In an edge-cloud multi-tier network, datacenters provide services to mobile users, with each service having specific latency constraints and computational requirements. Deploying such a variety of services while matching their requirements with the available computing resources is challenging. In addition, time-critical services may have to be migrated as the users move, to keep fulfilling their latency constraints. Unlike previous work relying on an orchestrator with an always-updated global view of the available resources and the users' locations, this work envisions a distributed solution to the above problems. In particular, we propose a distributed asynchronous framework for service deployment in the edge-cloud that increases the system resilience by avoiding a single point of failure, as in the case of a central orchestrator. Our solution ensures cost-efficient feasible placement of services, while using negligible bandwidth. Our results, obtained through trace-driven, large-scale simulations, show that the proposed solution provides performance very close to those obtained by state-of-the-art centralized solutions, and at the cost of a small communication overhead.

This study proposes an innovative approach to address the Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) by integrating Reinforcement Learning (RL) into Evolutionary Algorithms (EAs), forming the Reinforcement Learning-assisted EA (RL-EA). While traditional EAs struggle with scalability and convergence speed, RL offers promise in dynamic decision-making. By combining the strengths of both paradigms, RL-EAs aim to enhance the state-of-the-art solutions for CVRPTW. The proposed approach seeks to optimise fleet routes while adhering to capacity and time constraints, crucial for logistics and transportation sectors. This study contributes to the advancement of optimisation techniques in complex transport scenarios, offering potential improvements in efficiency and cost effectiveness.

Our benchmarking Llama 3 for Chinese news summarization is a novel approach that integrates cultural and ethical considerations into model evaluation, significantly enhancing the relevance and acceptability of the generated content. The study employs a comprehensive framework to assess accuracy, cultural understanding, and societal value compliance, providing a multifaceted evaluation of Llama 3's capabilities. The results demonstrate that Llama 3 outperforms traditional and contemporary models, achieving high scores in ROUGE metrics and specialized cultural and ethical indices. Key findings highlight the importance of fine-tuning on culturally rich datasets and the use of advanced evaluation metrics to capture the complex interplay between language, culture, and ethics. Challenges encountered during the research underscore the need for continuous dataset updates and metric refinement, suggesting directions for future studies. The insights gained from this evaluation contribute to the broader field of natural language processing by showcasing the potential of advanced models to produce high-quality, culturally aware, and ethically compliant summaries.

This paper delves into the realm of large language models (LLMs), which are potent tools within the domain of artificial intelligence, facilitating comprehensive comprehension and utilization of human language by computers. LLMs, exemplified by GPT-3, exhibit remarkable proficiency across various linguistic tasks, ranging from writing to language translation and customer service. However, despite their efficacy, these models are not without drawbacks; they often manifest biases, possess opaque workings, and consume substantial energy resources. Tracing the evolution of LLMs from rudimentary models to sophisticated constructs capable of intricate language processing, this paper explores their learning mechanisms, diverse applications, and associated concerns, including the imperative of fairness and privacy preservation. Looking towards the future, it advocates for research endeavors aimed at enhancing the environmental sustainability, comprehensibility, and impartiality of LLMs. The overarching objective is to harness these models effectively and ethically, cognizant of their transformative potential across technological and societal landscapes. In addition, it should be noted that the author extensively reviewed and studied over 40 articles within a span of 2 months to comprehensively acquaint themselves with the subject matter, ensuring the most effective presentation of large language models (LLMs) to the interested audience.

