Federated learning (FL) enables collaborative model training across distributed devices while preserving privacy. However, growing heterogeneity in device resources and communication links challenges conventional FL, especially when relying on a single central server. Hierarchical federated learning (HFL) mitigates these issues by organizing devices into clusters coordinated through intermediate aggregators. Yet, the effectiveness of HFL critically depends on how clusters are formed: intra-cluster communication must be efficient, device computational capacities should be balanced to reduce stragglers, and data heterogeneity must be managed to ensure stable convergence. In this work, we propose a learning-topology co-optimization framework for HFL in networks where nodes communicate with each other with links of varying quality (e.g., device-to-device (D2D) or mesh networks). Our method jointly optimizes device connection topology and learning directions, leading to communication-efficient clusters that remain well aligned in optimization space. We provide a convergence analysis under mild assumptions, showing how inter- and intra-cluster divergence affect learning stability. Extensive experiments demonstrate that our approach consistently improves HFL performance, yielding at least a$6 \%$accuracy gain under unbalanced data distributions and over$16\%$reduction in training time for regression tasks compared with existing clustering algorithms.

In this paper, we present an enhanced convolutional model for indoor radio map generation, focusing on the integration of a novel ray-marching feature. We describe our machine learning pipeline developed for the ICASSP 2025 Signal Processing Grand Challenge, specifically the First Indoor Pathloss Radio Map Prediction Challenge. Our method incorporates a ray-marching feature that, combined with a UNet architecture enhanced by dilated convolution layers, significantly improves indoor pathloss prediction accuracy. Our approach achieved a 3rd-place ranking in the challenge with a weighted RMSE of 10.33 on the test dataset.

Predictive Quality of Service (PQoS) makes it possible to anticipate QoS changes, e.g., in wireless networks, and trigger appropriate countermeasures to avoid performance degradation. A promising tool for PQoS is given by Reinforcement Learning (RL), a methodology that enables the design of decision-making strategies for stochastic optimization. In this manuscript, we present PRATA, a new simulation framework to enable PRedictive QoS based on AI for Teleoperated driving Applications. PRATA consists of a modular pipeline that includes (i) an end-to-end protocol stack to simulate the 5G Radio Access Network (RAN), (ii) a tool for generating automotive data, and (iii) an Artificial Intelligence (AI) unit to optimize PQoS decisions. To prove its utility, we use PRATA to design an RL unit, named RAN-AI, to optimize the segmentation level of teleoperated driving data in the event of resource saturation or channel degradation. Hence, we show that the RAN-AI entity efficiently balances the trade-off between QoS and Quality of Experience (QoE) that characterize teleoperated driving applications, almost doubling the system performance compared to baseline approaches. In addition, by varying the learning settings of the RAN-AI entity, we investigate the impact of the state space and the relative cost of acquiring network data that are necessary for the implementation of RL.

Federated learning is a fast-developing distributed learning scheme with promising applications in vertical domains such as industrial automation and connected automated driving. The heterogeneity of devices in data distribution, communication, and computation, when deployed in dynamic environments typically with wireless communication, poses challenges to traditional federated learning solutions, where successful learning depends on balanced contribution from participants. In this paper, we propose a flexible communication strategy for devices in asynchronous federated learning, which adapts the training and uploading actions based on the condition of the communication link. We propose a novel method of computing aggregation weight based on model distances and number of local optimizations, to control errors introduced in asynchronous aggregation while maximizing learning speed. We prove the convergence of the learning tasks analytically under the new scheme. The improved performance is rooted in the increased number of optimizations during training, which grows by 12% through opportunistically condensing model uploading during good link condition periods. By facilitating timely communication between devices and server, combined with the novel aggregation weight design, our method reduces the communication resources in dynamic environments by at least 5% while even slightly increasing the learning accuracy.

