation. This dynamic feedback-driven coordination helps reduce parameter dependence and enhances adaptability across diverse problem landscapes. The performance of RLDMGO is evaluated on the IEEE CEC2017 benchmark suite. Statistical evaluations, including the Friedman test, indicate that RLDMGO achieves a competitive ranking among fourteen evaluated state-of-the-art competitors. To demonstrate its practical applicability, a binary variant (bRLDMGO) is applied to feature selection across 12 public datasets, generally yielding lower classification error rates and smaller feature subsets compared to seven binary optimizers. Furthermore, RLDMGO is deployed as an automated Neural Architecture Search (NAS) engine for cellular traction force computation. It optimizes a 3D U-Net, reducing model parameters to 2.47 M while maintaining competitive prediction accuracy compared to existing deep learning models. Ultimately, this study suggests that RLDMGO can serve as a viable and adaptive solver for complex optimization problems.
The increasing volume and heterogeneity of real-time data generated by Internet of Things (IoT) devices in intelligent environments poses significant challenges for efficient data processing. Traditional centralized Complex Event Processing (CEP) deployments can become bottlenecks when geographically distributed sensors continuously generate heterogeneous events. To address this issue, this paper proposes an edge-fog-cloud Distributed Complex Event Processing (DCEP) system for scalable event detection in intelligent environments. The proposed system assigns raw event generation to the edge layer, local event abstraction to fog-layer CEP engines, and higher-level alert composition to a cloud-layer CEP engine. The implementation combines Apache Kafka, Esper, lightweight ARM devices, and a server, and is evaluated through functional tests, integration tests, a distributed stress test, and a controlled comparison between centralized and distributed DCEP deployments. The stress test shows that the distributed deployment reaches 68,027 processed events/s before saturation in the evaluated setup. The controlled comparison executes the same event patterns as a single-node DCEP deployment on an Orange Pi and as a distributed DCEP deployment across Raspberry Pi, Orange Pi, and server nodes. Across all tested rates, both deployments produce the same number of high-level alerts, while the distributed deployment achieves lower end-to-end latency and higher high-level-alert output rate. These findings show that the proposed system is beneficial when event patterns can be decomposed by domain and placed across heterogeneous nodes.
To address security threats such as signal interference and atypical data injection in modern aerial systems, this paper proposes an aircraft trajectory anomaly identification method based on Multimodal Large Language Models (MLLMs). Traditional statistical “outlier detection” often fails to distinguish between complex flight maneuvers and stealthy anomalies. To overcome this, we first construct a high-fidelity 3D trajectory simulation dataset based on aerodynamic and kinematic mechanisms across three flight mission profiles. A physics-based framework is then developed to simulate five typical anomaly patterns, including oscillatory noise, positional drift, and velocity perturbation. By leveraging the visual perception and logical reasoning of MLLMs, combined with Chain-of-Thought (CoT) and prompt engineering, trajectory data is transformed into “vision-text” multimodal inputs. Experimental results demonstrate that the proposed method achieves an overall accuracy of 89.73
The accurate detection of pavement cracks is crucial for improving the safety of road surface as well as for the durability of infrastructure. Compared with conventional crack detection techniques, the proposed Yolov8-MED model more accurately detects pavement cracks. Recent advances have demonstrated the effectiveness of MHSA, ECA, and DCNv3 in computer vision tasks, motivating their use for feature extraction in deep learning–based crack detection. The challenge of integrating these techniques into the Yolov8-MED model while minimizing the computational load has emerged as a vital area of study in the field of automated pavement monitoring. As a result of these development, this study creates a Yolov8-MED design integrate MHSA, ECA and DCNv3 to enhance the model’s ability to capture fine crack details and improve detection accuracy. The aim is to enhance the crack location accuracy and keep the computational cost simultaneously. Further, to fine-tune the bounding box regression some modifications to the SCYLLA-IoU loss function have been made to improve on the localization of the bounding box. We also constructed the Xi’an Crack Dataset (XCD), which contains 7790 images of various crack types captured under diverse environmental conditions and is used to evaluate model performance. The Yolov8-MED model was benchmarked using the CrackSeg9k dataset and compared with the state-of-the-art models with a precision of 98.4
Task failures in large-scale cloud clusters degrade reliability and waste resources, yet failure prediction studies often rely on subsampled traces because end-to-end processing at full scale is challenging. We present an at-scale, reproducible pipeline for task-failure prediction using the Google Cluster Trace 2019 by processing the complete “Instance events” records across all eight clusters (≈ 163 GB, 14,127 parquet files) into a compact, machine-learning-ready feature table. We formulate failure prediction as a regression problem over aggregated trace windows and evaluate regularized linear baselines (Ridge, LASSO), Extreme Learning Machine (ELM), and ELM hybrids tuned with nature-inspired optimizers such as Genetic Algorithm, Grey Wolf Optimizer, and Swarm-based simulated annealing. Models are assessed with time-ordered train/validation/test splits and repeated runs with mean ± SD, N = 5, to reduce temporal leakage and quantify variability. The best-performing model, ELM_SwarmSA, achieves R² = 0.9913 ± 0.0004 with MAE = 3,054 ± 31 and RMSE = 4,314 ± 40, outperforming simpler baselines. By enabling full-trace benchmarking and providing processing artifacts and a compact derived dataset, this work improves the reliability of empirical comparisons and supports practical failure-risk scoring to inform scheduling and resource management.
In Internet of Things environments involving autonomous drones, sustained collaboration and effective conflict mitigation are essential for reliable service delivery, particularly in mission-critical applications such as parcel logistics. This paper presents an applied-artificial-intelligence approach based on recommendation of things to support cooperation and conflict-aware decision-making among drones performing parcel-delivery tasks. Unlike existing task-allocation and optimization approaches that primarily focus on resource efficiency and assume uniform cooperation among drones, the proposed approach explicitly models inter-drone coordination relations (e.g., facilitate, constrain, and cause), along with historical interaction patterns, to capture both collaboration potential and conflict likelihood. The proposed system analyzes job-level context—including spatial proximity, drone capabilities and constraints, and historical interaction data—to model coordination factors such as inter-drone relationships, collaboration potential, conflict likelihood, and expected satisfaction. Using these factors, the system generates recommendations for drone pairings to support informed task execution. The approach is evaluated through extensive experiments on a parcel-delivery case study under controlled conditions, demonstrating improved coordination-aware recommendation behavior compared with baseline strategies, with more consistent prediction and identification of collaborative drone pairings and conflict-prone combinations across multiple time intervals.
While simple random graph models often have a strong theoretical basis, graphs with complex constraints can only be generated by heuristic algorithms. Using these methods, there is no guarantee that the generated graphs are sufficiently random. However, this is important knowledge in many applications of random graphs, such as creating realistic and diverse synthetic datasets. To address this problem, we propose a randomness measure based on pairwise graph distances, and we present four new feature-based graph distance measures tailored to graphs with bounded frequencies of small subgraphs (graphlets). Three of the distances use features derived from graphlet frequencies, while the fourth is derived from the joint degree distribution and therefore much easier to compute. We evaluate these distances in a series of experiments on synthetic and real networks. Our experimental results show that two graphlet-based distances do not reliably show good results and, in particular, do not reproduce the expected trends in experiments on measuring randomness. However, our novel Radial Graphlet Distribution Distance is effective, and comparable in performance to state-of-the-art methods. These findings highlight the importance of selecting an appropriate graph distance. Finally, we show that our easy-to-compute Joint Degree Distance is a viable alternative to graphlet-based distances, especially for measuring randomness in sets of very large networks.