The narrative that emerges from the current development and training of the artificial intelligence models has a rhetoric of unlimited potential of AI in its current state. This paper argues against that very point and will be looking at the serious constraints currently facing models of AI and machine learning, particularly the finitude of the data on which they rely. Its not just that the results are limited by the data at hand, but also by a host of other constraints: diminishing returns in scaling up model size, imposing costs on training, and making models harder to run. These are increasingly observed when AI is clad upon complex, real-world tasks that involve genuine generalization and adaptability. The paper also critically examines the commercial motivations driving the focus on automating white-collar jobs, highlighting that such priorities often stem from immediate financial incentives rather than a comprehensive evaluation of AI's broader applications. These practices raise profound ethical issues, including concerns about transparency and the responsible deployment of AI technologies. By dispelling the myth of AI's unlimited potential, this paper advocates for a more grounded and ethical approach to AI development, stressing the urgent need for breakthrough technological advancements that move us closer to achieving artificial general intelligence (AGI).

A document by WANG Shunchao. Click on the document to view its contents.

The unit cell dynamics of Liquid-Crystal (LC) based reflectarray antennas and Reconfigurable Intelligent Surfaces have been recently analyzed. However, the temporal evolution of the radiation pattern, which can differ from the individual cell response, has not been thoroughly studied. In this paper, a methodology to analyze the dynamic far-field radiation is proposed and tested for two LC antennas. This can open the door to the relaxation of temporal requirements and to achieve further antenna-level response time optimization.

This paper proposes a three-layer unmanned combat aerial vehicle (UCAV) dogfight frame where Deep reinforcement learning (DRL) is responsible for high-level maneuver decision. A four-channel low-level control law is firstly constructed, followed by a library containing eight basic flight maneuvers (BFMs). Double deep Q network (DDQN) is applied for BFM selection in UCAV dogfight, where the opponent strategy during the training process is constructed with DT. Our simulation result shows that, the agent can achieve a win rate of 85.75% against the DT strategy, and positive results when facing various unseen opponents. Based on the proposed frame, interpretability of the DRL-based dogfight is significantly improved. The agent performs yo-yo to adjust its turn rate and gain higher maneuverability. Emergence of "Dive and Chase" behavior also indicates the agent can generate a novel tactic that utilizes the drawback of its opponent.

Wide bandwidth requirements for multi-Gbps communications have prompted the global telecommunications industry to consider new mid-band spectrum allocations in the 4--8 GHz FR1(C) and 7--24 GHz FR3 bands, above the crowded bands below 6 GHz. Allocations in the lower and upper mid-band aim to balance coverage and capacity, however, there is limited knowledge about the radio propagation characteristics in the 4--24 GHz frequency bands. Here we present the world's first comprehensive propagation measurement study at 6.75 GHz and 16.95 GHz in mid-band spectrum conducted at the NYU WIRELESS Research Center spanning distances from 11--97 m using 31 dBm EIRP transmit power with 15 and 20 dBi gain rotatable horn antennas at 6.75 GHz and 16.95 GHz, respectively. Analysis of the omnidirectional and directional propagation loss using the close-in free space model with 1 m reference distance reveals a familiar waveguiding effect in indoor environments for line-of-sight (LOS). Compared to mmWave frequencies, the omnidirectional LOS and non-LOS (NLOS) path loss exponents (PLE) are similar, when using a close-in 1 m free space path loss reference distance model. Observations of the omnidirectional and directional RMS delay spread (DS) at FR1(C) and FR3 as compared to mmWave and sub-THz frequencies indicate decreasing RMS DS as the carrier frequency is increased. The RMS angular spreads (AS) at 6.75 GHz are found to be wider compared to 16.95 GHz, showing greater number of multipath components from a broader set of directions in the azimuthal spatial plane when compared to higher frequencies. This work also presents results from extensive material penetration loss measurements at 6.75 GHz and 16.95 GHz using co and cross polarized antenna configurations for ten common construction materials found inside buildings and on building perimeters, including concrete walls, low-emissivity glass, wood, doors, drywall, and whiteboard. Our findings show penetration loss increases with frequency for all of the ten materials and partitions tested, and suggest revisions of 3GPP material penetration loss models for at least infrared reflective (IRR) glass and concrete may be necessary. The empirical data and resulting models for radio propagation and penetration presented in this paper provide critical information for future 5G and 6G wireless communications.

