Dynamic decision systems increasingly depend on the integration of heterogeneous information sources-textual, visual, contextual, and relational-to achieve adaptive and context-aware intelligence. However, existing learning frameworks often process these signals in isolation, limiting their ability to adapt decisions over time. To address this challenge, we propose MAF-RL, a Multi-Source Actor-Critic Fusion Reinforcement Learning (RL) framework that formulates sequential recommendation primarily as an RL problem and uses multi-source fusion to construct expressive state representations for the agent. The principal novelty of MAF-RL lies not in the fusion operator itself, but in its role as a decision-aware state construction mechanism. Diverse data streams-including sequential histories, textual semantics, visual representations, contextual metadata, and relational signals-are integrated into a unified RL state that is optimized end-toend through long-horizon Actor-Critic policy learning rather than short-term prediction loss. The Actor-Critic architecture, optimized through Proximal Policy Optimization (PPO), learns dynamic policies guided by a multi-objective reward that balances immediate performance, novelty, and strategic repetition. By grounding policy learning on fused multi-source states, this formulation enables the agent to reason over multi-source evidence and adapt actions across evolving envi-ronments. Empirical evaluation on three large-scale multi-source benchmarks -MovieLens-1 M, Amazon-Books, and Yelp -demonstrates that MAF-RL consistently outperforms state-of-the-art baselines, achieving superior ranking accuracy (HR@10, NDCG@10) and a better trade-off be-tween repetition and novelty (RR@10, Novelty@10). Overall, MAF-RL should be viewed as an RL-based sequential decision framework whose effectiveness derives from its multi-source state construction, enabling more adaptive and principled behavior in dynamic recommendation settings.

The rapid proliferation of the Internet of Things (IoT) has enabled large-scale connectivity among heterogeneous and resource-constrained devices. Although this connectivity supports applications in healthcare, transportation, industrial automation, and cyber–physical systems, it also increases the difficulty of providing secure, efficient, and privacy-preserving group communication. Conventional group key management (GKM) schemes may impose substantial rekeying, communication, and credential-management overhead, while static identifiers or public credentials can expose devices to identity tracing and session linkage. This paper presents Lightweight and Privacy-Preserving Group Key Management (LPP-GKM), a group key management scheme for dynamic IoT groups. The revised design combines pseudonym-based ECC mutual authentication, transcript-bound session-key establishment, hierarchy-based encrypted rekeying, and controlled pseudonym refresh. During normal authentication, a device sends its current pseudonym, an ephemeral ECC point, and freshness information without transmitting its stable public key. The trusted Group Manager (GM) resolves the corresponding public key internally from a protected registration record and establishes an authenticated device–GM session key. For membership changes, LPP-GKM uses a balanced symmetric key hierarchy. The GM refreshes the affected keys from the changed member leaf to the root and encrypts replacement keys under keys held only by authorized sibling subtrees. Consequently, a revoked device may record public rekey-update messages but cannot decrypt the replacement hierarchy keys required to derive the new group key. The design requires $$O(\log N)$$ refreshed hierarchy keys and encrypted update components for a single membership change in a balanced group of N active members, while routine rekeying uses symmetric-key operations only. The security analysis scopes mutual authentication, session-key establishment, post-revocation key exclusion, backward secrecy for newly admitted members, replay resistance, and identity protection under the stated trust and cryptographic assumptions. Privacy is limited to protection against direct identity exposure and transcript-level linkage by external observers and honest-but-curious infrastructure; it is not claimed against the trusted GM, traffic analysis, physical-layer tracking, or GM compromise. The performance evaluation distinguishes device-side operations, GM-side processing, and total network-level rekey delivery cost.

The increasing number of connected vehicles and intelligent transportation systems has created a significant need for efficient offloading and resource management in vehicular networks. These networks face challenges such as high mobility, variable network conditions, and diverse resource types. Traditional centralized methods cannot handle these issues effectively. Scalable and decentralized solutions are necessary to reduce latency, energy use, and computational overhead while maintaining reliable task execution. This research introduces a Decentralized PDE-Guided Microservice Offloading and Resource Allocation Framework. It uses a Multi-Agent Deep Reinforcement Learning (DRL) approach. The system is modeled as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP). This allows vehicles and fog nodes to make localized decisions. We model vehicle density as Partial Differential Equations (PDEs) and Directed Acyclic Graphs (DAGs) to represent microservices at a fine-grained level. The framework applies Multi-Agent Proximal Policy Optimization (MAPPO) to optimize task splitting, offloading, and routing. Simulation results show the framework’s strong performance. Compared to baseline methods, it achieves up to 21.14% lower task completion time, 23.73% energy savings, and 17.75% reduced offloading latency.

