ABSTRACT. This talk addresses the locomotive assignment problem in freight rail networks under schedule uncertainty, defined as early or late train departures and arrivals. It presents a two-stage stochastic formulation on scenario-dependent space-time networks. The first stage determines a nominal weekly assignment of locomotives, work events (pick-ups and set-outs), and light-travel movements, while the second stage restores the plan’s feasibility under delay realizations by reconstructing locomotive flows on perturbed networks. The recourse model penalizes adjustments according to their operational impact, distinguishing minor flow shifts from costly ad-hoc light-travel movements and increases in the locomotive fleet. To solve the resulting large-scale stochastic fixed-charge network problem, a Benders decomposition approach is introduced to improve computational tractability. Computational experiments on large freight-rail instances show that the stochastic model substantially reduces disruption-recovery costs relative to deterministic planning, highlighting the value of resilient locomotive scheduling.

ABSTRACT. This paper studies a freight transportation setting in which an Intelligent Decision Support Platform (IDSP) coordinates time-sensitive shipper requests with carrier service offers in a Many-to-One-to-Many (M1M) system, with applications in LTL networks. Because the platform does not own transportation assets, it must jointly decide (i) strategic allocations of capacity across heterogeneous resource types and (ii) tactical scheduled service network design on a time–space network, including demand acceptance, routing, and consolidation. The model captures two key uncertainties: fluctuating shipper demand and the availability of non-contract capacity, which can trigger costly rerouting, unmet service, and higher environmental impact. We adopt a risk-averse formulation based on Conditional Value-at-Risk (CVaR) to hedge against unfavorable demand–capacity realizations and obtain plans that are robust beyond expected-value performance. We develop a tailored Benders decomposition algorithm, and computational experiments on realistic networks compare risk-neutral and CVaR solutions, quantifying the value of risk-averse planning for platform and carriers.

ABSTRACT. This paper studies a multi-period relay network design problem for long-haul freight transportation, motivated by driver shortages in the truckload trucking industry. A relay network planner determines the locations and capacities of relay points and commits to service-time guarantees, while freight carriers decide whether to adopt the relay service based on transportation cost and travel time. This interaction is formulated as a bilevel optimization model capturing network design and user adoption under uncertain preferences. User adoption is characterized using a probabilistic criterion comparing relay transportation with direct transportation. Congestion at relay points is modeled through service and waiting times using queuing theory. The bilevel model is reformulated as an equivalent single-level mixed-integer second-order cone program. A path-based reformulation and a logic-based Benders decomposition algorithm are developed to improve computational efficiency. Numerical experiments and a large-scale case study demonstrate the effectiveness of the proposed approach.

ABSTRACT. Over the past decade, ride-sharing services have become increasingly important, with platforms like Uber/Lyft facilitating billions of rides annually. This rise reflects their ability to meet users' convenience, efficiency, and affordability needs. However, in busy areas and surge zones, the benefits of these platforms can diminish, prompting riders to relocate to cheaper, more convenient locations. While much research has focused on the strategic behavior of drivers, the strategic actions of riders (walking outside of surge zones) remain under-explored. This paper examines the impact of rider-side strategic behavior on surge dynamics. We investigate how riders' actions influence market dynamics, including supply, demand, and pricing. We show significant impacts, such as spillover effects where demand increases and prices surge even in areas adjacent to surge zones. Our theoretical insights and experimental results highlight that rider strategic behavior helps redistribute demand, reduce surge prices, and clear demand in a more balanced way.

ABSTRACT. Private motorized transport contributes to congestion, pollution, and inefficient use of urban space. Bike-sharing systems offer a low-emission alternative, but their success depends on the system-wide availability of operable bikes and functional stations. In practice, bikes become defective and systems are subject to vandalism. As maintenance resource are often limited to repair all stations on a daily basis, issues accumulate over time. Providers face a trade-off between short-term operational goals (minimizing unresolved issues) and long-term tactical goals (balanced service levels across stations) to support system acceptance. In a multi-period routing framework, we jointly consider operational and tactical goals. Our approach integrates short-term issue minimization with long-term balancing of station states through a cost function approximation that combines historical and expected future issue accumulation. Using real-world data from New York City, we show that our method achieves low overall issue levels while substantially improving service balance compared to benchmarks.

ABSTRACT. Ride-sourcing studies are usually constrained by the lack of access to data owned by private platforms. We develop a methodology to estimate high-resolution ride-sourcing demand (e.g., at the census-tract level) using low-resolution public data, such as town-level records from Massachusetts.Publicly available socio-demographic, employment, land-use, and built-environment variables are used to estimate two models: (1) a trip generation model to produce robust estimation results across spatial aggregation levels using geographically weighted regression (GWR); and (2) a multinomial logit model for destination choice given origin. Results show that towns with shorter commute times, higher black population proportions, and higher transit commuter proportions tend to generate and attract more ride-sourcing trips. Our GWR model maintains relatively stable performance at the boundaries. Tract-level origin–destination flows are constructed by proportionally disaggregating town-level flows using estimated parameters, preserving aggregate patterns while capturing local variation, particularly between low-flow natural areas and high-flow urban centers.

ABSTRACT. In on-demand food delivery, order dispatching must balance immediate dispatch and order holding under competing objectives, including delivery efficiency, service timeliness, and courier utilization. Holding orders may improve consolidation opportunities but can increase the risk of late delivery, requiring real-time decisions about when and how to dispatch orders. We propose a hierarchical framework that integrates offline inverse optimization with on-policy deep reinforcement learning. The framework determines which orders to dispatch or hold and groups dispatched orders via a multi-objective weighted set partitioning model. Objective weights are learned using an inverse optimization approach and embedded into a Markov decision process. A dispatch policy is then trained using Proximal Policy Optimization. Using data from a major food delivery platform, the learned policy outperforms current practice and benchmarks across key operational metrics. The results show that strategic order holding improves consolidation and efficiency.

ABSTRACT. Online food delivery platforms increasingly rely on hybrid fleets comprising full-time employees and part-time contractors who make their own strategic decisions. This paper addresses the operational challenge of jointly optimizing courier dispatch and pricing while satisfying the incentive-compatibility constraints of strategic part-time couriers. We first characterize the optimal policy for a single-restaurant setting, showing that it generates a distance-based customer partition between courier types. We then extend this framework to a general multi-restaurant network with order batching, formulating the problem as a bilinear optimization program. Validated against the Meituan dataset from the First INFORMS TSL Data-Driven Research Challenge, our proposed solution achieves lower operational costs and shorter latency compared to benchmark policies.

ABSTRACT. We study an urgency-based dispatching problem for app-based meal-delivery platforms operating with a hybrid workforce of full-time couriers and third-party gig workers. Orders arrive over time, each with a distance and a promised delivery deadline, which together define an implied riding speed capturing the time pressure faced by the courier. We model courier safety and stress through a convex, non-decreasing penalty in implied speed for orders served by full-time couriers, and include a distance-based cost for orders outsourced to gig workers. The platform’s objective is to minimize long-run average operating cost by dynamically deciding whether each arriving order should be served by a full-time courier or by a gig worker. A semi-Markov Decision Process is formulated for this problem. We propose a threshold-based policy to solve it and benchmark it against an offline set-covering model solved by column generation. Numerical results demonstrate that our policy performs well relative to benchmark.

ABSTRACT. Distribution logistics faces heterogeneous and uncertain service times driven by repeated driver-customer interactions, familiarity with delivery locations, and social dynamics. Behavioral effects that significantly shape last-mile performance yet remain underexplored in routing models, limiting efficiency gains and understanding of underlying mechanisms. We develop a multi-period vehicle routing model that incorporates empirical customer-driver interaction data to capture service time heterogeneity and learning effects. Service time estimation is decoupled from routing, enabling the use of various learning and service time models while maintaining a generally applicable multi-period structure. A computational study using data from a German grocer’s distribution network shows that by applying strategic driver assignment service times are reduced by 5.5% within a single period. The relevant variable part of the service time can actually be reduced by up to 32%. This work integrates driver behavioral learning into distribution planning and demonstrates how (machine) learning models can enhance multi-period routing frameworks.

ABSTRACT. We present offline model-based routing optimization (Offline MBRO), a framework for optimizing routing and allocation decisions using historical data. Offline MBRO has two steps: (i) learn a surrogate that predicts routing costs, and (ii) embed the surrogate within a mixed-integer program to optimize allocations. Exploiting permutation invariance over sets of customers, vehicles, and depots, we propose a Deep Sets surrogate that generalizes across vehicle routing problem variants and can capture behavior-driven costs such as driver-specific travel times. On instances of the traveling salesman problem, capacitated vehicle routing problem, and vehicle routing problem with time windows (VRPTW), the surrogate attains under 1% mean percentage error, whereas classical continuous approximation methods consistently overestimate costs. On a real order assignment setting in last-mile delivery, Offline MBRO reduces late deliveries from 20% historically to 4%, compared to 12% for a VRPTW baseline, and decreases conditional delay when late.