Optimizing neural network architectures presents significant challenges due to the vast search spaces and computational costs involved. This study explores the integration of evolutionary algorithms (EAs) and mathematical modeling techniques to enhance neural network optimization. We propose a novel framework combining EAs with dimensionality reduction, surrogate modeling, and hybrid optimization strategies to reduce computational complexity and improve performance. Our results demonstrate that the adapted EAs significantly increase accuracy and F1-scores while reducing the number of generations required for convergence. The hybrid approach, combining EAs with local search techniques, achieves superior performance and robustness across various datasets. These findings provide a foundational basis for future research in advanced optimization methods for neural networks.

Background. Closed loop brain-computer interfaces dynamically adjust stimulation settings and/or timings based upon concurrently measured data. EEG (electroencephalography) is a widely used input. Particularly for closed loop applications in sleep monitoring, Phase Locked Loops have been used to track the narrowband phase of the EEG signal in real-time to provide a reference signal for closing the loop. During sleep, motion artifacts are minimal. However, there are many potential applications of real-time phase tracking of EEG when using more mobile EEG where artifacts are not necessarily negligible. Objective. To evaluate the robustness of PLL based EEG phase tracking when used with artifact corrupted EEG signals. We hypothesize that the intrinsic flywheel action of a PLL means that the tracked phase will not be perturbed by transient artifacts, leading to good tracking performance even in EEG situations without dedicated artifact removal stages being applied. Approach. We tested a PLL algorithm by using single channel (Fp1) EEG data from two datasets, each of which contains both cleaned and artifact contaminated versions of the same underlying EEG signal. We explored whether the PLL has similar phase tracking performance on both versions of the data. The Phase-Locked Value (PLV) was used for evaluating the phase tracking performance. Results. In general, PLV values in excess of 0.7 were obtained for phase tracking performance, irrespective of whether clean or artifact contaminated EEG data was passed to the PLL. Averaged across all EEG frequency bands, we found no statistically significant difference (p < 0.01) between the PLV values generated from clean and contaminated versions of the EEG signals. The phase tracking performance remained stable even as the number of artifacts present increased, with the decrease in average PLV being less than 0.1 even in high artifact cases. Tracking each EEG frequency band in isolation, Delta band tracking was more sensitive to low-frequency and high-amplitude artifacts, such as EOG and jump artifacts, with many PLV values below 0.6 being obtained. Differences in PLV values were much lower for the Theta, Alpha, and Beta bands, with differences increasing in the higher frequency Gamma band. Significance: Robust phase tracking in the presence of artifacts means that dedicated artifact removal processing may not be necessary for closed loop EEG, unless working with Delta or high Gamma band signals. Artifact removal algorithms are computationally complex and resource intensive, and operating without a dedicated removal step may reduce the processing time required, allowing a faster closing of the loop.

Chronic liver disease incidence and mortality have been rising worldwide. In many cases, liver disease is detected late in the symptomatic stage, while the earlier detection would be crucial for early initiation of preventative actions. "The Chronic Liver Disease score", CLivD, risk detection model has been developed with Finnish healthcare data and it predicts a person's risk of getting the disease in future years. In this study, real-world data repository (Kanta) was used as a data source for "The ClivD score" risk calculation model. Our dataset consisted of 96 200 individuals from Kanta. We had two main objectives: 1) to evaluate feasibility to implement automatic CLivD score with current Kanta platform, 2) to identify and suggest the improvements for Kanta that would enable accurate automatic risk detection. We found that Kanta currently lacks many CLivD risk model input parameters in the structured format required to calculate precise risk scores. However, the risk scores can be improved by utilizing the unstructured text in patient reports and by approximating variables by utilizing other health data like diagnosis information. With only utilizing structured data we were able to identify only 33 persons out of 51 275 persons to "Low risk" category and under 1% to "moderate risk" category. By adding the diagnosis information approximation and free text utilization we were able to identify 37% of persons to "Low risk" category and 4% to "moderate risk" category. In both cases we were not able to identify any persons to "high-risk" category because of the missing waist-hip ratio measurement. We evaluated three scenarios to improve the coverage of waist-hip ratio data in Kanta and these yielded the most substantial improvement in prediction accuracy. We conclude that the current structured Kanta data is not enough for precise risk calculation for CLivD or other diseases where obesity, smoking and alcohol use are important risk factors. Our simulations show up to 14% improvement in risk detection when additional data sources are considered. Kanta shows potential for implementing nationwide automated risk detection models that could result in improved disease prevention and public health.