Terahertz (THz) communication is considered as one of the potential candidate technologies in the sixth generation (6G) wireless systems. This paper introduces channel measure-ments in two sub-THz bands, i.e., 140- and 220-GHz bands, in an outdoor street canyon environment for both line-of-sight (LoS) and non-line-of-sight (NLoS) scenarios with a frequency-domain vector network analyzer (VNA)-based sounder. Based on the measurement results, we computed and analyzed the statistical features of wireless propagation channels, including path loss, root mean square (RMS) delay spreads (DSs), azimuth spreads of arrival (ASA), and elevation spreads of arrival (ESA). Moreover, we observe the birth and death of clusters over a straight trajectory under the NLoS condition. A high-resolution param-eter estimation algorithm, i.e., the space-alternating generalized expectation-maximization (SAGE) algorithm, was employed to eliminate the effects of antenna patterns and the density-based spatial clustering of applications with noise (DBSCAN) algorithm was used to clustered the multipath components (MPCs). The obtained statistical properties of measured channels and obser-vations on channel evolution can be employed in THz channel modeling and system design.

Federated Learning is a fast-developing distributed learning scheme that has promising applications in vertical domains such as industrial automation and connected automated driving. In this paper we address the heterogeneity of the participation of devices in federated learning caused by: i) non-uniform distribution of local data; ii) uneven and varying computational resources across the devices; and iii) dynamic communication link. We propose a quasi-dynamic simulation scheme allowing realistic approximation of these three factors of heterogeneity. Aggregation schemes at the server based on the clients’ work status are implemented. We show that the new asynchronous aggregation algorithm does not require tuning of hyper-parameters such as the round time in synchronous federated learning and the aggregation weight in classic asynchronous aggregation, while providing better or comparable performance in terms of accuracy and convergence speed.

Advanced vehicular applications are foreseen to rely on wireless connectivity. By predicting wireless network performance, application adaptations can be done a priori to ensure robust and efficient operations. However, determining accurate predictions of network performance or factors affecting it is challenging in highly dynamic vehicular environments. In this paper, we address the problem of determining prediction intervals for future observations. Specifically, prediction intervals are determined for wireless link quality between two vehicles that communicate directly with each other. Evaluations based on real-world vehicle-to-vehicle dataset show that the proposed mechanism is able to determine prediction intervals with the desired confidence level in dynamic scenarios. Furthermore, the mechanism is capable of identifying the intervals that are unreliable with respect to coverage requirements.

We derive an analytical expression of the error vector magnitude (EVM) in terahertz (THz) communication links affected by two performance limiters, namely antenna alignment error and the receiver's in-phase/quadrature imbalance (IQI). Through the sole use of elementary functions, we show how to compute the EVM in closed-form approximately, but accurately and efficiently. Furthermore, we show that antenna alignment errors and IQI can severely degrade the performance of THz communications. Investigations of the special and limiting cases of the regimes pertaining to high signal-to-noise ratio (SNR) and low antenna alignment error variance are carried out, where we also demonstrate the possible separation of the IQI and THz channel effects on the performance of the EVM.

Motivated by the teleoperation and local map sharing vehicular communication use cases, we investigate whether uplink throughput can be predicted by different machine learning approaches. First, we perform measurements of the vehicle to infrastructure (V2I) uplink throughput in Munich, Germany. Then, we use the collected measurements to evaluate whether linear regression (LR), deep neural network (DNN), and random forest (RF) can predict the uplink throughput. Our results show that, while very easy to train, LR is overly simple in describing the relationship of the input features and the predicted uplink throughput. On the other hand, DNN and RF can provide a very good prediction of uplink throughput (below 0.5 Mbps mean absolute error for a 40 Mbps uplink connection), while requiring longer training. Irrespective of the employed model, our results show that the best indicator of uplink throughput is signal to interference and noise ratio (SINR). When location information is added to SINR, the prediction error can be further reduced, albeit slightly. In the absence of SINR, location information is the second best in predicting uplink throughput. However, it can be employed only for locations that were available in the training dataset. On the other hand, SINR allows for generalization to locations different to those observed in the training dataset.