ABSTRACT. In the Close Enough Traveling Salesman Problem (CETSP), one must find the shortest tour that travels through every given customer-centered neighborhood, usually a disk centered on each customer or target. In this study, we combine prior work on CETSP estimation models and on using results from a randomized heuristic on combinatorial optimization problems such as the TSP, VRP, Split Delivery VRP (SDVRP), and Knapsack Problem in order to obtain strong solution value estimates efficiently. Here, we use heuristics based on the Generalized TSP (GTSP) and merging nodes in Steiner zones. With our regression-based approach, we are able to obtain estimates on instances that would otherwise take hours to solve to optimality. A broader motivation for this research stream is to show that we can obtain excellent solution value estimates to a wide variety of hard combinatorial optimization problems.

ABSTRACT. Severe rainstorms create nonstationary, heavy-tailed travel time variability that violates the stationary i.i.d. assumptions commonly used in stochastic traffic assignment. We propose a risk- and ambiguity-averse truncated-logit stochastic user equilibrium (TSUE) framework for hazard-prone networks, where OD-specific admissibility thresholds eliminate unrealistic assignment to unacceptable routes. Tail-risk sensitivity is incorporated through a mean–CVaR certainty equivalent (confidence level α, mixing weight λ). To hedge against distributional misspecification and regime shifts, we further develop a 1-Wasserstein distributionally robust TSUE over ambiguity sets with radius ρ, including a structured regime-dependent extension. Reformulations based on Rockafellar–Uryasev and Wasserstein duality yield tractable conic programs solvable via Benders decomposition with analytic cuts. In a downtown Chicago experiment using 2023 precipitation data, conservative CVaR reduces flow on vulnerable radial routes by 40.5%, while DRO robustness increases flow on reliable corridors by 27.3% and reduces flow on hazard bottlenecks by 17.5%, demonstrating protective, stability-oriented reallocation under extreme weather.

ABSTRACT. We propose a distributional view of Wardrop’s user equilibrium (UE) by defining a Gibbs posterior over feasible link flows, using the Beckmann potential as an energy function. The inverse temperature $\beta$ controls how strongly the distribution concentrates around near-optimal solutions, turning UE from a single point estimate into a controllable family of feasible alternatives. A key challenge is feasibility: samples must satisfy flow conservation and nonnegativity at every step. We address this with a projected Langevin sampler in link-flow space, where each noisy gradient update is followed by an Alternating Direction Method of Multipliers (ADMM) Euclidean projection based on an origin--destination (OD) decomposition. Experiments on the Sioux Falls benchmark show a stable, predictable contraction of posterior uncertainty as $\beta$ increases, reveal that dependence across links is largely local, and visualize leading modes of variability through principal component analysis (PCA) loadings on the network.

ABSTRACT. Betweenness centrality is widely used in transportation as a proxy for infrastructure criticality and vulnerability, with route choice models motivating visitation-based betweenness measures. We instead study uncertainty in network state through evolving realizations with random edge availability and travel times. Each realization defines a betweenness maximizer. When disconnected, we use the maximizer within the connected component containing the current node. We show reveal order in a realization tree does not affect the compressed dynamics, yielding an absorbing Markov chain (AMC) on $V \cup \{\perp\}$. AMC represents normalized occupancy before absorption and equals probability of a uniformly sampled step before absorption under sampling proportional to episode length. When the kernel is available only through simulation, we estimate AMC and multi-reward variants from pooled episodes with consistency and asymptotic normality. We propose robust AMC using row-wise Kullback-Leibler neighborhoods, yielding exponentially tilted worst-case kernels and ranking stability diagnostics.

ABSTRACT. Cargo-hitching allows Public Transit operators to exploit an additional revenue stream by offering excess capacity to Logistic Service Providers. In this context, we propose the first methodology for estimating the total cargo throughput of a cargo-hitching system. Such a methodology can then be used as a tool in public policy decision-making regarding infrastructure investments. To this end, we formulate a novel stochastic network design problem and apply Sample Average Approximation. We develop a Progressive Hedging algorithm to find solutions for instances of realistic size in which we solve the single scenario subproblems via a Column Generation-based heuristic. Moreover, we compiled ridership data from the Metropolitan Transportation Authority to derive managerial insights for a case study on New York City, United States. We present a suitable network design for our case study and assess the potential of cargo-hitching as well as its performance drivers.

ABSTRACT. Transportation demand estimation models are typically calibrated using coarse numerical data with low spatial and temporal granularity. By building low-dimensional representations of complex travel patterns, demand estimation methods using coarse data often yield incomplete estimates of true demand. Meanwhile, qualitative data offer opportunities to capture traveler experiences at more granular spatiotemporal levels. Yet, integrating qualitative data into demand estimation models remains an open question, as such data may often under- or over-represent certain populations. In this paper, we address the question of how to integrate qualitative data into demand estimation. We consider the example of detecting overcrowding levels from comments in transit systems, which become key metrics in recovering historical demand. Through an analytical model and a case study for Washington D.C. Metrorail, we characterize how comment bias impacts demand estimation quality. Our findings highlight the importance of representative comment collection and the impact of comment-based demand estimation methods.

ABSTRACT. We introduce the Expeditionary Logistics Network Design Problem (ELNDP), a new formulation for operational-level planning in expeditionary environments where multi-modal vehicle coordination is critical and penalties for unmet demand dominate transportation costs. ELNDP extends the classical Scheduled Service Network Design Problem by incorporating flexible commodity sourcing and heterogeneous vehicle capabilities, both essential in military logistics. We propose an iterative refinement algorithm based on dynamic discretization discovery (DDD) that iteratively constructs consolidation plans on partially time-expanded networks. Unlike the classical DDD framework, our approach overestimates arc travel times and introduces backward recovery arcs to compute relaxed solutions. We develop a new procedure for eliminating illegal vehicle cycles arising from explicit vehicle management, and introduce acceleration techniques based on capacity factors and a multi-commodity maximum-flow heuristic. A case study with the U.S. Marine Corps shows that our method doubles demand fulfillment and reduces solve times by nearly 30% compared to benchmark approaches.

ABSTRACT. Solving large-scale service network design problems is a central challenge in transportation logistics. While Dynamic Discretization Discovery (DDD) has proven effective when using time-expanded models, existing methods struggle with complex operational constraints, such as maximum wait times at terminals and First-In-First-Out (FIFO) loading models. We introduce a new DDD methodology based on a two-layer relaxation graph that separates arrival and departure events into distinct layers. This novel structure allows us to systematically identify and repair three types of temporal infeasibilities that arise due to the sparse time discretization. Our approach iteratively refines this sparse time-expanded network, guaranteeing convergence to a feasible operational solution without necessarily constructing the full model. The framework is applied within a two-stage planning process, translating frequency-based weekly delivery plans into detailed, executable schedules. Initial numerical results show our method is effective, demonstrating its potential to solve previously intractable, high-fidelity consolidation scheduling problems.

ABSTRACT. The Continuous-Time Service Network Design Problem (CTSNDP) aims to optimize transportation planning over a continuous-time horizon at minimal total cost. To provide greater flexibility and achieve cost reductions, the CTSNDP with Splittable Deliveries (SCTSNDP) allows each commodity to be divided into smaller shipments which may follow different routes and schedules. Although the dynamic discretization discovery (DDD) approach, the best-known exact algorithms for the CTSNDP with Unsplittable Deliveries (USCTSNDP), can be applied to the SCTSNDP by dividing each commodity of multiple units into multiple identical commodities of one unit, this direct approach significantly increases the commodity size and undermines computational efficiency. To address this challenge, we develop a new DDD algorithm for the SCTSNDP, which demonstrates how the DDD approach can be applied in the presence of splittable deliveries. It achieves an efficiency in solving the SCTSNDP that is comparable to that of the best-known DDD algorithm for the corresponding USCTSNDP.

ABSTRACT. Supply chains are exposed disruptions more frequently, while firms observe these events independently, leading to an information asymmetry. Although information sharing can improve coordination and performance, competing firms may strategically keep information due to competitive concerns. This study examines a voluntary information-sharing platform that enables competing firms to exchange disruption-related information while maintaining operational decision autonomy. We model firms’ one-time participation decisions as a static game in a two-echelon supply chain network subject to stochastic disruptions and private signals. The analysis characterizes equilibrium platform participation and shows how information power, competition, and network structures jointly impact firms' incentives. Numerical experiments demonstrate that firms with strong information advantages may hesitate to participate when competition is intense or networks are dense, whereas lower information gaps support participation. The results provide insights into when voluntary platforms can effectively mitigate information asymmetry and improve supply chain performance.

ABSTRACT. As supply chains become more interconnected with broader reach and access, customer expectations have risen. As a result, supply chains have had to adapt and become more efficient to meet these expectations. These efficient designs leave little room for things to go wrong, but due to increasing disruptions, companies must consider resilience as part of their supply chain design to mitigate the impact of these disruptions and remain competitive. In this paper, we investigate the vulnerability of e-commerce distribution networks in three disaster-prone regions of the U.S., through a three-step methodological approach that involves: 1) realistic network generation, 2) network disruption that simulates the impacts of natural disasters, and 3) network evaluation through causal mediation analysis. We show that the most cost-efficient distribution networks are also the most vulnerable, and we conclude by highlighting the pivotal role of network structures in preserving resilience and informing more resilient network design.