In flying ad hoc networks (FANETs), high node mobility, dynamic topology, and limited resources, such as energy and bandwidth, lead to unstable links and short-lived routes. In such an environment, although Q-learning-based routing methods are adaptable, they face serious challenges in practice due to large state space, high computational load, and slow convergence. To address these issues, this paper proposes a two-level Q-learning-based geographic routing protocol called TLQ-Geo for FANETs. This protocol integrates hierarchical decision-making with adaptive reinforcement learning. TLQ-Geo divides the routing process into two layers: the guided region selection (GRS) layer and the Q-learning-based routing (QRL) layer. The GRS layer determines a bounded search corridor between the source and the destination using a chain of intelligent decision points (IDPs), while the QRL layer performs distributed path optimization within this virtual corridor via Q-learning. By restricting the state space to the region guided by IDPs, TLQ-Geo significantly reduces convergence time and computational overhead. In addition, a dynamic inter-layer feedback mechanism periodically evaluates the performance of each IDP chain and adaptively reconfigures it under topology variations. Extensive simulations demonstrate that when the node density varies, TLQ-Geo achieves higher network lifespan (approximately 4.51

Optimization is an important branch of engineering, artificial intelligence, and data science that aims to find the best solution among possible solutions. Given the increasing complexity of optimization problems, using traditional methods is often inefficient and challenging. In contrast, metaheuristic algorithms, inspired by natural, biological, and social phenomena, can strike a good balance between global and local search and provide optimal answers in a reasonable time. One of the algorithms invented in 2021 is the Chameleon Swarm Algorithm (CSA), which has achieved remarkable performance in solving complex optimization problems. The CSA literature lacks a comprehensive, structured, and analytical review. So far, no source has reviewed research on improved versions, hybrid algorithms, parameter modification methods, chaotic, fuzzy, OBL, and quantum models, as well as extensive practical applications of CSA in a single framework. This paper presents a comprehensive and integrated picture of the scientific status of CSA and future research opportunities, systematically reviewing 130 papers. The statistical analyses show that the largest number of papers are published by Elsevier (31

In mobile edge computing (MEC)-enabled Internet of Things (IoT) networks, optimizing task offloading is critical for improving energy efficiency, latency, and resource use in dynamic, resource-limited environments. This paper presents a novel artificial intelligence-based optimization algorithm, Celestial Swarm Optimization (CSO), inspired by gravitational dynamics for efficient task allocation in MEC-enabled IoT networks. Unlike traditional swarm intelligence methods, CSO integrates dual-attraction forces: global best guidance and a dynamically computed centre of mass, to balance exploration and exploitation adaptively. The proposed method decides, in real time, whether to offload tasks by checking device constraints and the state of the network, so it can jointly reduce latency, energy, and memory usage. Simulation results show that CSO surpasses standard techniques, including PSO, GA, and GWO, reaching as much as 21% energy savings, 17% less latency, and 22% better memory efficiency. Therefore, CSO appears to be a robust and scalable solution for intelligent task scheduling in IoT environments with strict energy budgets and heavy computation needs.

ABSTRACT In recent years, there has been a rise in the use of ChatGPT for education, healthcare, smart cities and emerging technologies. However, no available studies or reviews have provided a consistent and reliable depiction of the situation regarding its usage and evaluation. The reporting of datasets, evaluation indicators, factors influencing performance and conditions of deployment has also varied from study to study. This fragmented state of affairs seriously inhibits attempts to assess ChatGPT's capabilities and limitations and thus improve the design of future versions. Earlier reviews were often conducted in a way pertaining to a single area or were mainly descriptive, with less emphasis on methodological evaluation and issues in deployment and ethics. To fill this void, we undertook a systematic review, according to PRISMA guidelines, limiting our searches to English‐language journal articles published during 2021–2025 by reputable publishers. Such studies focused directly on GPT models, provided assessment conditions and were cited extensively as preprints, while the exclusion criteria encompassed poorly linked studies, those not in English and studies employing ChatGPT as an adjunct. An analysis of these studies revealed that the vast majority of research has taken place in the area of education (32%) and health (28%). The review revealed significant variation in assessment accuracy across domains, frequent challenges with doubtful sensitivity, unpredictable and rapid changes and risks associated with specific domains that impact reliability and safety. This study's primary contribution is an effort to develop an integrated analytical framework that puts together these interdisciplinary results in a streamlined manner for interpreting the capabilities and limitations of ChatGPT. Because of the methodological heterogeneity of existing studies, the results can be viewed as qualitative trends instead of standard quantitative evidence. The results thereby accentuate the need for consistent criteria, domain‐informed evaluation practices and stronger methodological reporting to underpin a more reliable deployment of ChatGPT‐based systems.

