The model compression techniques enable deploying compressed large language models (CMs) at network edge, facilitating convenient provision of AI-generated content (AIGC) services. To ensure timely delivery of these services, efficient placement of CMs across resource-constrained edge networks is essential. In this work, we investigate how to obtain latency-efficient CM placement across resource-constrained edge networks. With the objective of service latency optimization, we formulate the CM placement in resource-constrained network (CPRN) problem and establish its NP-hardness. We propose the Migration-agent-based Deep Reinforcement Learning (M-DRL) approach, which incorporates a specially designed migration agent tailored for such placement problems. To enhance training efficiency, we incorporate efficient heuristic placement results into the environment of M-DRL, developing our Heuristic-guided M-DRL (HM-DRL) approach. Our extensive simulation results demonstrate that HM-DRL outperforms an extended benchmark in service latency, while maintaining a low training overhead.

With the widespread adoption of large language models (LLMs) like GPT, user experiences in various interactive applications have significantly improved. However, reports from OpenAI highlight that GPT clients are now facing high response delays and frequent interruptions, particularly during peak usage hours, due to limited computation resources. This challenge is expected to escalate as machines are interacting with GPT models at higher frequencies, with greater data volumes, and over longer lifecycles. A promising solution is to deploy LLMs across edge networks to efficiently distribute the huge resource demands. This work presents the very first efforts at exploring how to cost-effectively deploy expert models from a mixture of experts (MoE) LLM within edge networks. We introduce the expert models deployment in edge networks (EMD-EN) problem, focusing on optimizing deployment costs. To address this, we propose a novel least cost gain (LCG) measure for selecting appropriate physical nodes to host expert models and present a corresponding LCG-based expert models deployment (LCGEMD) algorithm. Extensive simulations show that our approach outperforms the benchmarks by an average of 17.31% and 36.98% in terms of deployment cost reduction.

Global navigation satellite system (GNSS) positioning can be significantly degraded due to multipath and non-line-of-sight (NLOS) signals in urban areas. Cellular vehicle-to-everything (C-V2X) technology provides new opportunities to enhance GNSS performance from a single intelligent vehicle by leveraging roadside GNSS (RSG) and C-V2X. Inspired by this, we propose an RSG-aided GNSS/LiDAR/IMU (RSG-GLIO) method to achieve reliable odometry and mapping, which leverages the high-quality double-differenced (DD) measurements provided by nearby RSG, effectively mitigating shared random errors such as multipath and NLOS. Our RSG-GLIO first estimates the absolute state of the vehicle using onboard sensors. Utilizing this initial positioning estimate, the proposed method introduces a coarse-to-fine selection scheme to identify consistent DD observations from available RSG measurements. Finally, the consistent roadside DD constraints are jointly optimized into factor graph optimization (FGO). Static and dynamic data are extensively evaluated using multiple RSG receivers deployed in the Hong Kong C-V2X testbed to evaluate the effectiveness of roadside-aided positioning. The results demonstrate a significant 36.6% improvement in terms of absolute positioning accuracy compared to the state-of-the-art GLIO method. Furthermore, we showcase the potential for employing RSG as low-cost base stations in dense urban areas. The data of our work is publicly accessible at https://github.com/DarrenWong/RSG-GLIO.

Global Navigation Satellite Systems (GNSS) are crucial for intelligent transportation systems (ITS), providing essential positioning capabilities globally. However, in urban canyons, the GNSS performance could significantly degraded due to the blockage of direct GNSS signals. The pseudorange measurements are largely affected and the conventional model of weighting observations is not suitable in urban canyons. This paper addresses these challenges by integrating Artificial Intelligence (AI), specifically deep learning, into GNSS positioning process to enhance positioning accuracy. Traditional methods have primarily focused on pseudorange correction due to the absence of ground truth for weight estimation. In response, we propose an innovative indirect training approach using deep learning to optimize both pseudorange bias and weight estimation, aiming to minimize the positioning errors. To support this integration, we developedpyrtklib, a Python binding for the open-source RTKLIB tool, bridging the gap between traditional GNSS algorithms, typically developed in Fortran or C, and modern Python-based AI frameworks. Comparative analyses demonstrate that our method surpasses established tools like goGPS and RTKLIB in positioning accuracy, marking a significant advancement in the field. The source code of tightly coupled deep learning and GNSS integration, along with pyrtklib, is available on GitHub at https://github.com/ebhrz/TDL-GNSS and https://github.com/IPNL-POLYU/pyrtklib.