ABSTRACT. Ensuring the resilience of critical supply chains, particularly for essential commodities such as fuel, pharmaceuticals and emergency response supplies is crucial when facing disruptions ranging from natural disasters to intentional interdictions. While the strategic pre-allocation of safety stock is an effective mitigation strategy for short-term disruptions, it entails significant costs, requiring an optimized balance between limited budgets and operational continuity. This study addresses this challenge by proposing a tri-level Defender-Attacker-Defender (DAD) model applied to a multi-echelon supply chain. The formulation seeks to determine an optimal safety stock allocation plan that minimizes system costs against worst-case network interdiction. To address the computational complexity of this problem, we implement an exact Column-and-Constraint Generation (C&CG) algorithm. Finally, we validate our model through a case study on the Alaska Fuel Distribution Network. Our results reveal managerial insights that quantify the relationship between inventory pre-positioning investment and worst-case system resilience.

ABSTRACT. Committing a fleet size at the tactical level in advance of daily operations requires accounting for day-to-day variability in customer requests. We address this problem by modeling fleet sizing as a newsvendor problem, in which per-item costs are replaced by per-vehicle costs and unserved demand costs are represented by penalties for unserved requests. A key challenge is estimating the number of requests that can be fully served by different fleet sizes. To tackle this, we propose continuous approximation models that capture the structure of the problem from a high-level perspective. Our approach requires low computational effort while providing actionable managerial insights. Specifically, we show that our total cost function is convex. We use our methods to assess the optimal fleet size and its cost under different levels of information availability and for different parameter configurations. We also validate our approximations on a case study in a real-road network.

ABSTRACT. Collection and delivery points (CDPs) enable logistics operators to consolidate multiple customer delivery requests on a single vehicle stop, thereby reducing distribution costs. However, for customers to adopt CDPs, they must be willing to travel to a nearby CDP to pick up their parcels. This choice depends on the customer’s proximity to CDPs and the economic incentives they receive for using CDPs. We propose a continuous approximation model that jointly recommends the optimal size of the CDP network and economic incentives to customers choosing CDPs, with the goal of minimizing total expected costs. In our experiments, we show that CDPs alone can reduce costs by 16.9%, while offering incentives to customers increases potential savings to 28.0%. We find that CDPs are most beneficial in high-density areas, leveraging economies of scale, whereas incentives are more effective in low-density regions.

ABSTRACT. This study investigates a hybrid last-mile delivery system in which a truck continuously dispatches and retrieves sidekicking drones while in motion. To move beyond instance-specific routing decisions, we develop a continuous approximation framework that characterizes average system performance under spatially stochastic demand. The hybrid routing problem is represented using a striping approximation, while service congestion arising from limited drone availability is captured through a queueing-theoretic model. The resulting formulation links truck traversal time to key system parameters, including demand density, vehicle speeds, and the number of drones. We validate the proposed approximation through simulation and heuristic routing solutions, demonstrating strong agreement across a wide range of operating conditions. Overall, the framework enables analytical characterization of the performance envelope of truck–drone collaborative delivery and informs strategic system design decisions relative to alternative delivery operations.

ABSTRACT. The \emph{Elementary Shortest‑Path Problem} (ESPP) seeks a minimum cost path from $s$ to $t$ that visits each vertex at most once. The presence of negative-cost cycles renders the problem $\mathcal{NP}$‑hard. We present a probabilistic method for finding near-optimal ESPP, enabled by an unsupervised graph neural network that jointly learns node value estimates and edge-selection probabilities via a surrogate loss function. The loss provides a high probability certificate of finding near-optimal ESPP solutions by simultaneously reducing negative-cost cycles and embedding the desired algorithmic alignment. At inference time, a decoding algorithm transforms the learned edge probabilities into an elementary path. Experiments on graphs of up to 500 nodes show that the proposed method surpasses both unsupervised baselines and classical heuristics, while exhibiting high performance in cross-size and cross-topology generalization on unseen synthetic graphs.

ABSTRACT. Rounded capacity cuts (RCCs) are among the most effective cutting planes for vehicle routing problems, but their exact separation is computationally intractable on large instances, forcing modern solvers to rely on heuristics. We propose a learning-based RCC separator built on a transformer architecture operating on a novel augmented graph induced by fractional LP solutions. The augmented graph preserves the sparse structure of the LP graph while incorporating selective multi-hop connectivity via bottleneck flow features, enabling efficient capture of long-range interactions without dense attention or recursive coarsening. Trained to imitate an exact separation oracle, our model predicts high-quality RCC-defining subsets in a single forward pass. Computational experiments on standard benchmarks demonstrate strong scalability and substantial improvements over existing neural separators on large instances, while a hybrid integration with classical heuristics yields compact relaxations and competitive performance on small and medium instances.

ABSTRACT. Routing problems such as the TSP and CVRP are computationally challenging due to their exponentially growing search spaces. This study proposes a supervised machine learning–based approach for edge classification in routing problems, predicting the likelihood that a given edge appears in an optimal solution. The resulting confidence scores are used as a preprocessing mechanism to reduce the decision space and to guide candidate selection similar to granular neighborhoods. Several supervised learning models are evaluated with respect to their predictive performance and their downstream impact on exact solvers and state-of-the-art metaheuristics. Computational experiments show that machine learning–based edge pruning reduces solver runtimes by over 50% on average for TSP instances solved to optimality with Gurobi, while decreasing average optimality gaps from 2.98% to 1.91% on harder instances, and yields consistent 13–16% runtime reductions for state-of-the-art metaheuristics without degrading solution quality.

ABSTRACT. This study introduces the perturbed utility Markovian traffic assignment model and establishes its associated equilibrium concept, termed perturbed utility Markovian equilibrium (PUME), as a variational inequality defined in the less constrained cost space. PUME generalizes existing Markovian traffic assignment models with more flexible and richer behavioral and utility structures. It features link flows as a continuously differentiable function of link utilities, which is further extended to accommodate more general flow-cost relationships. Moreover, computationally efficient algorithms are developed and demonstrated on medium-scale transportation networks with superior performance relative to state-of-the-art methods.

ABSTRACT. First-order continuum traffic models, including the Lighthill–Whitham–Richards (LWR) model, are traditionally formulated deterministically, leaving a gap in our ability to describe how stochastic upstream fluctuations propagate and interact with downstream traffic. This work proposes a novel framework for solving the LWR model with stochastic boundary conditions, where the upstream density evolves as a Markov process with upward jumps representing shock-inducing fluctuations. We proved that the resulting solution evolves according to a kinetic equation. This equation describes the statistical dynamics of traffic density over space and time, accounting for shockwave dynamics in a unified way. We validate our approach through numerical experiments and with real-world congestion trajectory data. Results show that our framework accurately reproduces shockwave statistics at significantly reduced computational cost, without requiring space-time discretization. This framework offers an analytical tool for modeling traffic under uncertainty, enabling more robust strategies for congestion prediction and control.

ABSTRACT. Large-scale traffic assignment is computationally challenging due to the combinatorial growth of feasible paths. While Beckmann’s formulation yields a convex user equilibrium problem, its path-based representation becomes intractable for metropolitan networks. This paper reformulates traffic assignment as a reduced-order convex optimization problem by exploiting the rapid singular value decay of the path–link incidence matrix. A major–minor flow decomposition is introduced, where minor path flows are compressed using singular value decomposition (SVD), yielding substantial dimensionality reduction. Combined with an augmented Lagrangian framework, the approach achieves over 90% variable reduction while preserving convexity and solution accuracy. Numerical experiments on large-scale GMNS benchmark networks, including the Chicago Regional network with over 2.8 million paths, demonstrate significant computational speedups with negligible loss in solution quality.

ABSTRACT. Accurately predicting when inbound passengers reach key terminal checkpoints is critical for proactive airport operations, yet is challenging due to behavioral heterogeneity and incomplete passenger data. This paper proposes a temporal–spatial passenger flow modeling framework that combines a flow-timing model with a complementary facility-attractiveness network. The flow-timing model learns the distributions of the checkpoint-reach times of inbound passengers using pre-determinable flight-level attributes, and enables flexible, nonparametric modeling. While fragmented and unobserved trajectories can induce bias in timing-based inference, we introduce a node-level attractiveness parameter that redistributes timing-based passenger flows across the terminal network via a gravity-style mechanism, which ensures spatial flow consistency. Using operational data from Singapore Changi Airport, the flow-timing model achieves strong agreement with empirical data (81.74% PDF overlap), a mean peak-arrival timing error of 2.78 minutes, and an average arrival passenger misalignment rate of 5.6%, highlighting the framework’s potential for scenario analysis and passenger-centric resource allocation.

