Large Language Models (LLMs) have revolutionized natural language processing, but they lack autonomous interaction capabilities. LLM agents address this by integrating three key modules, such as memory, planning, and tool use modules, around the LLM core. This survey systematically analyzes LLM agents through their architectural components, control flow paradigms, and application domains spanning data analysis, software engineering, and social simulation. We identify critical challenges in reliability, evaluation, and scalability, and outline future research directions. This work provides a structured overview of the rapidly evolving landscape of LLM agents.
In this research, we are the first to examine the potential and practicality of achieving a quantum computing advantage within communication relay networks. Barrage Relay networks (BRNs) offer a low-latency and resilient network structure capable of preventing collisions through self-directed cooperation, thus enabling a robust low-latency broadcast system. We extend BRN by designing a multi-hop quantum teleportationbased relay network. This work discovers a theoretical framework for a multi-relay strategy for quantum teleportation tailored for tactical ad-hoc networking in the post-quantum era, utilizing quantum devices along with preliminary experimental analyses to support our proposed protocol using Qiskit.
In this paper, we propose a robust optimizationbased mobile base station placement method for efficient store-carry-forward data collection using unmanned aerial vehicles (UAVs) in environments with limited or disrupted communication infrastructure. We focus on scenarios in which communication demand is temporally fluctuating and uncertain, with the aim to determine a placement strategy that maintains stable performance under varying future conditions. We formulate the base station placement problem as a robust optimization problem that explicitly accounts for demand uncertainty. We obtain the optimal solution using the column-and-constraint generation algorithm. To evaluate the effectiveness of the proposed approach, we conducted simulation experiments with predefined candidate sites and demand points, and compared the results with those of benchmark methods. The proposed method significantly reduces both the total deployment cost and number of installed base stations while maintaining a high data collection rate and balanced demand coverage, which demonstrates its robustness and efficiency, even under probabilistic demand fluctuations.
Cloud native architecture is about building and running scalable microservice applications to take full advantage of the cloud environments. Managed Kubernetes is the powerhouse orchestrating cloud native applications with elastic scaling. However, traditional Kubernetes autoscalers are reactive, meaning the scaling controllers adjust resources only after they detect demand within the cluster and do not incorporate any predictive measures. This can lead to either over-provisioning and increased costs or under-provisioning and performance degradation. We propose NimbusGuard, an open-source, Kubernetes-based autoscaling system that leverages a deep reinforcement learning agent to provide proactive autoscaling. The agents perception is augmented by a Long Short-Term Memory model that forecasts future workload patterns. The evaluations were conducted by comparing NimbusGuard against the built-in scaling controllers, such as Horizontal Pod Autoscaler, and the event-driven autoscaler KEDA. The experimental results demonstrate how NimbusGuard's proactive framework translates into superior performance and cost efficiency compared to existing reactive methods.
In this paper, we introduce the SDIR (Susceptible-Delayable-Infected-Recovered) model, an extension of the classical SIR epidemic framework, to provide a more explicit characterization of user behavior in online social networks. The newly merged state D (delayable) represents users who have received the information but delayed its spreading and may eventually choose not to share it at all. Based on the mean-field approximation method, we derive the dynamical equations of the model and investigate its convergence and stability conditions. Under these conditions, we further propose an approximation algorithm for the edge-deletion problem, aiming to minimize the influence of information diffusion by identifying approximate solutions.
Driver behavior monitoring is essential for advancing driver assistance systems, particularly in detecting highrisk or distracted actions. This study introduces ResBoot-50, an enhanced ResNet-50-based model designed for driver behavior detection, trained and tested on a dataset State-Farm, MRLEye, and Drive&Act datasets to capture a diverse range of driving behaviors. To ensure robust evaluation, we incorporated bootstrap sampling techniques, which provided varied training and validation splits, enabling a more comprehensive assessment of model performance and generalizability. ResBoot-50 achieved exceptional performance, with a validation accuracy, precision, recall, and F1-score all approximately at 99.52%, underscoring its reliability across multiple behavior categories. The use of bootstrap testing has proved beneficial in reducing overfitting and enhancing model robustness, supporting the models readiness for real-world applications. These findings highlight the impact of bootstrap-based evaluation in driver behavior analysis and suggest significant potential for integrating ResBoot-50 into driver assistance systems to improve road safety.