The recent surge in the development of autonomous vehicles has increased the need for reliable dynamic positioning of road agents in urban areas. Doppler frequency measurement of the global navigation satellite system (GNSS) can provide dynamic information and be used to estimate velocity. However, similar to pseudorange, the accuracy of Doppler frequency is degraded in dense urban areas, due to signal reflections from obstacles, resulting in substantial velocity errors. Existing methods tend to directly exclude non-line-of-sight (NLOS) Doppler frequency, which in turn leads to insufficient measurement numbers. 3D mapping aided (3DMA) GNSS ray-tracing method is commonly used to estimate extra delay from NLOS. The angle of arrival (AOA) of the reflected signal is obtained while tracing the propagation path, which can also be used to simulate NLOS Doppler frequency. Thus, this paper investigates the potential of using NLOS Doppler frequency as a feature for estimating velocity. A novel candidate-based 3DMA GNSS velocity estimation framework is proposed using Doppler frequency to examine the consistency between simulation on each candidate and measurement. Experimental results show NLOS Doppler frequency feature can enhance velocity estimation accuracy, reducing the root-mean-square error of velocity estimation by 46% from 1.06 m/s to 0.57 m/s in dense urban areas.

Traffic classification is a fundamental tool for network management, measurement and security. As the new services with diversified QoS requirements are evolving, traffic classification plays a more significant role in ensuring end-to-end performance guarantees. Several efforts have introduced deep learning (DL) to train traffic classifiers without manual features, however, these classifiers depend heavily on numerous samples and their quality. Indeed, it is unrealistic to obtain sufficient and representative samples in the underlying real network. In this paper, we propose a highly accurate traffic classification model by an Ensemble Learning framework, where Convolutional Recurrent Neural Networks are integrated into the Bagging to train the classifier using only small-scale datasets. Specifically, ensemble learning employs a combinatorial design of three base models to build more accurate and robust learning models so that absorbing features opens up the possibility of training classifiers on small sample sets. We conduct comprehensive experiments with a real-world dataset encompassing 20 applications. Extensive experiment results demonstrate that even with a mere 10% of training samples, our proposed model attains a classification accuracy of 93.94% and a classification precision of 94.33%, outperforming multiple other cutting-edge methods.

To provide dedicated protection in Network Function Virtualization (NFV), a reliability-aware Service Function Chain (SFC) can be deployed using two disjoint paths: a primary path and a backup path. In the event of network failures along the primary path, the working traffic is switched to the backup path to maintain service continuity. The process of accommodating reliability-aware SFCs is commonly referred to as SFC Deployment and Protection (SFCDP), which is proven to be NP-hard. In this work, we introduce a novel Network Service Deployment (NSD) transformer that can effectively incorporate and utilize fine-grained information regarding the network re-sources and the SFC requests. We develop a deep reinforcement learning framework based on NSD transformer (DRL-NSD) to effectively optimize the process of SFCDP. We conduct extensive simulations to validate the NSD transformer and compare the proposed NSD transformer with two benchmark neural network architectures. Our experimental results demonstrate that the NSD transformer outperforms the benchmarks across a variety of network topologies and network load settings.

Smooth and accurate global navigation satellite system (GNSS) positioning for pedestrians in urban canyons is still a challenge due to the multipath effects and the non-line-of-sight (NLOS) receptions caused by the reflections from surrounding buildings. Factor graph optimization (FGO) attracts more and more attention in GNSS society for improving urban GNSS positioning by effectively exploiting the measurement redundancy from historical information to resist the outlier measurements. Unfortunately, the FGO-based GNSS standalone positioning is still challenged in highly urbanized areas. As an extension of the previous FGO-based GNSS positioning method, the potential of the pedestrian dead reckoning (PDR) model in FGO to improve the GNSS standalone positioning performance in urban canyons is exploited in this paper. Specifically, the relative motion of the pedestrian is estimated based on the raw acceleration measurements from the onboard smartphone inertial measurement unit (IMU) via the PDR algorithm. Then the raw GNSS pseudorange, Doppler measurements, and relative motion from PDR are integrated using the FGO. Given the context of pedestrian navigation with a small acceleration most of the time, a novel soft motion model is proposed to smooth the states involved in the factor graph model. This paper verified the effectiveness of employing the PDR model in FGO step-by-step through two datasets collected in dense urban canyons of Hong Kong using smartphone-level GNSS receivers. The comparison between the conventional extended Kalman filter, several existing methods, and FGO-based integration is presented. The proposed method shows better results than the conventional FGO method in all test datasets, with at least a 22% decrease in the mean value of positioning error. The proposed method reduces the average localization error from 31.64 m to 18.51 m in a deep urban area.

Positioning in urban areas is still a challenge due to non-line-of-sight (NLOS) and multipath reception. This paper explores the geometrical characteristics of the GNSS ranging measurement by a graphical representation to better indicate the pseudorange consistency, which can be used to mitigate the NLOS and multipath receptions. The graphical representation is created by the grid-based method combined with the single differenced technique, which is called the single differenced residual map (SDRes Map). With the graphical properties of the SDRes Map, four main focuses of the NLOS/multipath problems, including positioning, signal status prediction, satellite weighting calculation, and NLOS/multipath error calculation, are able to be tackled simultaneously and demonstrated to have superior performance against the conventional or even state-of-the-art method methods.