ABSTRACT. Airport capacity expansion planning is important yet challenging due to the stochastic factors inherent in long-term air demand growth. This paper examines two key factors affecting expansion decisions. The first is demand uncertainty, which encompasses different demand growth patterns, short-term volatility, and sudden downward shocks. The second is present bias, whereby decision makers prefer short-term utility relative to long-term benefits. Ignoring present bias may lead to delayed project delivery due to different preferences among future authorities. We derive and compare expansion decisions for five types of decision makers: time-consistent, naïve conservative, naïve optimistic, sophisticated conservative, and sophisticated optimistic authorities. The results show that higher demand uncertainty and more conservative planning behaviors lead to higher trigger demand levels for capacity expansion.

ABSTRACT. This study addresses the integrated Airline Timetabling and Fleet Assignment Problem with Passenger Mix (ATFP), which simultaneously optimizes flight departure times and fleet assignments while accounting for passenger choice behavior. Despite its profitability potential, the ATFP often leads to computationally intractable Mixed Integer Programming (MIP) models for large-scale networks. To overcome this challenge, we propose a novel algorithm: Volumetric Adaptive Network Tailoring via Guided Exploration (VANTAGE). By leveraging information from solution pools and LP relaxations, VANTAGE iteratively refines the time-space network, focusing computational resources on high-potential regions. Benchmarked against commercial solvers and state-of-the-art methods using real-world data from Frontier Airlines, VANTAGE demonstrates superior scalability and efficiency, achieving higher profits with minimal memory requirements. Our approach effectively bridges the gap between complex theoretical models and large-scale practical airline operations.

ABSTRACT. This work develops a MILP model to support the long-term transition from diesel to battery-electric feeder buses in Quito, a system dominated by subsidised diesel and subject to ambitious decarbonization targets. The model integrates fleet replacement, charging infrastructure planning and environmental constraints within a unified optimisation framework, extending Pelletier et al. (2019) to include multiple electric technologies, heterogeneous charging strategies and container-based centralized charging.

Using operational and cost data from the El Labrador feeder system, we construct a 30-year transition plan and analyse how technology choices, charging configurations and investment timing respond to policy and energy-price scenarios. Results show that centralized charging reduces infrastructure needs and that transition trajectories are highly sensitive to diesel and electricity prices, especially early in the horizon. Sensitivity experiments further indicate that reducing diesel subsidies strengthens the case for early electrification. The framework offers a practical decision-support tool for agencies planning large-scale fleet electrification.

ABSTRACT. Service region design (SRD) is a fundamental challenge in operations planning, influencing sectors such as e-retailing and urban mobility systems. We study a sequential SRD problem where at most k regions can be invested per period, and model regional demand using a geometric Brownian motion with Poisson jumps to capture both gradual fluctuations and spillover-driven surges. Each investment sequence is evaluated using real options analysis (ROA), which determines optimal timing and yields its option value through Monte Carlo simulation. Because exhaustive ROA-based optimization becomes intractable at scale, we develop a dual-backward recursion evaluation algorithm and propose HS-GMSAC, a hierarchical search framework that combines a deep reinforcement learning (DRL) sequence generator with iterated local search heuristics. The DRL generator learns to produce high-value sequences using ROA-derived rewards. Numerical experiments demonstrate that HS-GMSAC achieves superior investment sequences and computational efficiency compared with benchmark approaches, enabling scalable optimization of large SSRD instances.

ABSTRACT. Urban freight transport generates significant congestion, pollution, and inefficiencies. To address these issues, policymakers are increasingly exploring incentive-based mechanisms that encourage shifts toward sustainable modes such as inland waterways and rail-based scheduled services. Their effectiveness, however, hinges on the strategic interactions among transportation authorities, logistics service providers (LSPs), and end customers. This study develops a multi-agency game-theoretic framework that captures these interdependencies by integrating regulatory decisions, LSP routing and pricing choices, and customer demand responses. The authority sets road taxes and subsidies for sustainable modes under a budget constraint; LSPs optimize profits by selecting modes and prices; customers choose services through a utility-maximizing logit model. The resulting bilevel formulation represents policy design at the upper level and market equilibrium at the lower level. A case study of Amsterdam evaluates three implementation strategies, offering comparative insights into their efficiency and practical feasibility for sustainable urban freight.

ABSTRACT. Precision agriculture depends on efficient data collection and treatment, requiring Unmanned Aerial Systems (UAS) to match traditional operational quality while managing unique energy and meteorological constraints. Dynamic factors such as wind, precipitation, and traffic often render pre-planned sorties infeasible. To address this, we propose a novel optimization framework for agricultural drone deployment that integrates high-resolution weather forecasts directly into the routing decision process.

ABSTRACT. This paper presents a dynamic optimization framework for Dockless E-scooter Sharing Systems (DESSS) designed to maximize service levels and profit. Using a mixed-integer linear programming model, the framework manages operator recharging, truck repositioning, and rider incentives while tracking individual scooter states and locations across distinct parking and non-parking zones. The study compares four scenarios: a baseline with no intervention, charging only, charging with repositioning, and a fully integrated approach. Results indicate that while charging alone doubles served trips compared to the baseline, the fully integrated scenario, combining charging, repositioning, and user incentives, achieves the highest served demand, lowest unmet demand, and greatest net return. The findings demonstrate that integrating granular tracking with operational strategies and incentives is essential for high-performance, cost-effective DESSS deployment, providing key insights for operators and urban planners.

ABSTRACT. Multimodal transportation addresses the first- and last-mile problem by integrating fixed-route transit services with flexible travel modes to support complete user trips. However, compared to direct car trips, multimodal trips may be less convenient, discouraging user adoption. To promote multimodal utilization, we propose the Incentive Bundle Program (IBP), which allocates subsidies to participating users. We develop a game-theoretic tri-level optimization framework that captures the interactions among three stakeholders in the IBP: the city-wide operator, the ridehailing operator, and users. The city-wide operator determines subsidy allocations, the ridehailing operator sets pricing, and users select travel modes based on their individual disutilities. We propose two plans that characterize the interaction between the city-wide and ridehailing operators and design a corresponding solution methodology for each plan. Using real-world data, we conduct numerical studies to analyze the impact of each plan on system-level VMT and ridehailing operator profits.

ABSTRACT. Connected and autonomous vehicles (CAVs) enable new forms of fleet coordination for on-demand mobility and delivery services, yet most deployment strategies still rely on fixed depots or unconstrained spatial positioning. This study proposes a novel fleet deployment strategy, the Base Loop with Coordination (BLC), in which idle vehicles are distributed along a loop-shaped road corridor and rebalanced to maintain uniform spacing. We develop queuing models that derive service times from analytically obtained trip distance distributions in a stylized radial–ringed city and evaluate performance through stochastic simulations. Results show that coordinated strategies significantly reduce blocking probability and system congestion compared with uncoordinated point-based deployments. Moreover, simulation results for point-based coordinated strategies are closely approximated by simplified M/M queuing models, supporting their use for practical evaluation. The proposed framework highlights the potential of loop-based spatial coordination as a scalable deployment approach for CAV-based on-demand services.

ABSTRACT. A corporate sponsor is evaluating a proposed “fast barge” container service that transloads at a Gulf Coast terminal and serves five inland Mississippi River corridors. We model weekly, multi-corridor pricing as a 52-week Markov decision process with Poisson demand whose mean follows a log-linear price elasticity curve, corridor capacities, and an optional shared fleet-hours constraint coupling corridors. A data-informed simulator combines shipment-manifest seasonality with sponsor-provided operating costs. We compare PPO and A2C (discrete price updates) and SAC (continuous updates) against cost-plus, a myopic demand-tracking heuristic, and an interpretable shadow-price (bid-price) baseline. Under baseline fleet-hours (780 h/week), SAC attains mean weekly reward of USD 3.18M, outperforming myopic (USD 1.14M) and matching cost-plus (USD 3.17M); the shadow-price baseline achieves USD 3.24M. Under tightened fleet-hours (500 h/week), the shadow-price baseline reaches USD 1.94M and retrained A2C and SAC remain competitive, highlighting the value of adaptive pricing when capacity binds.

ABSTRACT. There has been a global push toward the adoption of electric vehicles, which is driving the expansion of charging infrastructure. However, the availability of chargers and the limited power capacity at local grids are not expanding at the same pace as EV demand, creating an operational challenge of serving more vehicles with constrained resources. We study a captive charging station that primarily serves a dedicated fleet but admits external EVs for additional revenue. The chargers are flexible in their operational mode and can operate in either a fast or a slow mode. Using a Markov decision process framework, we model the joint problem of admitting external vehicles, assigning charging modes to incoming EVs, and dynamically adjusting charging modes during service to accommodate incoming requests. Our analysis proposes a priority-aware admission and power-adaptive mode switching policy to jointly enhance the station’s utilization and service performance under limited power.