In recent years, autonomous networks have been designed with Predictive Quality of Service (PQoS) in mind, as a means for applications operating in the industrial and/or automotive sectors to predict unanticipated Quality of Service (QoS) changes and react accordingly. In this context, Reinforce-ment Learning (RL) has come out as a promising approach to perform accurate predictions, and optimize the efficiency and adaptability of wireless networks. Along these lines, in this paper we propose the design of a new entity, integrated at the RAN level that implements PQoS functionalities with the support of an RL framework. Specifically, we focus on the design of the reward function of the learning agent, able to convert QoS estimates into appropriate countermeasures if QoS requirements are not satisfied. We demonstrate via ns-3 simulations that our approach achieves better results in terms of QoS and Quality of Experience (QoE) performance of end users in a teleoperated driving scenario.

Terahertz (THz) spectrum band has garnered a lot of interest recently, due to the advances in supporting radio electronics and the existence of hundreds of Gigahertz of available spectrum. Therefore, it is considered as one of the key enablers for the upcoming sixth generation (6G) wireless networks, which target to provide Terabit per second communication services. However, the propagation loss due to the high carrier frequency, obstacles, and the atmospheric attenuation in this frequency band need to be compensated to enable usable coverage ranges. In this paper, we investigate the potential of using Intelligent Reflecting Surface (IRS) for enhancing the coverage of Terahertz communication. We conduct a case study for an indoor cellular communication scenario with multiple Transmission and Reception Points supported by IRS. The numerical result show that the IRS can enable significant coverage extension. Besides, by optimizing the frequency resource allocation and IRS selection jointly considering the frequency-dependent atmospheric absorption, the throughput of underserved users can be improved significantly.

Advanced wireless communication use cases relying on sidelinks require guaranteed network performance. However, fulfilling the desired Quality-of-Service (QoS) requirements can-not be always ensured due to factors such as varying interference and dynamicity in the network. To reduce the impact of varying network performance on sidelink applications, we address the problem of long-term proactive adaptation of sidelink applications based on network performance. Specifically, we propose a modular algorithmic framework comprising of data-driven prediction models and model-based optimization to predict the feasibility of given application data rate requirements. Further-more, the framework is capable of predicting the simultaneously achievable data rate demands (in the max-min sense). A vehicular communication scenario is considered as an example to show the applicability of the framework. Evaluations show that variations in feasibility and achievable data rate demands can be well captured to enable robust application layer adaptations.

Cellular networks enable high reliability of vehicle-to-vehicle (V2V) communications thanks to centralized, efficient coordination of radio resources. Collision-free transmissions are possible, where base stations could allocate orthogonal resources to the vehicles. However, in case of limited resources in relation to the data traffic load, the resource allocation task becomes a challenge. Current solutions propose heuristic algorithms that focus on resource reuse, often based on the location of the vehicles. Such schedulers are mainly designed assuming ideal network coverage conditions and are prone to performance degradation in case of coverage loss. Further, they typically rely on frequent scheduling updates, which increases the dependency on coverage. In this paper, we propose a reinforcement learning-based approach to scheduling V2V communications. Our solution, called iVRLS, delivers higher reliability than an enhanced version of a state-of-the-art benchmark algorithm in case of intermittent coverage conditions, while requiring less frequent scheduling. Following this approach, we enable a unified scheduler deployment irrespective of coverage, which offers graceful performance behavior across varying coverage conditions, thus making iVRLS a robust alternative to existing schedulers.

To meet the Quality-of-Service (QoS) requirements of vehicular applications, some knowledge of future wireless channel statistics is essential. We address the problem of predicting channel quality between vehicles in terms of path loss which, exhibits strong fluctuations over time due to highly dynamic vehicular environment. We propose a framework for data-driven path loss prediction models that are obtained from datasets comprising information related to message transmissions and the communication scenario. By combining changepoint detection method and online learning, the proposed framework adapts the current prediction model based on its performance, thus accounting for the dynamics in the environment and the cost of re-training. Evaluations using real world Vehicle-to-Vehicle communications datasets show that adapting the prediction function using the proposed framework can achieve prediction accuracy comparable to that of online learning case, while significantly reducing the number of data samples required for re-training.