The prediction of anomalies and the diagnosis of various machines using Internet of Things (IoT) technology is a topic of extensive research. In this study, IoT technology was utilized to develop a system for collecting data and diagnosing existing air conditioners, with a focus on anomaly detection. Building on our previous research, which explored strategies for deploying vibration sensors in various locations to optimize seasonal diagnosis models, this study introduces a method for improving anomaly detection accuracy. The method leverages air-conditioner-specific sensor data analyzed using the MahalanobisTaguchi (MT) method, which calculates the Mahalanobis distance for normal sensor values and vibration data. Discrepancies in sampling rates between vibration and air-conditioner-specific sensor data were addressed using the proposed analysis method.
With the development of the ICT industry, endpoint security issues are emerging. To quickly detect and respond to evolving attacks, it is necessary to improve threat detection performance without degrading endpoints’ performance. Conventional data learning techniques have the problem of severe performance degradation when learning unbalanced and sparse datasets. In this paper, we propose a privacy-preserving generative adversarial network (PP-GAN) technique to solve the problems of conventional methods. In addition, we minimize damage caused by dataset leakage by filtering out privacy features that may cause information leakage in advance. According to the experimental results, even if the dataset is unbalanced and sparse, we improved the accuracy by 29% compared to the model that did not apply GAN by augmenting the dataset with privacy data deleted, achieving an average attack classification accuracy of 75.75%. In addition, we reduced data leakage by about 10% while reducing the accuracy by 0.55% compared to the general GANbased dataset generation model.
This paper introduces a joint network and computation resource management framework for mobile networks, aiming to minimize task completion times in offloading scenarios. A task management function (TMF) is designed to coordinate resources across network entities, optimizing both network and computation resources while ensuring tasks meet their deadlines. In addition, a low-complexity heuristic algorithm is developed. It operates with two sub-algorithms: initial resource allocation algorithm and resource adjustment algorithm.
Deep learning (DL)-based medical image classification has become a pivotal research area in computer vision, significantly enhancing the diagnostic process across the medical field. DL-based image classification can be aided by supplementary data, further improving its performance. One prominent example of such data is gaze-points, which involve eye-tracking techniques to document radiologists’ interaction with images. Recently, gaze-point data has been used as a pre-processing method to map an image into a graph, enabling the use of graph neural networks (GNNs) in computer vision domain. Such techniques have demonstrated significant improvements compared to conventional convolution neural networks (CNNs); however, their reliance on human involvement makes them labor-intensive, thus limiting their applications. To address this, we propose to leverage DL-based automatic segmentation mask generation to prepare image data as input for the GNN. This technique uses the segmentation mask as the attention information to guide the classifier. The results demonstrate that the proposed segmentation-aided classification model surpasses conventional CNN models and delivers the same performance as the existing supplementary data-aided techniques while reducing manual labor.
Mobile devices, such as smartphones and tablets, are often referred to as resource-constrained devices due to the limitations imposed by the hardware, interaction modalities, and the network environment. Expensive capped data plans and unreasonable data roaming charges often impose a constraint on the usage of cellular data for mobile device users. And yet, mobile devices remain the most popular platform for consuming social media content given the high popularity of platforms like TikTok, Instagram, YouTube, and Facebook. This work reviews existing approaches to media caching and focuses on a novel mobile data caching strategy leveraging semantic properties of data on mobile social media platforms.
In this paper, we propose an intelligent user localization application that adjusts based on the number of active base stations in the mobile network. To achieve this, we collect mobile network data using a simulator that supports the Open-RAN framework. Evaluation results demonstrate that the proposed application accurately predicts user locations when the number of active gNodeBs (gNBs) exceeds three.
Traffic classification is crucial for various aspects of network management, including network security, Quality of Service (QoS), and resource allocation. While deep learning (DL) models have demonstrated effectiveness in traffic classification, each type of DL model captures different feature sets, potentially missing essential information. For example, a dense network might excel at identifying general patterns, while a CNN focuses on spatial features, and an LSTM captures temporal dependencies. However, relying on a single DL model may result in incomplete feature extraction. To address this limitation and enhance classification performance, we propose an ensemble model that combines three DL architectures: a dense network, a CNN, and an LSTM. This ensemble method leverages the strengths of each model, enabling comprehensive feature extraction that incorporates spatial, temporal, and generalized patterns. By integrating these diverse feature sets, our ensemble model improves traffic classification accuracy and enhances overall system performance.
During search and rescue operations carried out for emergency response or disaster relief, there are many targets requiring rescue operations in a given environment, whose number and locations are unknown. In this paper, we investigate how to control the mobility of swarm of autonomous robots to efficiently search for and approach to targets to execute given tasks. We propose a swarm control algorithm, which includes the search and approach operations based on swarm intelligence, called mobile sensing cluster (MSC), combined with target localization employing particle filter (PF). We design mobility control that takes full advantage of the integrated operations of MSC and localization using PF. With computer simulations, we show that the proposed swarm control reduces the time required for searching targets and for completing their required tasks in comparison to reference schemes.