ABSTRACT. As cities accelerate the transition to zero-emission bus fleets, a critical “digital divide” excludes data-scarce regions (Global South) from advanced planning tools. Existing transit electrification methods mainly rely on GTFS data and neglect the stochastic realities of extreme weather and underserved communities. This study proposes “Yeraz,” a dual-workflow decision support framework that utilizes Mixed-Integer Programming (MIP) to select a subset of bus routes for electrification. Unlike traditional cost-minimization models, we introduce a weighted objective function that simultaneously maximizes emissions reduction and a spatial equity score, allowing planners to generate a Pareto Frontier of optimal investment strategies. We incorporate chance constraints for energy consumption, explicitly addressing high variability of weather conditions worldwide. The framework is tested on the Yerevan, Armenia network. The results indicate that while ridership-driven heuristics prioritize high-frequency arterial routes (e.g., Route 41), the proposed equity-weighted optimization shifts investment to topographically challenging, underserved districts (e.g., Route 13).

ABSTRACT. Online Freight Exchange (OFEX) platforms have the potential to enhance freight transport efficiency by consolidating transportation demand. However, current algorithmic solutions do not adequately accommodate time-window constraints and fail to deliver sub-second decision performance required for operational deployment. To bridge this gap, we formally introduce the Multi-commodity One-to-one Pickup-and-Delivery Selective Travelling Salesperson Problem with Time Windows (m1-PDSTSP-TW). We present a novel learning-seeded hybrid search pipeline that incorporates geographical decomposition and a rolling-horizon approach. Our pipeline accelerates the solution process by using Lagrangian relaxation and a Deep Neural Network, Transformer Architecture, to rapidly generate constraint-aware seed solutions. These seeds are subsequently refined by a Large Neighborhood Search featuring a custom softmax-biased removal mechanism. Preliminary experiments on synthetic data confirm the pipeline's strength: the hybrid method consistently yields the highest revenue, with the neural constructor significantly cutting down computation time compared to standard metaheuristics.

ABSTRACT. All-electric ships are a promising solution for decarbonizing inland waterway transport, but their limited sailing range relative to conventional vessels requires explicit integration of energy refueling decisions into operational planning. This study investigates the joint optimization of cargo transport and energy refueling for AES operations in inland waterways. We first construct an integrated shipping network that connects the main waterway with its tributaries. Based on this network, we formulate a mixed-integer linear programming model that jointly optimizes pickup-and-delivery cargo flows and flexible on-site energy refueling decisions, including charging during port stays and battery swapping, to minimize total operational cost. To solve instances of practical scale, we develop a tailored branch-and-price framework enhanced by acceleration strategies, including a state-reduction-based construction heuristic, a graph-reduction-based heuristic-then-exact pricing scheme, and a multi-state labeling algorithm with two customized dominance rules. Computational experiments confirm the superior efficiency of the proposed algorithm and provide insights.

ABSTRACT. As autonomous vehicles (AVs) increasingly operate on public roads, understanding their real-world behavior is essential for safety analysis, policy design, and public trust. We propose SVBRD-LLM, a zero-shot framework that discovers, verifies, and applies interpretable behavioral rules from traffic videos. The system extracts trajectories via YOLOv8 and ByteTrack, computes kinematic features, and prompts GPT-5 to generate 35 structured hypotheses comparing AVs with human-driven vehicles. Hypotheses are validated and iteratively refined using failure cases to remove spurious correlations, yielding a high-confidence rule library. On an independent test set, SVBRD-LLM supports speed-change prediction, lane-change prediction, and AV identification. Using 1,500+ hours of real-world video, it achieves 90.0% accuracy and a 93.3% F1 Score for AV identification. The resulting rules highlight smoother speed control, more conservative lane changes, and more stable acceleration, each with context, semantics, and confidence.

ABSTRACT. The shipping industry is essential to global supply chain management. However, its significant carbon emissions pose a serious threat to the environment. In this context, Onboard Carbon Capture and Storage (OCCS) has emerged as a critical solution, providing an effective measure to lower emissions and support the transition toward green supply chain in shipping industry. Considering the interdependent decisions of ports and shipping companies, our research focuses on selecting ports and timing for carbon processing systems to maximize ports revenue, identifying vessels on specific routes for OCCS installation, and determining optimal ports for offloading captured carbon to minimize processing costs. To address these problems, we develop an integer programming model to optimize the decisions of both ports and shipping companies. Furthermore, we conduct numerical experiments and perform sensitivity analyses. Finally, we provide valuable insights for OCCS optimization and scheduling in the maritime industry.

ABSTRACT. Urban rail transit planning typically overlooks small-scale OD trips ("SmallODs"). While these "long tail" demands may have limited impact on network-wide averages, their stochastic path choices can create localized volatilities. This study proposes a "Boundary Stress Test" to screen for sections sensitive to SmallOD uncertainty. Using a Spatiotemporal Evaluation Matrix, we quantify the potential deviation in flow magnitude and peak timing under different routing scenarios. Results from the Nanchang Metro show that while the network remains generally stable, specific "Busy & Misaligned" critical sections are identified where SmallOD volatility causes peak hours to drift. This framework serves as a diagnostic tool, enabling operators to pinpoint and manage these hidden vulnerabilities without overhauling the entire schedule.

ABSTRACT. We propose a machine learning–based approach for the Multi-Depot Vehicle Routing Problem that learns customer-to-depot assignments from high-quality solutions. Our method relies on a graph neural network to capture routing-aware assignment patterns and generalizes well across different spatial customer distributions. Empirical results on synthetic instances show that the proposed model outperforms classical geometry-based heuristics in challenging settings and provides a scalable alternative for multi-depot vehicle routing solution methods.

ABSTRACT. Emerging agricultural supply chains often face uncertainty regarding when demand growth justifies investment in distribution infrastructure, such as regional warehouses. This study examines how uniform demand scaling affects the economic justification and timing of opening the first warehouse in a nationwide distribution network for industrial hemp products. Using real-world order data from a U.S.-based hemp processor, we develop a mixed-integer linear programming (MILP) model that determines optimal warehouse location and shipment routing decisions while accounting for transportation costs and carbon emissions. GIS-derived road distances are used to evaluate routing patterns under increasing demand levels. Through systematic demand-scaling experiments, we identify the point at which variable transportation cost savings from warehouse routing offset the fixed cost of facility establishment. The results quantify how firms’ distribution strategies change with increasing demand, identifying the demand level at which opening a warehouse and routing a portion of shipments through warehouses becomes optimal.

ABSTRACT. This paper proposes a computationally efficient dual-based inverse optimal transport approach for estimating preferred arrival time (PAT) distributions in bottleneck models. Although the Vickrey bottleneck model is a fundamental framework for analyzing departure-time choice under congestion, empirical applications remain limited due to the difficulty of identifying heterogeneous scheduling preferences. Building on recent results that formulate departure-time equilibrium as a one-dimensional optimal transport problem, we exploit the analytical characterization of equilibrium queueing delays as optimal dual variables. By parameterizing the PAT distribution and fitting the implied queueing-delay profile to observed data, we formulate a minimum-distance estimation problem defined entirely on the dual side of the equilibrium. Unlike inverse approaches based on observed flows, the proposed method remains informative even when the bottleneck operates at full capacity. Numerical experiments using synthetic data demonstrate that the method accurately recovers the shape and location of unimodal PAT distributions, highlighting its effectiveness and practical applicability.

ABSTRACT. We study the design of autonomous-vehicle (AV)–based last-mile transportation services in distributed regions where driver shortages and sparse demand make conventional operations costly and unreliable. AV deployment is constrained by geofences induced by Operational Design Domain (ODD) requirements, while long-term planning favors stable service partitions that may become misaligned as spatial demand evolves. We propose a generalizable spatial-temporal demand batching framework that jointly optimizes the depot location, district partitioning, fleet sizing, and dispatch scheduling under spatial demand ambiguity. Demand uncertainty is addressed using Wasserstein distributionally robust optimization, while ODD costs are modeled through a modular, feature-based structure capturing district-specific driving complexity. We develop a discretization-based reformulation that yields neat convex inner programs and computationally tractable solution methods despite complex modeling considerations. A case study using a nonprofit last-mile transportation provider in East Pittsburgh demonstrates significant performance improvements over fixed-route designs.

ABSTRACT. A critical requirement for dynamic traffic signal optimization (TSO) is a traffic dynamics model that is both accurate and computationally efficient. Existing analytical models often fail to capture complex traffic behaviors due to simplifying assumptions, while microscopic traffic simulations, though high-fidelity, are expensive for use within optimization loops. This study proposes a physics-structured modular surrogate modeling paradigm that preserves the analytical framework of traffic dynamics models while replacing analytically challenging components with data-driven modules trained on simulation data. A demonstration based on a store-and-forward structure is presented. The resulting link dynamics model is embedded in a large-scale dynamic TSO problem under a tight simulation budget. Numerical experiments show improved replication of traffic dynamics and, consequently, stronger dynamic signal plans compared with the original store-and-forward model. The proposed modeling paradigm is general and can be extended to other analytical models and transportation applications.