The rise of battery-powered Internet of Thing (IoT) fleets in buildings and campuses requires policies that manage sensing, communication, and edge-cloud offloading while considering energy, carbon, privacy, and cost limits. In this paper, we frame this challenge as a Markov Decision Process (MDP) and design a controller using Deep Reinforcement Learning (DRL). We present a Rainbow-based IoT controller that retains distributional value learning, dueling networks, NoisyNets, n-step returns, Prioritized Experience Replay (PER), and double selection, and contributes four novelties: dual-budget Lagrangian control with warm-up, connectivity-robust distributional targets reweighted by outage/queue risk, federated sketch-guided replay for underrepresented regimes, and realistic ISAC-aware macro-actions with integrated DP/COQ accounting and budget-aware training/logging. Simulations show that the proposed algorithm achieves ti88 % higher anomaly detection, ti39 % higher packet success, ti52 % less energy consumption, and ti74 % lower cloud cost than the best baseline, demonstrating superior utility, reliability, and sustainability in IoT workloads.

Nowadays, Federated Learning (FL) enables collaborative model training without sharing raw data, making it suitable for smart grid applications. We propose an adaptive energy-aware FL framework for predicting appliance-level energy consumption using the UCI Appliance Energy Prediction dataset. In our approach clients transmit either full-precision or 8-bit quantized updates based on reliability and energy levels, under a 0.2 MB budget. Our method improves prediction accuracy and reduces communication cost, supporting scalable and energy-efficient learning in smart home environments due to its implementation with a multilayer perceptron (MLP). We implemented our model in Python on Google Colab, and achieved superior performance compared to baseline full-precision and quantized approaches.

With the development of smart objects, Social IoT (SIoT) has become more complicated. Mobility and service management issues have become the most challenging problems for mobile objects acting as service providers in the SIoT environment, particularly in service management, including service discovery, selection, and composition. To provide an optimal service management approach, a conceptual model is needed to separate the different layers of SIoT, from the end-user's request for service to the detection and composition stages. Moreover, timing QoS factors are crucial elements that sufficiently demonstrate the better performance of the used approaches in this scope. This paper considers a new approach to service management in SIoT mobile objects by displaying a model regarding the QoS factors (Response Time (RT), Wait Time (WT), Discovery Time (DT), Composition Time (CT), and Selection Time (ST)) and service parameters. The proposed approach directly compares static and mobile objects based on the datasets related to mobile and fixed private objects, which will be described later. It determines the mobile object's Search Area (SA) by estimating the object's distance from the requester. Simulation results present improvement in dedicated QoS factors in the mobile SIoT environment. It is concluded that it is essential to consider the timing criteria of service quality in the SIoT. This allows for timely improvements that help reduce service management time in service discovery, selection, and composition.

Although cloud computing has introduced valuable capabilities and features for the IoT, several challenges still persist in this field. The long distance between IoT devices and cloud platforms often leads to data transmission delays, resulting in degraded service quality, particularly for latency-sensitive applications. In the fog platform, the negative impact of this distance is significantly reduced by performing certain data management operations, such as decreasing the volume of data sent to the cloud. In this research, we proposed a prediction-based scheme in fog computing to reduce the volume of time-series IoT data transmitted to the cloud. The proposed scheme consists of two approaches: data filtering and prediction. In the data filtering approach, the amount of data received from a sensor node is reduced by discarding its redundant measurements. In the prediction approach, a fog-based prediction model is employed to reduce data transmission to the cloud. In this approach, historical IoT time-series data are modeled; then, based on the state space model approach associated with the Kalman Filter, modeled historical data, and a machine learning method, a prediction model is developed to accurately predict the future values of time-series data. This model is deployed on both fog and cloud layers, and similar predictions are made simultaneously in both layers. In both layers, in place of the new measurement at time t, the predicted value at time t is used; just the new measurement, outside the confidence range, is transmitted from fog to cloud. The experimental results indicate that the proposed scheme reduces the volume of data transmitted to the cloud by 50.30