Predicting the performance of vehicular communication networks is challenging due to the interplay of multiple factors. One prominently influencing factor is the wireless channel between the transmitter and the receiver. We address the problem of predicting the path loss between two communicating vehicles by using a non-parameterized, data-driven approach. Specifically, we apply Random Forest, a non-parametric learning method, to real world vehicle-to-vehicle communications dataset and evaluate it with respect to its prediction accuracy and generalization capability. We show that availability of additional information to the non-parametric model results in better performance than the well known parameterized log distance path loss model. We further discuss the relative contribution of different features for the model accuracy and conclude that careful selection of features can achieve results nearly as accurate as using all available features. Finally, we discuss several aspects that need to be considered while using such data-driven prediction models along with applications of V2V path loss prediction.

We perform simultaneous vehicle-to-vehicle (V2V) multi-band double-directional measurements at 6.75 GHz, 30 GHz, and 60 GHz in a street canyon environment. We analyse both line-of-sight channels and those obstructed by vehicle blocker. For the different bands, we analyse delay spread, angular spread, vehicle blockage loss, and path loss. Finally, we discuss how the system aspects influence the radio channel.

Reliable and accurate multipath clustering results are essential to design and parameterize geometry based (stochastic) channel models proposed by organizations such as 3GPP, COST, WINNER and others. Clustering within wireless channel models assumes that multipath components arrive/depart with similar properties in the considered parameter domains. We propose a new approach combining tracking and initialization of the classic K-Means clustering algorithm. The results show a significantly improved consistency for cluster estimation, which subsequently allows for analyzing the lifetime of these clusters. Finally, we apply the proposed approach on two urban macro channel sounding datasets and compare them to the current 3GPP standard. We found that the angular parameters are rather similar to the standard whereas the cluster delay spreads and the number of rays per cluster differ significantly.

We explore a new approach to radio resource allocation for vehicle-to-vehicle (V2V) communications in case of out-of-coverage areas that are delimited by network infrastructure. By collecting and predicting information such as vehicle velocity, density and message traffic, the network infrastructure ensures reliability of the V2V services. We propose reserving required amount of resources for services that cannot be pre-scheduled (e.g., emergency braking, crash notifications, etc.), and scheduling those services that can be pre-scheduled (e.g., platooning). We analyze the resource reservation as a function of target reliability under varying vehicle densities and sizes of out-of-coverage area. For pre-scheduled services, we explore how variations in the vehicle velocities and predictions affect successful transmissions. The results indicate that increase in required reliability does not penalize the system prohibitively. On the other hand, speed prediction errors decrease the transmission success rate considerably, thus calling for a more flexible scheduler design.

Presents the introductory welcome message from the conference proceedings. May include the conference officers' congratulations to all involved with the conference event and publication of the proceedings record.

Due to the differences in terms of antenna height, scatterer density, and relative speed, V2I links exhibit different propagation characteristics compared to V2V links. We develop a geometry-based path loss and shadow fading model for V2I links. We separately model the following types of V2I links: line-of-sight, non-line-of-sight due to vehicles, non-line-of-sight due to foliage, and non-line-of-sight due to buildings. We validate the proposed model using V2I field measurements. We implement the model in the GEMV2 simulator, and make the source code publicly available.

We investigate the evolution of line of sight (LOS) blockage over both time and space for vehicle-to-vehicle (V2V) channels. Using realistic vehicular mobility and building and foliage locations from maps, we first perform LOS blockage analysis to extract LOS probabilities in real cities and on highways for varying vehicular densities. Next, to model the time evolution of LOS blockage for V2V links, we employ a three- state discrete-time Markov chain comprised of the following states: i) LOS; ii) non-LOS due to static objects (e.g., buildings, trees, etc.); and iii) non-LOS due to mobile objects (vehicles). We obtain state transition probabilities based on the evolution of LOS blockage. Finally, we perform curve fitting and obtain a set of distance- dependent equations for both LOS and transition probabilities. These equations can be used to generate time-evolved V2V channel realizations for representative urban and highway environments. Our results can be used to perform highly efficient and accurate simulations without the need to employ complex geometry-based models for link evolution.