ABSTRACT. The integration of passenger and freight transport in urban areas has gained increasing attention in recent years, driven by the rising urban logistics demand and the long-standing challenges in transit operations. Yet, existing practices are hardly sustainable due to high capital investment and operational costs. This paper proposes to use modular autonomous vehicles (MAVs) to leverage the flexible and efficient coupling and decoupling of mobility and logistics modules. An integrated operational model is developed to optimize MAV schedules, vehicle composition, and delivery routes. A Nash bargaining mechanism is then implemented to determine the unit MAV delivery price that balances the net profits of transit and logistics operators, considering their negotiation power. The results show that integration demonstrates significant savings in freight delivery, while maintaining the transit service quality regardless of the passenger demand variations. Besides, both transit and logistics operators enjoy higher profitability thanks to the integrated system.

ABSTRACT. Weaving ramps are critical bottlenecks in highway networks due to conflicting flows and complex interactions among heterogeneous vehicles. This study investigates the aggregate lane-choice behavior of mainline vehicles approaching weaving zones and its implications for system performance. We first develop a Wardrop-equilibrium-based lane-choice model for human-driven vehicles (HDVs), validated using microscopic traffic simulations. Building on this foundation, we propose a bilevel Stackelberg–Wardrop framework for mixed traffic, in which connected and autonomous vehicles (CAVs) act as strategic leaders and HDVs respond according to Wardrop equilibrium. We analyze how CAV penetration influences equilibrium outcomes and overall congestion. The framework is further extended to heterogeneous settings using Social Value Orientation (SVO) to capture varying degrees of altruism. Analytical results identify interpretable thresholds for CAV penetration and altruistic behavior, beyond which system delay transitions from inefficiency toward socially optimal performance, highlighting the potential of strategically designed CAV coordination in mixed-autonomy traffic systems.

ABSTRACT. Efficient airport slot allocation is critical for mitigating congestion and reducing schedule displacement in coordinated airport networks. Most existing approaches either focus on single airports or assume fixed flying times, limiting scheduling flexibility under high demand. This study introduces a strategic stage airport network slot allocation model that explicitly incorporates flying time flexibility. Minimum and maximum flying times are derived from historical data, allowing arrival slots to adjust within feasible windows relative to departures. To solve the resulting large scale mixed-integer problem, a hybrid algorithm integrating variable-weight local search and deep reinforcement learning is developed. Computational experiments using one week of real flight data from 51 major Chinese airports demonstrate substantial reductions in total slot displacement compared with a fixed flying time benchmark, particularly during peak periods and at major hubs. Results highlight flying time flexibility as an effective operational resource for improving network level slot allocation efficiency.

ABSTRACT. This study investigates the efficiency–urgency tradeoff in on-demand food delivery services, specifically focusing on how the estimated time of arrival (ETA) influences driver behavior under network effects. We develop a micro-foundation model linking ETA tightness to bundle size and driver switching probabilities. A key contribution is the modeling of the "switching" behavior, where drivers move between service regions, using a space-time geometry to capture dynamic acceptance probabilities. We integrate this into a queueing framework (M/M/S) coupled with network percolation theory to demonstrate that tight ETAs not only reduce individual bundling efficiency, increase matching time, but also fragment the driver network, amplifying matching frictions.

ABSTRACT. Hub selection plays a key role in the design of less-than-truckload (LTL) freight transportation networks. Carriers must improve freight consolidation efficiency, reduce transportation costs, and remain resilient to changing customer demand. Although traditional integer programming models can provide optimal solutions, they are computationally expensive and not easily accessible to industry practitioners. In this project, we aim to develop machine-learning-assisted heuristics for hub selection. Using exact solutions from an integer programming model, we generate training datasets for machine-learning algorithms to learn patterns behind optimal solutions. These datasets encapsulate LTL network structure through key topological features, including hub popularity, customer flexibility, granularity, and proximity. The resulting heuristic rules are therefore based on a small set of features and remain effective as additional hubs and customers are introduced. We are also synthesizing other LTL network features and investigating how their interactions influence optimal network design decisions.

ABSTRACT. Communities often rely on mutual trust and cooperation to meet collective needs. This cooperation involves sharing and moving resources among members, but time or capacity constraints can make exchanges difficult to coordinate. To address this, we introduce a framework enabling fast, decentralized resource sharing and routing among trusted participants under strict time horizons. Extending the SWAP platform for collaborative exchange (Huang et al. 2023), the SWAP-and-Route mechanism lets communities declare needs and available resources and provides optimized exchange and routing plans. Once participation is sufficient, the framework identifies feasible, valuable exchanges while assigning and scheduling pickup and delivery tasks. By embedding vehicle routing collaboration, redundant trips are eliminated, and coordination efficiency is enhanced. We strengthen formulations through enhancement strategies that improve scalability and assess them using a factorial design. We demonstrate applicability through response coordination and exchange among food pantries.

ABSTRACT. High-frequency traffic time series are essential for congestion monitoring and operational decision making, yet are often incomplete due to sensor failures and communication issues. This paper proposes an interpretable and efficient framework for traffic time series imputation that explicitly captures underlying temporal dynamics while remaining robust to localized disruptions. The method leverages structured Hankel representations to model long-term temporal coherence and integrates robust temporal regularization to preserve abrupt events such as congestion onsets and recoveries. An efficient optimization algorithm is developed to solve the resulting formulation at scale. A representative case study on real traffic speed data demonstrates that the method accurately reconstructs long-range trends and major patterns under severe missingness.

ABSTRACT. This study presents a novel mathematical model that treats the airport declared capacity as an endogenous decision variable, aiming to simultaneously minimize operational delays (aircraft waiting time in the queue to utilize the runway), schedule delays (request displacements), and maximize declared capacity utilization. Furthermore, we proposed a multi-objective method, which integrates Column Generation (CG) with Simulated Annealing (SA) to handle large-scale instances of the problem. This study presents experimental results using a synthetic dataset (based on real schedule data) for São Paulo–Guarulhos International Airport in Brazil, analysing the trade-offs among objectives (such as declared capacity utilization versus operational delay) and conducting a sensitivity analysis regarding turnaround time flexibility. Furthermore, our analysis extends to the temporal distribution of delays. For each solution representative (with its associated declared capacity level), we have analysed the frequency and magnitude of total delays across different days and hours over the entire scheduling season.

ABSTRACT. The Transportation Science and Logistics (TSL) community will benefit from a unified, open data standard that spans dynamic traffic assignment, traffic signal optimization, and multimodal network modeling. Such a standard enables reproducible benchmarks, transparent model validation, and seamless data exchange among researchers and planning agencies.

The General Modeling Network Specification (GMNS) enables the sharing of routable road networks with varying resolutions. GMNS follows four core design principles: it defines data tables rather than software implementations; it is extensible with few required files (nodes, links) and optional extensions for dynamic networks; it represents infrastructure, services, and policies including physical roads, traffic controls, tolls, and time-of-day restrictions; and it uses simple tables for human and machine readability.

More than 40 curated GMNS networks have been assembled, with sizes ranging from a few dozen nodes to 20 million nodes, to enable benchmarking, and to demonstrate multi-scale optimization and multimodal connectivity.

ABSTRACT. In large-scale evacuations, a major challenge is managing charging delays for electric vehicles (EVs) as they converge on a vulnerable charging network. This paper presents a real-time, cooperative dispatching framework using mobile truck charging stations to support fixed infrastructure during evacuations. The task is modeled as a finite-horizon Markov Decision Process (MDP) that accounts for random EV arrivals, charging demand, and charger outages. The objective is to optimize truck dispatch decisions to maximize cumulative rewards. We propose a Multi-Agent Proximal Policy Optimization (MAPPO) algorithm, which operates under a centralized training, decentralized execution model. Simulations on a South Florida transportation network show that MAPPO outperforms rule-based and non-cooperative methods, reducing wait times and minimizing queued vehicles. These results highlight the potential of this learning-based, cooperative approach to improve charging network resilience during evacuations.

ABSTRACT. Although many well-known algorithms can solve each bipartite matching problem instance efficiently, it remains an open question how one could estimate the expected optimal matching distance for arbitrary numbers of randomly distributed vertices in $D$-dimensional spaces (referred to as a random bipartite matching problem, RBMP). This paper proposes a comprehensive modeling framework that yields closed-form approximate formulas for estimating the expected optimal matching cost across three interrelated but increasingly complex versions of RBMPs: (i) RBMP-I, where edge costs are independently and identically distributed (i.i.d.); (ii) RBMP-S, where edge costs represent distances between vertices uniformly distributed on the surface of a hyper-sphere in a $D$-dimensional Euclidean space; and (iii) RBMP-B, where the vertices are uniformly distributed in a hyper-ball within a $D$-dimensional L$^p$ metric space. A series of Monte-Carlo simulation experiments are conducted to verify the accuracy of the proposed formulas under varying parameter combinations.

