Clustered vehicle routing problems (CluVRPs) represent a complex class of combinatorial optimization problems with significant real-world relevance. They extend classic VRPs by introducing pre-specified customer clusters and requiring effective routing both between clusters and within each cluster. While numerous deep learning approaches have been developed to address the standard VRP, research on CluVRPs remains relatively limited, presenting opportunities and challenges for advancing solutions to more practical VRPs with cluster-related constraints. This paper offers a deep reinforcement learning (DRL) approach to solving CluVRPs. We propose a cluster-aware attention module in the encoder, along with inter-cluster and intra-cluster decoders to specialize the constructive policies within and between clusters. Symmetrical data augmentation is adopted in the training to improve the performance. Empirical results in different CluVRP variants manifest that the DRL method outperforms existing approaches, consistently offering advantages for various instances.

Electric buses (EBs) play a crucial role in achieving global greenhouse gas emission targets. However, efficiently operating an electric bus fleet (EBF) requires a comprehensive approach that considers both mobility and energy systems, particularly when implementing opportunity charging strategies. Existing literature and many real-life implementations often focus on only one of these systems, oversimplifying the other, which can lead to inefficiencies, operational challenges, or even unfeasible implementations. To fill this gap, we propose a framework to assess the impact of bus opportunity charging strategies on the power grid by integrating a traffic simulation model (SUMO) and a power grid simulation model (Gaia). SUMO evaluates the energy consumption and charging needs of the EBs, while Gaia assesses the impact of the transformer load in the distribution grid. The integrated method is applied to Rotterdam’s bus line 36 to demonstrate the practicality of this approach. Results indicate that designing an electric bus route with opportunity charging is feasible only when both mobility and energy systems are carefully coordinated.

Existing activity-based and agent-based simulations alone often failed to capture the interaction between individual activity scheduling and detailed urban traffic dynamics. ActivitySim provides a representation of individual activity schedulings but often lacks detailed traffic dynamics, whereas MATSim can capture detailed interactions between travellers and mobility systems but often overlooks several decision-making factors, such as activity scheduling shift, household interactions and land-use influences. To address these limitations, this paper presents an Activity- and Agent-based Co-simulation framework that integrates ActivitySim and MATSim, both of which are open-source software popularly adopted in each research community. ActivitySim generates individual activity schedules and location choices, which serve as synthetic travel demand input for MATSim. MATSim then simulates detailed mobility interactions, with its outputs aggregated into zonal level-of-service matrices and fed back to ActivitySim for iterative scheduling adjustments. The feedback loop bridges the strengths of both models and is applied to the MRDH (Rotterdam-The Hague Metropolitan) region in the Netherlands. The initial MRDH model for the base-year reference scenario demonstrates that the proposed co-simulation framework effectively replicates existing mobility patterns, paving the way for fine-grained intervention evaluations like ride-hailing services in the future.

Artificial Intelligence (AI) agents capable of autonomous learning and independent decision-making hold great promise for addressing complex challenges across various critical infrastructure domains, including transportation, energy systems, and manufacturing. However, the surge in the design and deployment of AI systems, driven by various stakeholders with distinct and unaligned objectives, introduces a crucial challenge: How can uncoordinated AI systems coexist and evolve harmoniously in shared environments without creating chaos or compromising safety? To address this, we advocate for a fundamental rethinking of existing multi-agent frameworks, such as multi-agent systems and game theory, which are largely limited to predefined rules and static objective structures. We posit that AI agents should be empowered to adjust their objectives dynamically, make compromises, form coalitions, and safely compete or cooperate through evolving relationships and social feedback. Through two case studies in critical infrastructure applications, we call for a shift toward the emergent, self-organizing, and context-aware nature of these multi-agentic AI systems.