It is our great pleasure to welcome you to the 3rd IEEE Workshop on Smart Vehicles: Connectivity Technologies and ITS Applications (SmartVehicles'16), which is held in conjunction with the 13th IEEE International Symposium on a World of Wireless Mobile and Multimedia Networks (WoWMoM'16). The aim of this workshop is to bring together academics, industry professionals, and application developers to presents recent developments, current research challenges, and future directions in the area of connected and smart transportation.

Vehicular networks will enable a number of active safety and traffic efficiency applications. At the core of many of those applications is cooperative awareness: the ability to detect location, speed, and heading of surrounding vehicles. We empirically analyze three key metrics that shed light on the communication performance available to applications: Packet Delivery Ratio: link quality in terms of the proportion of received messages over distance; Neighborhood Awareness Ratio: the proportion of detected neighbors within a given distance, which serves as an indicator of the effectiveness of cooperative awareness message exchange; and Neighborhood Interference Ratio: the proportion of neighbors above the desired range of interest, which can provide insight into the interference levels of fully deployed systems. By analyzing the measurement data collected within the scope of the DRIVE-C2X project, we conclude that the link layer delivery and neighborhood awareness criterion can be fulfilled for safety applications: in the analyzed datasets, the cooperative awareness ratio is close to 100% up to 100 meters. Depending on the desired region of interest, the interference from far-away vehicles can be considerable, thus requiring effective congestion control to balance between neighborhood awareness and interference.

Vehicle-to-Vehicle (V2V) communication is a core technology for enabling safety and non-safety applications in next generation intelligent transportation systems. Due to relatively low heights of the antennas, V2V communication is often influenced by topographic features, man-made structures, and other vehicles located between the communicating vehicles. On highways, it was shown experimentally that vehicles can obstruct the line of sight (LOS) communication up to 50 percent of the time; furthermore, a single obstructing vehicle can reduce the power at the receiver by more than 20 dB. Based on both experimental measurements and simulations performed using a validated channel model, we show that the elevated position of antennas on tall vehicles improves communication performance. Tall vehicles can significantly increase the effective communication range, with an improvement of up to 50 percent in certain scenarios. Using these findings, we propose a new V2V relaying scheme called tall vehicle relaying (TVR) that takes advantage of better channel characteristics provided by tall vehicles. TVR distinguishes between tall and short vehicles and, where appropriate, chooses tall vehicles as next hop relays. We investigate TVR's system-level performance through a combination of link-level experiments and system-level simulations and show that it outperforms existing techniques.

We propose a probabilistic key distribution protocol for vehicular network that alleviates the burden of traditional public-key infrastructures. Roadside units act as trusted nodes and are used for secret-sharing among vehicles in their vicinity. Secure communication is immediately possible between these vehicles with high probability. Our performance evaluation, which uses both analysis and simulation, shows that high reliability and short dissemination time can be achieved with low complexity.

A thorough understanding of the communications channel between vehicles is essential for realistic modeling of Vehicular Ad Hoc Networks (VANETs) and the development of related technology and applications. The impact of vehicles as obstacles on vehicle-to-vehicle (V2V) communication has been largely neglected in VANET research, especially in simulations. Useful models accounting for vehicles as obstacles must satisfy a number of requirements, most notably accurate positioning, realistic mobility patterns, realistic propagation characteristics, and manageable complexity. We present a model that satisfies all of these requirements. Vehicles are modeled as physical obstacles affecting the V2V communication. The proposed model accounts for vehicles as three-dimensional obstacles and takes into account their impact on the LOS obstruction, received signal power, and the packet reception rate. We utilize two real world highway datasets collected via stereoscopic aerial photography to test our proposed model, and we confirm the importance of modeling the effects of obstructing vehicles through experimental measurements. Our results show considerable obstruction of LOS due to vehicles. By obstructing the LOS, vehicles induce significant attenuation and packet loss. The algorithm behind the proposed model allows for computationally efficient implementation in VANET simulators. It is also shown that by modeling the vehicles as obstacles, significant realism can be added to existing simulators with clear implications on the design of upper layer protocols.