In this paper, we propose a novel framework for leveraging digital twins to enhance ON/OFF-LINE blockchain networks in drone and IoT communication systems. The proposed solution addresses critical challenges in environments where internet connectivity is intermittent or restricted, such as construction sites, remote areas, or military operations. We present a resilient communication architecture that integrates Digital Twins (DTs) to mirror the state of offline blockchain nodes in the cloud, enabling efficient synchronization and reducing data conflicts when transitioning between online and offline states. The framework also introduces a method for efficient data management and secure, decentralized identity verification by using packaged certification with DID documents and selfcontained verifiable credentials. By maintaining a synchronized digital twin in the cloud, the proposed system ensures that offline nodes can operate independently while seamlessly integrating updates when connectivity is restored.
Serverless computing has emerged as a revolutionary paradigm in cloud computing, allowing developers to focus on code development without managing the underlying infrastructure. However, the cold start problem remains a significant challenge and can negate many of the benefits of serverless computing. This paper comprehensively reviews three different reuse techniques, namely container reuse, data caching, and function reuse, which are resource reuse paradigms in serverless computing systems, to mitigate the cold start problem and optimize performance. This review paper emphasizes the importance of resource reuse strategies in improving the efficiency and responsiveness of serverless applications, enabling wider adoption of serverless computing in more areas.
Low-earth orbit (LEO) satellite networks enable communication services in areas beyond the coverage of terrestrial base stations. This system is considered a key technology for 6G communications as it offers lower latency and less signal attenuation compared to geostationary satellites, achieving high throughput. By integrating these satellite networks with terrestrial base stations, a network structure can be created that combines the wide coverage of satellites with the data processing capabilities of terrestrial base stations, providing a more reliable and flexible communication environment. Therefore, to address the increasing number of connected devices and the rise in network density due to the growth of IoT, recent research on integrated satellite-terrestrial network systems is actively being conducted. In this paper, we investigate the latest research topics, such as space-air-ground integrated networks (SAGIN), RIS-integrated systems, security issues, and semantic-aware SAGIN systems.
Deployment of IoT services consists of two principal types: the horizontal type which employs a common platform for multiple services, and the vertical type which utilizes a specialized environment for each service. In this paper, authors describe the location-aware service for mobile robots in construction fields as an exemplary of vertical IoT services. The authors have developed a position recognition system using the odometry of mobile robots to support the operation of existing remote patrol systems for construction sites. In the system, the odometry of the mobile robot is employed to display the current position of the robot within the construction site on a map generated by BIM, thereby providing support for the operation of the remote patrol system. This system improves the efficiency of remote patrols by allowing the operator to know the exact current location of the robot on the site, in addition to the camera images from the existing robots.
Recent years, with the introduction of 5th generation mobile communication networks, bursty traffic and upstream traffic characteristics have changed significantly. The deployment of small cells and passive optical networks for highspeed broadband services are considered. Passive optical network technology can apply to multiplex many small cells. The authors have proposed BCOM as a bandwidth control scheme for uplink traffic and have carried out bandwidth allocation, assuming the case where mobile vehicles pass according to schedule. In this paper, BCOMs evaluate assuming that mobile vehicles do not arrive on schedule.
Lower production costs have inspired studies on unmanned aerial vehicles (UAV) for wireless communication. However, limited transmission power and size of the UAV make it challenging to use advanced communication models while meeting the growing need for high data rates and energy efficiency (EE). In this paper, we study an energy-efficient UAV network enhanced by an intelligent reflecting surface (IRS) with simultaneous wireless information and power transfer (SWIPT), where the IRS is employed to improve the EE of ground user equipment (GUE). The goal is to maximize the average EE by jointly controlling the UAV’s flying route, IRS phase steer, UAV transmission power, and power splitting (PS) ratio of the energy transfer technology. The formulated problem of maximizing the average $E E$ is non-convex and thus challenging to be solved. To address this problem, we propose a deep reinforcement learning (DRL) approach. The modified reward function is implemented to enhance the efficiency of the DRL agent, which is formulated based on the expected signal-to-interference-plusnoise ratio (SINR) map. Simulation results demonstrate that the proposed DRL algorithm achieves lower energy consumption, higher data rate, and improved EE compared to the comparison algorithm.