From the initial phases of human-computer interaction (HCI), where the computer was unaware of the users' mental states, we are now progressing towards cognition-aware user interfaces. One crucial cognitive state considered by research on cognition-aware user interfaces is the cognitive load. Eye-tracking has been suggested as one particularly unobtrusive method for estimating cognitive load. Although the accuracy of cognitive load detection has improved in recent work, it is still insufficient for cognition-aware user interfaces, which require high accuracy for getting accepted by the user. This paper introduces two new eyetracking metrics for estimating perceived cognitive load based on benign anisocoria (BA). Unlike previous pupil-based metrics, our metrics are based on pupil size asymmetry between the left and right eye. As a case study, we illustrate the effectiveness of the proposed metrics on a recently published eye-tracking dataset recorded under laboratory conditions. The results show that our proposed features based on BA can improve the performance of classifiers for detecting the perceived mental workload associated with an N-back test. The best classification accuracy was 84.24% while the classification accuracy in the absence of the proposed features was 81.91% for the light gradient boosting classifier.

As the volume of end-of-life lithium-ion batteries (LIB) used in electric vehicles increases, so does the demand for recycling to recover valuable active materials as secondary resources, reducing the need for primary raw materials and minimizing supply dependencies. In addition to pyro- and hydrometallurgy recycling processes, an increasingly examined approach is the direct recycling route, which involves disassembling battery systems into modules, modules to cells, and ultimately cells to electrodes. To minimize passive materials such as various metals from the housing, joints, arresters, plastics, foams, and adhesives in the subsequent crushed mass, the introduction of module disassembly processes into the existing LIB recycling process chain is promising for the resulting increase in recyclate purity. This study presents the conceptualization of a LIB module disassembly process chain which aims to elevate the depth of disassembly of LIB during early recycling stages. A reverse engineering approach is used to recover individual cells from two battery modules of one specific automotive manufacturer by removing their casing through milling and disconnecting the cells from the arresters. The experimental results are analyzed for influences on the disassembly possibilities for individual components as well as the interactions within the system, enabling the elaboration of a priority graph and a disassembly plan. In order to allow for scalable recycling process chains suitable for very high quantities, a concept for a fully automated and variant-compatible robot application based on the experiment is derived. Furthermore, the depth of disassembly is compared with the corresponding reduction of passive materials in the crushed mass. The findings indicate that several module components can be disassembled, resulting in cell recovery and a reduction of metal and plastic in the crushed mass by about 17 %, highlighting the importance of robotic disassembly in enhancing material separation efficiency during LIB recycling.

Objective: the paper aims to address the surveillance of long-term bedridden patients. Although often necessary, continuous monitoring cannot be carried out due to cost constraints and limited personnel availability. Current solutions involve wearable devices and cameras, but they have limitations, as discomfort and concerns about privacy. Method: firstly, a thorough review of existing literature has been carried out. Then, it was developed the innovative system, focusing on bed monitoring instead of direct patient control. After identifying necessary patient condition data, various sensors were tested. To leverage data potential, the authors explored methods of centralizing and analyzing information. A platform was designed to assist operators in anticipating accidents. Finally, Digital Twins and Cyber Physical Systems were considered for their potential value in the system. Results: The proposed system monitors patient's bed without direct contact, in contrast to wearable devices. This innovative approach utilizes applied sensors capable of identifying risky behaviors, signs of specific pathologies, unusual movements, tremors, or abnormal humidity in the bed. The benefits include improved service levels, reduced operator surveillance, freeing up time for value-added activities, timely intervention where necessary, prevention of bed falls and sores, detection of wet beds, enhancement of sleep quality, and monitoring of weight trends. Conclusions: Literature indicates that lack of bed surveillance in healthcare is a global issue expected to worsen due to the progressive aging of the population. Authors’ 4.0 solution can effectively address this problem, improving service levels by monitoring and simulating patient behavior rather than direct monitoring, thus minimizing patient discomfort.