ABSTRACT. Within the maritime transport system, ship safety is central to maritime safety. Under limited Port State Control (PSC) resources, comprehensive inspection of all arriving ships is rarely feasible, highlighting the importance of accurately identifying high-risk vessels within limited inspection resources. This study proposes a data-driven approach based on ensemble prediction to support more efficient and precise ship selection in PSC. In our ensemble model, unlike approaches that treat base prediction models independently, we relax this constraint and treat the base prediction models dependently, calculating the mixture joint computation. Building upon this foundation, we obtain the comparison relationship matrix. We further adopt the lexicographic optimization with the predicted comparison relationship between any two ships, generating executable inspection priority sequences and selection schemes. Experimental results conducted on the real data of the Hong Kong Port demonstrate that the proposed method enhances the decision effectiveness and precision of ship selection.

ABSTRACT. Digital labor platforms rely on a heterogeneous workforce whose participation responds dynamically to platform policies. We develop an agent-based model grounded in multi-method, multi-year empirical research to examine the interaction between worker learning and platform wage adjustment in a spatial gig market. Workers differ by employment type and behavioral strategy, capturing variation in economic dependence and risk orientation. Results show that when workers learn, hourly earnings increase as workers reject low-value offers, while completion rates decline. When the platform responds by adapting wages, completion recovers to its highest level, but at substantial cost, driving platform profits negative. At the worker level, higher earnings do not translate uniformly into greater participation. Some income targeters work fewer hours after reaching goals more quickly, while full-time wage maximizers expand participation. These nonlinear responses highlight the importance of accounting for worker heterogeneity and motivate future work on targeted incentives and collective worker learning

ABSTRACT. Airport slot allocation is a critical problem that directly related with operational efficiency, congestion, and environmental performance. This paper proposes a Stackelberg game-theoretic framework for airport slot allocation in which the leader airline in the airport commits to a slot choice first, and the remaining airlines optimally respond under capacity, congestion, and emissions-related constraints. A mixed-integer optimization problem is solved for the followers, while the leader anticipates these responses to maximize its own payoff. Numerical experiments based on an airport scenario demonstrate that the proposed approach reallocates traffic from peak to off-peak periods when environmental and congestion costs are internalized, favoring lower-emission airlines in peak slots. Results show that the effectiveness of the Stackelberg framework in balancing economic efficiency with environmental sustainability, offering a more realistic alternative simulated the airline competition to valuation-only slot allocation mechanisms.

ABSTRACT. This paper develops a hybrid hyper-heuristic approach for generating fast and near-optimal solutions to the variant of the VRP with a mixture of closed-loop and open-path routes, heterogeneous vehicle types, fixed costs and time-dependent travel and service times. This is applicable in emerging markets where last-mile distributors frequently rent a portion of their fleet from the spot market to meet daily uncertain demand (and these rented vehicles do not need to return to the depot), and face heavy, time of day-dependent traffic congestion. This problem setting called the COMVRP-TD encompasses the Open VRP, the VRPTW, the HVRP-FSM and the TDVRP as special cases. The algorithm developed here incorporates operators from the genetic algorithm and the large neighborhood search applied in a statistically learned sequence and produces solutions that are within 5% of exact MIP optimal solutions.

ABSTRACT. Community shuttle services are often planned separately for each residential community in a hub-and-spoke manner, which can cause redundant trips and low capacity utilization. We propose an integrated community shuttle service network design (CSSND) that serves on-demand feeder requests across communities. The system uses modular autonomous vehicles (MAVs) that can operate alone or physically couple into complete MAVs (CMAVs), allowing capacity to be shared and adapted to demand. We formulate CSSND as a two-stage stochastic integer program: the first stage selects the network configuration and dispatch schedule, and the second stage optimizes scenario-based routing itineraries using a path-based formulation. To solve the resulting model to optimality with integer decisions in both stages, we develop a CG-BBC method that embeds column generation within a branch-and-Benders-cut framework. Singapore case studies show CG-BBC outperforms Gurobi in solution quality under the same time limit.

ABSTRACT. We consider a planning problem for freight transportation carriers that seeks to profitably match supply with demand under uncertain shipment volumes. On the supply side, transportation network design decisions regarding hub locations and vehicle dispatches across the planning horizon are determined. On the demand side, the carrier’s ability to expand service coverage by selectively accepting additional customer demands beyond existing contracts is incorporated, including long-term and transactional customers. This problem, referred to as the Demand-driven Hub and Service Network Design under Uncertainty, is formulated as a two-stage stochastic program. An enhanced Benders Decomposition–based solution method is developed, inspired by Partial Benders Decomposition, embedding subsets of subproblem variables and constraints into the master problem and incorporating valid inequalities. An extensive computational study demonstrates that the proposed method outperforms benchmark adaptations. The benefits of solving the integrated model are validated using a Sample Average Approximation–based analysis.

ABSTRACT. This study addresses the integrated optimization of strategic hub location, tactical transportation planning, and operational routing for less-than-truckload freight carriers operating under demand uncertainty. We propose a novel multi-stage stochastic programming framework that captures the hierarchical nature of these decisions across an extended time horizon—ranging from long-term infrastructure investments to seasonal capacity planning and daily shipment flows. To tackle the computational complexity of this large-scale problem, we develop a Nested Partial Benders Decomposition (NPBD) algorithm. This method enhances standard nested decomposition by integrating partial Benders cuts to improve convergence and solution quality. The proposed model and solution methodology are validated using a realistic dataset from a multi-state U.S. freight carrier. Computational experiments demonstrate the efficiency of the NPBD algorithm and quantify the significant value of the stochastic solution (VSS), highlighting the benefits of explicitly incorporating uncertainty into multi-echelon network design decisions.

ABSTRACT. Battery electric buses (BEBs) are increasingly adopted by transit agencies, yet full electrification is typically preceded by a prolonged transition phase in which BEBs operate alongside diesel buses under limited charging infrastructure. This paper studies the operational scheduling and charging problem for mixed bus fleets under block-based transit operations. We develop a detailed mixed-integer linear programming formulation that explicitly integrates vehicle assignment, state-of-charge dynamics, and shared charger capacity constraints at high temporal resolution. To overcome the computational intractability of the monolithic formulation, We propose an accelerated logic-based Benders decomposition framework with novel structural energy lower bound cuts that proactively eliminate energy-infeasible assignments and provide tight lower-bound estimates without explicitly solving the charging subproblem. Computational experiments based on a real-world case study from the Alexandria Bus Company demonstrate that the proposed approach scales to realistic problem sizes and outperforms standard decomposition approaches and commercial solvers in terms of solution time.

ABSTRACT. Urban fixed-route transit systems struggle to adapt to changing passenger demand patterns and exhibit persistent gaps in accessibility, driving interest in flexible, demand-responsive microtransit. Motivated by a Chinese microtransit operator slated to scale nationally, this paper studies an operating model that uses high-capacity vehicles and requires passengers to walk to designated bus stations to enable better demand consolidation. However, the use of high-capacity vehicles and additional vehicle-customer coordination requirements induce a highly complex routing problem. In response, we formulate a subpath-based optimization model on a time–space network to minimize operating costs and maximize passenger service metrics in microtransit operations. For scalability, we propose an efficient label-setting algorithm with dominance rules for subpath enumeration. Using real-world data from microtransit operations in Zhuzhou, we demonstrate that microtransit with vehicle-customer coordination significantly outperforms fixed-route transit and ride-hailing in efficiency and passenger service, achieving higher coverage with modest walking requirements.

ABSTRACT. To reduce energy consumption and achieve eco-driving of emerging maglev train services, we study the maglev train speed-profile optimization (M-TSPO). Different from wheel–rail trains, maglev operation is stator-based: Fixed stators are divided into controllable intervals with stator-level speed limits and power-supply capabilities, which renders interval length an additional decision variable. In this paper, we develop a constrained nonconvex M-TSPO model based on train motion equations. By applying variable separation and piecewise approximations, we reformulate the M-TSPO model into a speed–time–location dependent model (STL-DM). We then propose a two-stage learning-based framework, in which a multi-gate mixture-of-experts network tunes interval entry speeds and a relaxed STL-DM optimizes interval lengths to construct the speed profile, guided by a customized loss function. Case studies on the Beijing maglev line achieve energy reductions of 14.26% and 8.91% over location-time and speed-time benchmarks. This yields 1.05 or 0.62 tons CO2/day emission reduction, respectively.

ABSTRACT. Rapid e-commerce growth is straining last-mile delivery, the costliest segment of the parcel logistics chain. We study service points (SPs) with automated lockers as a cost-saving alternative to home delivery. We propose a simple pre-dispatch policy: inform recipients of nearby SP availability and let them choose pickup location and time. A disutility model captures heterogeneous sensitivities to distance and delay. In a simulation calibrated with real data, the preference-based policy outperforms rule-based industry baselines, consistently lowering recipient disutility and balancing SP loads, especially when location preferences vary widely or congestion is moderate. The approach requires no estimation of individual utilities, remains effective under conservative availability forecasts, and improves service quality without additional physical infrastructure.

