Predicting the future trajectories of dynamic traffic actors is a cornerstone task in autonomous driving. Though existing notable efforts have resulted in impressive performance improvements, a gap persists in scene cognitive and understanding of complex traffic semantics. This paper proposes Traj-LLM, the first to investigate the potential of using pre-trained Large Language Models (LLMs) without explicit prompt engineering to generate future motions from vehicular past trajectories and traffic scene semantics. Traj-LLM starts with sparse context joint encoding to dissect the agent and scene features into a form that LLMs understand. On this basis, we creatively explore LLMs' strong understanding capability to capture a spectrum of high-level scene knowledge and interactive information. To emulate the human-like lane focus cognitive function and enhance Traj-LLM's scene comprehension, we introduce lane-aware probabilistic learning powered by the Mamba module. Finally, a multi-modal Laplace decoder is designed to achieve scene-compliant predictions. Extensive experiments manifest that Traj-LLM, fueled by prior knowledge and understanding prowess of LLMs, together with lane-aware probability learning, transcends the state-of-the-art methods across most evaluation metrics. Moreover, the few-shot analysis serves to substantiate Traj-LLM's performance, as even with merely 50% of the dataset, it surpasses the majority of benchmarks relying on complete data utilization. This study explores endowing the trajectory prediction task with advanced capabilities inherent in LLMs, furnishing a more universal and adaptable solution for forecasting agent movements in a new way.

The ever-increasing and unbalanced traffic load in cellular vehicle-to-everything (C-V2X) networks have increased the network congestion and led to user dissatisfaction. To relieve the network congestion and improve the traffic load balance, in this paper, we propose an intelligent software defined C-V2X network framework to enable flexible and low-complexity traffic offloading by decoupling the network data plane from the control plane. In the data plane, the cellular traffic offloading and the vehicle assisted traffic offloading are jointly performed. In the control plane, deep learning is deployed to reduce the software defined network (SDN) control complexity and improve the traffic offloading efficiency. Under the proposed framework, we investigate the traffic offloading problem, which can be formulated as a multi-objective optimization problem. Specifically, the first objective maximizes the cellular access point (AP) throughput with consideration of the load balance by associating the users with the APs. The second objective maximizes the vehicle throughput with consideration of the vehicle trajectory by associating the delay-insensitive users with the vehicles. The two objectives are coupled by the association between the cellular APs and the vehicles. A deep learning based online-offline approach is proposed to solve the multi-objective optimization problem. The online stage decouples the optimization problem into two sub-problems and utilizes the ‘Pareto optimal’ to find the solutions. The offline stage utilizes deep learning to learn from the historical optimization information of the online stage and helps predict the optimal solutions with reduced complexity. Numerical results are provided to validate the advantages of our proposed traffic offloading approach via deep learning in C-V2X networks.

Line-of-sight (LOS) blockage is a crucial problem in millimeter-wave vehicle-to-vehicle (V2V) communication due to its severe penetration loss and high vehicle mobility. In order to overcome the LOS blockage problem, in this paper, we propose using neighbor vehicles as relays to forward the blocked traffic flows. Specifically, a traffic-aware relay vehicle selection is investigated. First, analytic hierarchy process (AHP) is adopted to handle the situation that different traffic types have different preferences on the performances of rate, delay and data dropping ratio. Second, we introduce a coalitional game (CG) to evaluate the relative performances of the relay vehicles. Finally, by combining the results of AHP and CG, a heuristic relay selection scheme is devised by selecting the relay vehicle with the best rationality degree. Simulation results show that the proposed scheme can adapt to the requirements of different traffic types and improve the performances of rate, delay and data dropping ratio when LOS blockage occurs.

The air ground integrated networks can leverage unmanned aerial vehicle (UAV) communications to tackle the ever-increasing and unbalanced traffic load in future communication systems. This paper investigates the UAV enabled traffic offloading problem in air ground integrated networks with mixed user traffic. The problem jointly maximizes the system load balance and the total UAV reward, which can be formulated under a two-layer network graph model. In the cellular network graph, the association between the delay-sensitive users and the access points (APs) as well as the association between the UAVs and the APs are formulated. In the UAV network graph, the association between the delay-insensitive users and the UAVs is formulated. By observing the coupling relationship of the decision variables, we decouple the problem into three sub-problems and solve the first two sub-problems with reduced complexity. Then, we devise a Deep Neural Network (DNN) empowered genetic algorithm to solve the last sub-problem. The DNN can be leveraged to filter out the non-optimal solutions in the initialization operator of the genetic algorithm for improving the efficiency. Performance comparisons are provided between the proposed traffic offloading scheme and the existing ones, which validate the advantages of the DNN empowered genetic algorithm regarding its convergence, accuracy, and robustness.

To provide a better support for various vehicular applications, digital twin (DT), as an emerging technology, can enable a virtual presentation of physical vehicular networks to reflect the current network state through real-time data updating. However, the constrained resources and high data updating cost may degrade the performance of DT. In this paper, we trade off the data updating cost and the performance of DT to adaptively determine the resource management and computation offloading in vehicular networks. Specifically, we propose a novel vehicle to vehicle pairing prediction algorithm assisted by DT to improve the offloading decision efficiency and investigate the effect of data updating frequency on prediction accuracy. Based on the prediction results, we formulate a joint data updating frequency selection, offloading decision and channel allocation problem with the objective of minimizing the computation and communication costs. To solve the formulated problem, we propose a prediction-based stability maximum pairing algorithm to obtain the proper task offloading strategy. Moreover, a deep Q-learning network algorithm is proposed to select the optimal DT data updating frequency according to the real-time vehicular network state. Based on the obtained optimal solution, we further propose an alternating direction method of multipliers-based iteration algorithm to optimize the computation and channel resource allocation and minimize the total costs. Numerical results are provided to validate the effectiveness and efficiency of our proposed algorithms.

In this paper, a social-aware virtual medium access control (SV-MAC) is devised to integrate virtualization and social awareness to D2D communications underlying heterogeneous cellular networks. SV-MAC is a converged network layer. Through SV-MAC, heterogeneous networks can be unified in a single protocol stack, providing a common abstraction for users and mobile network operators (MNOs) to share their network resources. Leveraging users' social information, SV-MAC can realize social-aware D2D discovery, association, and resource allocation for improved D2D communication efficiency. In order to validate the advantages of the proposed SV-MAC protocol, we study the cellular and D2D resource allocation problems under the SV-MAC protocol. The problems are formulated as maximizing the cellular and D2D energy efficiency (EE). The cellular resource allocation problem is first solved considering its priority to D2D users. The resource allocation problem for D2D users is solved after cellular users to control the D2D-to-cellular interference within a tolerable threshold. In the simulation, we introduce real mobile user traces to simulate the user social information. Simulation results show that the proposed SV-MAC resource allocation schemes obtain a beneficial EE improvement comparing with conventional schemes.