Ride-pooling has the potential to offer a sustainable solution for urban mobility by reducing vehicle use and emissions through shared trips. However, its adoption remains limited due to poor matching performance. Many requests fail to form feasible pools, and even successful matches often involve long detours or minimal cost savings. These inefficiencies largely arise from fragmented market structures: most operators act independently, restricting matching to their own request pools and limiting the formation of beneficial coalitions. Aggregation platforms improve efficiency by integrating regional operators through unified dispatch systems, but raise concerns over long-term stability. Differences in operator cost structures and market shares may incentivize deviation, at the same time, passengers may reject assigned payments if more attractive alternatives exist. To address these challenges, we propose a multi-level coalition formation game that jointly models operator and passenger collaboration. At the upper level, operators play a non-cooperative game to decide coalition partners. At the lower level, passengers are grouped into shared trips through a cooperative game that ensures individually rational payments. The two layers are coupled via constraint propagation, forming a unified decision-making process. We evaluate our framework using real-world data from three Chinese regions—Chengdu, Haikou, and the Ningxia Hui Autonomous Region—chosen to reflect diverse urban and regional contexts. Compared to independent operations, our approach increases vehicle occupancy by 14%–28%, reduces total costs by 10%–15%, and shortens average travel distances by 4%–5%. The system maintains stable coalition structures with operator deviation rates below 6.81% and near-zero passenger deviation rates.

Demand prediction is essential for effective management of Mobility-on-Demand (MoD) systems, as accurate forecasts enable better resource allocation, reduced wait times, and improved user satisfaction. Beyond that, probabilistic prediction methods that explicitly account for uncertainty are particularly valuable, as it allows decision-makers to assess risk and make robust plans under uncertain operational environments. However, most existing approaches focus on point predictions, which fail to capture the full spectrum of possible future outcomes. For probabilistic prediction, many methods typically rely on strong parametric distributional assumptions that may not accurately reflect the complex real-world environments. Nonparametric methods proposed in the literature, although promising, often suffer from high computational costs and model complexity, limiting their practical applicability. To overcome these challenges, we propose the Spatial-Temporal Graph Convolutional Network Variational Autoencoder (STGCN-VAE), a novel deep learning framework designed for uncertainty-aware probabilistic travel demand prediction in MoD services. The STGCN-VAE effectively captures complex spatial-temporal dependencies and inherent uncertainties in MoD demand data, generating diverse and realistic future demand scenarios and constructing comprehensive demand distributions. Specifically, the proposed framework integrates three key components: a Spatial-Temporal Graph Convolutional Network (STGCN) to learn complex spatial-temporal dependencies, a Variational Autoencoder (VAE) to compress these patterns into a latent space, and a Kernel Density Estimation (KDE) module to accurately construct probabilistic demand distributions and quantify uncertainties. Experiments on four different real-world MoD datasets including both rideshare and bikeshare services across different cities demonstrate that STGCN-VAE consistently outperforms state-of-the-art baselines in both point and probabilistic prediction, highlighting its robustness and broad transferability across service modes and urban contexts.

To reduce the vehicle relocation rate considering relieving disequilibrium of the supply-demand ratios across regions for car-sharing systems, in this paper, we propose a data-driven optimization framework by integrating the non-parametric learning algorithm and two-stage stochastic programming modeling technique to address the one-way station-based car-sharing relocation problem. In contrast with the most existing work that deals with demand uncertainty using predefined probability distributions, the learning-based framework is capable of handling demand uncertainty by learning the intrinsic pattern from large-scale historical data and computing high quality solutions. To validate the performance of our proposed approach, we conduct a group of numerical experiments based on New York taxicab trip record data set. The experimental results show that our proposed data-driven approach outperforms the parametric approaches and deterministic model in terms of business profit, relocation rate, and value of stochastic solution (VSS). Most significantly, compared with the deterministic approach, the vehicle relocation rates are reduced by approximate 80%, 70% and 40% under small fleet size, medium fleet size and large fleet size, respectively. In addition, the VSS of our approach is more than 3 times higher than the one of Poisson distribution by average.