Wireless Sensor Networks (WSNs) in IoT environments face severe constraints that threaten both performance and lifetime. The primary challenge is the limited battery capacity of sensor nodes, which frequently causes premature network partitioning. Dynamic topologies driven by node mobility, hardware failures, or environmental interference further complicate reliable routing. When combined with fluctuating traffic demands in critical applications (healthcare, smart grids, environmental monitoring), these factors demand highly adaptive routing strategies. This paper proposes ARFOR, a hybrid framework that combines an enhanced Random Forest model with the Owl Optimization Algorithm (OOA). The Random Forest component, augmented by shared memory and an energy-aware attention mechanism, predicts topology changes and suggests initial routes. OOA then refines these paths using multi-modal network data while dynamically adjusting its parameters. Through continuous closed-loop adaptation, ARFOR aligns its behavior with real-time network conditions. Simulation results demonstrate that ARFOR increases residual energy by over 33

Autonomous vehicles, most notably self-driving cars, are seen by many as a generational shift in transportation, offering the potential to reduce crash rates and relieve congestion. In order to operate, autonomous vehicles combine data from cameras, radar, and LiDAR and need to act on this data in real-time creating a heavy and bursty computational demand. Processing these heavy and bursty demands within the autonomy system's power and thermal envelope will require considerable effort in determining which tasks to perform when and in what order. In some cases, offloading workloads to edge or cloud servers creates an opportunity to offload compute, reduce end-to-end latency, and enhance overall responsiveness if workloads are offloaded under strict latency constraints. In this work, we survey state-of-the-art offloading methods, identify significant challenges such as latency management, network reliability, and security, and outline future improvements within the area of vehicular systems. From an analysis of the state of the literature, we also objectively evaluate when and how offloading can enhance multi-faceted computational demands of autonomous driving stacks, with the overall goal of support safer and more capable vehicular systems.

Federated Learning (FL) is a relatively new concept, one that allows distributed training of models without having to rely on raw data being centralized. FL preserves privacy and fosters cooperation in learning. However, across the Internet of Things (IoT) and edge environments, a significant challenge remains: the high communication overhead that results from frequent model updates, especially under heterogeneous and non-IID data. To solve this, we propose an adaptive, communication-efficient framework that implements dynamic sparsification together with fine-grained layer-wise quantization-and we call it FedAdaptQ. The sparsification rate is adjusted as a function of training loss and gradient variance-sending more information in the earlier, less stable stages, and less in the later stable stages. At the same time, large gradients are selected by local metrics and quantized to conserve communication cost while maintaining accuracy. Experiments under non-IID frameworks manifest that the proposed method achieved competitive accuracy against FedAvg. These results raise the promise of FedAdaptQ for resource-constrained environments such as IoT and mobile devices.

Recommender systems (RSs) still struggle to effectively model complex, high-dimensional user-item interactions and incorporate contextual information. We propose SimGOL (Similarity-based Graph Optimization and Learning), a graph-based framework that addresses these challenges by modeling user and item interactions as separate graphs via Graph Convolutional Networks (GCNs) and integrating them through a Flexible Fusion Mechanism (FFM). This approach captures similarity-based connections in each graph and uses the FFM’s attention and gating to dynamically fuse the user and item representations. We evaluated our proposed model on six benchmark datasets, ranging from 82,790 to 1,000,209 interactions, comparing its performance against eight state-of-the-art baseline models. SimGOL achieved up to a 4.6

Wireless Sensor Networks (WSNs) integrated with the Internet of Thing (IoT) face critical challenges. These include limited energy resources leading to rapid node depletion, dynamic topologies due to node failures or mobility disrupting connectivity, and fluctuating traffic demands in applications like healthcare, smart grids, and environmental monitoring. These factors collectively result in reduced reliability, increased latency, and scalability limitations. These issues are exacerbated by unpredictable node behavior, heterogeneous device capabilities, and the need for real-time data delivery in mission-critical scenarios, necessitating robust, adaptive routing solutions. This paper proposes ASGRR (Adaptive Swarm-Guided Graph Policy Routing), a novel hybrid routing framework that integrates Message Passing Neural Networks (MPNN), Policy Gradient Reinforcement Learning (PGRL), and the Artificial Bee Colony (ABC) algorithm in a self-adaptive hybrid form. The MPNN employs a long-term memory module to predict topology changes, capturing complex spatial-temporal node relationships. The Policy Gradient method uses shared experience memory and predictive link stability metric to optimize multi-criteria path selection. The Bees Algorithm leverages entropy-based dynamic clustering to refine paths locally, adjusting scout bee allocation based on network volatility. ASGRR dynamically tunes parameters to adapt to real-time network changes. Simulation analysis demonstrates up to 38.06