The rapid expansion of generative artificial intelligence (AI) technologies is projected to have a significant effect on electricity use in the global data center sector. Some earlier research has proposed using data centers for load-balancing the future power grid to allow higher integration of variable renewables. In this paper, we review the expected future electricity consumption of AI and evaluate AI data center’s potential roles in a clean energy system. Our work finds that the levelized cost of computing (LCOC) favors higher load factors and shows relatively low sensitivity to electricity price levels. Under our baseline cost assumptions, a baseload electricity price of $125/MWh benefits load factors higher than 64%, depending on market price conditions and variations. The findings reveal that a boom in AI energy use could drive significant baseload power demand in the future power system.

Brain-Computer Interface (BCI) is a technology that enables direct communication between the brain and external devices, typically by interpreting neural signals. BCI-based solutions for neurodegenerative disorders need datasets with patients’ native languages. However, research in BCI lacks insufficient language-specific datasets, as seen in Odia, spoken by 35-40 million individuals in India. To address this gap, we developed an Electroencephalograph (EEG) based BCI dataset featuring EEG signal samples of commonly spoken Odia words. Using the OpenBCI Cyton device, EEG recordings are collected from a volunteer who speaks Odia language. The dataset is divided into 4 parts: (i) vocal Odia words, (ii) English translations of these Odia words, (iii) sub-vocalization of the Odia words, and (iv) sub-vocalization of English words. The dataset contains information about 100 different words. Each word is recorded with ten trials. By training the dataset using Machine Learning (ML) and Deep Learning (DL) methods, a BCI system can be designed to translate EEG signals into both vocal and subvocal for the Odia and English languages. This can enhance the communication and quality of Odia-speaking patients with neurodegenerative diseases.

Robot Vision is the technique of enabling robots to process visual data from the environment by utilizing a combination of camera hardware and computer algorithms. Advanced deep neural networks have significantly played a vital role in indulging robots to make more sense out of complex visual data at different circumstances, especially in object detection and continuous tracking. In this research, we initiated a unique and cutting-edge backbone neural network for the conventional YOLO algorithm which we named as SBHK-Net. The network boosted up the performance of the existing YOLO algorithm drastically which manifests a strong potential of improving tracking and recognition accuracies of other conventional algorithms in the robot vision industry as well. It has the greatest accuracy 59.2% AP among all known real-time object detectors with 30 FPS or above on GPU RTX3060, and it outperforms all other known object detectors in the range of 5 FPS to 160 FPS. We used YOLOv7 as our reference point for the core research. The transformer-based detector SWINL Cascade-Mask R-CNN (9.2 FPS A100, 53.9% AP) and the convolutional detector ConvNeXt-XL Cascade-Mask R-CNN (8.6 FPS A100, 55.2% AP) are both outperformed by the SBHK-Net core object detector (56 FPS RTX3060, 56.4% AP) in terms of speed and accuracy, respectively. In terms of speed and accuracy, it surpasses a number of other object detectors, including DINO-5scale-R50, ViT-Adapter-B, Scaled-YOLOv4, YOLOv5, DETR, Deformable DETR, and YOLOR and YoLOX. The source code of this research is available at https://github.com/ac005sheekar/SBHK-Net.