ABSTRACT. This paper studies a last-mile delivery problem in which customer requests are served through a network of parcel lockers rather than by traditional home delivery. While locker-based delivery enables route consolidation and increased flexibility for customers, it also introduces operational challenges due to the uncertain availability of locker space, which depends on stochastic customer pickup behavior and is only revealed upon vehicle arrival. We model this setting as a dynamic vehicle routing problem with customer-specific locker preferences and stochastic locker availability. To address the resulting decision-making under uncertainty, we propose a model-based lookahead policy that combines offline scenario-based estimation of locker availability with an online anticipatory optimization scheme. The proposed approach dynamically updates delivery decisions in real time while respecting operational constraints such as route duration and service quality. Computational experiments on realistic urban instances illustrate the effectiveness of the approach compared to a baseline policy.

ABSTRACT. We present a two-stage stochastic location-routing model for the placement of parcel lockers that explicitly accounts for behavioral uncertainty in the customer choice of locker delivery or home delivery. We model each customer’s choice of delivery location as an independent Bernoulli random variable. First-stage decisions determine locker locations, while second-stage decisions determine scenario-dependent routing plans to serve both locker and home-delivery customers. We tailor a Monte Carlo scenario sampling approach to account for the non-equiprobable scenarios and demonstrate that this approach is more accurate at estimating the expected cost than completely random Monte Carlo sampling. Ongoing work examines a sampling strategy that stratifies the scenario space by the number of customers requesting home delivery.

ABSTRACT. We introduce a new variant of the Undirected Team Orienteering Arc Routing Problem that incorporates three key features: required edges, capacitated vehicles, and multiple services. In this problem, demand is placed at some edges of a given undirected graph and served demand edges produce a profit. Feasible routes must start and end at a given depot and there is a time limit constraint on the maximum duration of each route and a capacity limit on the demand served by each vehicle. The problem asks for a given number of maximum profit routes while ensuring all required edges are served. We propose a new unified undirected formulation with binary variables. We also introduce a logic-based Benders decomposition and show how to strengthen the logic-based Benders cuts. Furthermore, we design several new families of valid inequalities. Extensive computational tests are conducted to examine the performance of the proposed formulations and valid inequalities.

ABSTRACT. The traveling salesman problem with drone (TSP-D) is a drone-assisted delivery route optimization problem. In this study, we provide a comprehensive understanding of the TSP-D polytope. First, we suggest four different families of valid inequalities, such as Truck subtour elimination inequalities (TSEIs), Drone subtour elimination inequalities (DSEIs), Drone external matching inequalities (DEMIs), and CG-strengthened Truck subtour elimination inequalities (CG-TSEIs). Second, among the suggested inequalities, we show that TSEIs and DSEIs are facet-defining inequalities for the TSP-D polytope. Third, we prove that the exact separation problems of TSEIs and DSEIs are in class P. Fourth, we implement a cutting-plane algorithm to compute LP bounds for the compact TSP-D formulation strengthened with the TSEIs, DSEIs, DEMIs, and CG-TSEIs. Our computational experiments demonstrate that the strengthened polyhedron by the suggested families of inequalities significantly outperforms the LP-relaxation of the TSP-D compact formulation, providing the much tighter LP bounds over all generated instances.

ABSTRACT. Full-truckload (FTL) transport, where goods occupy an entire trailer, is vital for Europe’s economy, yet accounting for roughly six percent of Europe’s greenhouse-gas emissions. Long combination vehicles (LCVs), which pull multiple trailers with a single truck, offer opportunities to consolidate shipments and reduce emissions, but their effective use requires complex planning due to heterogeneous vehicle configurations and configuration changes at equipment yards. We introduce the Full-truckload Pickup and Delivery Problem with Configurations (FTL-PDPC), which explicitly models vehicle configurations. We present compact and extended formulations and develop a tailored Large Neighborhood Search (LNS) heuristic. Computational experiments show that the LNS outperforms state-of-the-art heuristics, reducing the average optimality gap from 0.48% to 0.19%. A case study at a European FTL carrier with 444 feasible configurations shows that the LNS reduces emissions by 11.54% compared to current operations. Disallowing LCVs increases emissions by 34.65%, while equipment standardization yields an additional 11.71% reduction.

ABSTRACT. Network revenue management (NRM) maximizes expected revenue by allocating limited resources to sequentially arriving customer requests in passenger transportation networks. Bid price controls, which accept a request whenever its revenue exceeds the estimated opportunity cost, are widely used due to their simplicity: they can be computed offline from the optimal dual solution of the deterministic linear program (DLP), and during real-time operation the policy reduces to a simple threshold comparison. However, when the DLP is primal degenerate, multiple optimal dual solutions exist, leading to potentially suboptimal acceptance decisions. Yet its impact on NRM has not been systematically examined. This paper closes this gap, analyzing how degeneracy affects bid price controls and proposing two efficient control approaches that retain correct acceptance behavior. Our analysis shows that degeneracy occurs frequently across passenger transportation networs. In an extensive numerical study, we compare one of our approaches with the classic dynamic bid price control.

ABSTRACT. Technological advancements in telematics enable insurers to price risk contingent on monitoring signals of the driver's action, ostensibly mitigating moral hazard. However, canonical models of Usage-Based Insurance (UBI) assume fully attentive drivers, overlooking the cognitive loads inherent to driving. We bridge this gap by embedding rational inattention into a principal-agent framework with monitoring. We show that information frictions introduce a behavioral wedge, compelling the principal to design "high-powered" incentive schedules to overcome cognitive loads. Our analysis reveals: while high attention costs facilitate rent extraction by the insurer at the expense of the driver's surplus, improvements in monitoring precision generate Pareto improvements even in the presence of cognitive frictions. These findings underscore that the welfare gains from UBI depend critically on the interplay between signal quality and the agent's information processing costs.

ABSTRACT. Vehicle sharing providers traditionally offer spontaneous short-term rentals. However, providers started to additionally offer long-term rentals which customers can reserve in advance, including their specification of departure location and time, and therefore allowing them to reliably plan trips. For the provider, novel challenges for operational control arise: to ensure the availability of a suitable vehicle for each long-term rental reservation, the provider can either block an adequately located vehicle or relocate a vehicle. The former can cause lost revenue due to missed short-term rentals while the latter causes relocation cost. We formulate a sequential decision process to address a profit maximizing provider's problem of controlling vehicle availability for short-term rentals under the consideration of given long-term rental reservations. We develop a dynamic anticipative solution approach using a mixed-integer program which results in substantially higher profit compared to benchmarks motivated from industry.

ABSTRACT. Cities worldwide are exploring station electrification to further scale up e-bike-share systems, which currently rely on battery swapping that is labor- and carbon-intensive. In light of this vision, we develop an optimization model that jointly determines which stations to electrify and how to operate battery swapping at non-electrified stations. The model captures two key features: network effects and endogenous demand. Decisions at individual stations propagate through the system via bike flows, shaping bike and dock availability across locations and time and shaping user demand. The resulting formulation is a mixed-integer nonlinear program, with nonlinearity arising from spatially and temporally heterogeneous demand substitution between regular and electric bikes. To enable solutions at scale, we develop a machine-learning-based approach that yields a compact, high-quality mixed-integer linear approximation. Using data from New York City, we demonstrate the scalability of our approach and derive insights into when to prioritize station electrification versus battery swapping.

ABSTRACT. We present analytical and empirical arguments that competition over a nearly homogeneous service forces scooter firms to oversupply. Using deployment and ride data from scooter firm APIs, city transportation departments, and operational data from one major operator, we estimate average daily scooter utilization across nine major U.S. cities at 4.4%—almost half the 7.7% minimum utilization target set by policymakers. If consumers choose among available scooters roughly at random, equilibrium requires firms to inflate fleet sizes to preserve market shares. Using granular consumer-choice data from Kansas City, we estimate this relationship and provide evidence for random choice. Counterfactual simulations for Kansas City show: (i) a 228% oversupply of scooters; (ii) the interaction between consumer choice and competition explains 196 percentage points of this oversupply; and (iii) entry by one more firm would raise oversupply by an additional 87 points. Policymakers should be frugal with the number of firms they permit.

ABSTRACT. Regulators increasingly use minimum-service mandates to improve mobility in historically disadvantaged neighborhoods, yet their effectiveness remains untested. Shared micromobility—electric scooters and bikes—has become a prominent testing ground, with 36 programs across 34 United States cities adopting minimum deployment thresholds by 2024. We study eight programs across six cities between 2019 and 2023, jointly covering 42\% of the population in cities with such mandates. Because thresholds apply within administratively drawn boundaries that cut through disadvantaged neighborhoods, we can evaluate their effectiveness using a spatial regression discontinuity design. Analyzing 3.76 million status updates from 19,466 scooters deployed by an operator with over 30\% market share, we find precisely estimated null effects on availability, ridership, essential trips, and other outcomes despite full compliance. Thresholds are generally not binding, suggesting mandates must become more targeted and stricter to alter operator behavior.