Internet of Things (IoT), fog computing, and cloud services are revolutionary technologies. These technologies enable efficient collection, processing, and management of data. Today, as the importance of data is growing, the proliferation of IoT devices has accelerated. Fog computing acts as a decentralized intermediary and facilitates real-time data processing. And cloud services provide robust storage and scalable computing capabilities. However, data management, especially large-scale data, poses a significant challenge in the integration of IoT, fog computing, and cloud services. It requires efficient management of the vast amount of real-time data generated by IoT devices while ensuring seamless interoperability and scalability across decentralized and centralized systems. Given the importance of integrating IoT, fog, and cloud technologies for efficient data management, we present a systematic literature review (SLR) of 127 research articles and 17 review articles published up to 2025 that focus on key areas of these technologies. We propose two classifications; one addressing data management strategies and the other exploring real-world applications. We also review and discuss real-world applications, common evaluation environments, programming languages, simulators, and evaluation criteria in IoT, fog computing, and cloud services in recent papers. In addition, we identify common challenges, future directions, open issues, and a roadmap for researchers.

The Internet of Things (IoT) significantly impacts various industries, enabling better connectivity and real-time data exchange for applications ranging from smart cities to healthcare. Integrating cloud, fog, and edge computing is essential for managing increased data and processing needs as IoT networks become complex. Cloud computing provides extensive storage and powerful computing capabilities but can experience delays due to the distance data must travel. Fog computing addresses these delays by processing data closer to its source, while edge computing reduces them even further by processing data directly on IoT devices. Effective management of these computing layers requires strategic task offloading, which involves moving tasks to the most appropriate computing layer to balance latency, energy consumption, and operational efficiency. Several strategies have been developed to optimize network communication and task offloading, with metaheuristic algorithms emerging as promising approaches. Inspired by natural processes, these algorithms are skilled at searching complex spaces to find near-optimal solutions for efficient and dynamic task offloading. This review provides a detailed analysis of how metaheuristic algorithms optimize task offloading. It evaluates their effectiveness in improving system performance, managing resources, and reducing costs. The review also identifies the current challenges in this area and suggests future research directions to advance this field.

Respiration Rate (RR) is a biomarker for several illnesses that can be extracted from biosignals, such as photoplethysmogram (PPG) and accelerometers. Smartwatch-based PPG signals are more prone to noise interference, particularly within their lower frequency spectrum where respiratory data is embedded. Therefore, existing methods are insufficient for extracting RR from PPG data collected from wrists reliably. Additionally, accelerometer sensors embedded in smartwatches capture respiration-induced motion and can be integrated with PPG signals to improve RR extraction. This paper proposes a deep learning-based model to extract RR from raw PPG and accelerometer signals captured via a smartwatch. The proposed network combines dilated residual inception module and Multi-Scale convolutions. We propose a pre-trained foundation model for smartwatch-based RR extraction and apply a transfer learning technique to enhance the generalizability of our method across different datasets. We test the proposed method using two public datasets (i.e., WESAD and PPG-DaLiA). The proposed method shows the Mean Absolute Error (MAE) of 2.29 and 3.09 and Root Mean Squared Errors (RMSE) of 3.11 and 3.79 across PPG-DaLiA and WESAD datasets, respectively. In contrast, the best results obtained by the existing methods are an MAE of 2.68, an RMSE of 3.5 for PPG-DaLiA, an MAE of 3.46, and an RMSE of 4.02 for WESAD datasets.

Given the crucial role of teachers in the education system, robust mechanisms are necessary to enhance their teaching effectiveness. Through leveraging advanced technological methods, both teacher and student evaluation processes can be conducted with high accuracy. This study proposes an IoT-based automated teacher performance evaluation system that utilizes machine learning algorithms and computer vision techniques to provide immediate feedback on teaching performance to supervisors. The system analyzes key elements such as hand movements and the teacher's position in the classroom. By enhancing teaching performance, the model aims to improve student learning outcomes. In addition, to develop and test the system, a hypothetical dataset - called the teacher dataset - was created for this proposed model by collecting 35 publicly available videos from YouTube. This approach employs a ResNet50 pre-trained neural network for transfer learning and feature extraction to classify teacher behavior into 8 classes. Fuzzy logic converts the predictions into three teaching quality ratings (poor/medium/good). Using this custom dataset, the model achieved an accuracy of 84.8%, indicating strong performance. This approach enables automated feedback on teaching style, reducing the need for in-person evaluations by educational supervisors. The proposed system has the potential to significantly enhance the overall quality of teaching and learning.