ABSTRACT. This paper examines the two-stage fixed-charge transportation problem (TS-FCTP), in which a commodity is shipped from a set of capacitated source nodes to satisfy the demand at a set of destination nodes through some transshipment nodes. Every link between a source node and a transshipment node, and between a transshipment node and a destination node, incurs both fixed and variable costs, and the objective of the problem is to minimize the total costs. The fixed costs lead to a mixed-integer linear programming (MILP) formulation for TS-FCTP, which is known to be NP-hard. One approach to solve TS-FCTP more efficiently is by strengthening its lower bound using valid inequalities, also known as cutting planes. In this paper, we propose two classes of valid inequalities, referred to as PRQ Type1 and Type2, tailored for the TS-FCTP. Through computational experiments, we demonstrate the impact of our proposed valid inequalities in solving the TS-FCTP.

ABSTRACT. In this paper, we study a distribution network design problem that explicitly accounts for adversarial disruptions to intermediate facilities. We consider a two-echelon capacitated network connecting supply and demand nodes, where the designer locates facilities to minimize total costs under both normal and disrupted operations. The problem is formulated as a sequential Stackelberg game between a designer and an interdictor. Unlike deterministic interdiction models, we propose a stochastic mixed-integer programming formulation in which interdiction outcomes are uncertain and the interdictor seeks to maximize expected disruption. We develop an exact solution approach based on Benders decomposition, with stochastic subproblems solved efficiently via dual decomposition. Two families of valid inequalities are introduced to accelerate convergence. Computational experiments demonstrate the effectiveness of the proposed method. Results further show that the stochastic model yields less conservative network designs than deterministic counterparts, locating fewer facilities while incurring only marginal increases in post-disruption costs.

ABSTRACT. Traditional network interdiction problems assume that the interdictor is the only player that affects the network's topology or parameters. In reality, followers may adapt to interdictions by changing certain aspects of the network. This behavior is representative of applications in human trafficking, where traffickers aim to respond to interdictions to recover lost revenues. We present a multiperiod max - min cost flow network interdiction problem in which the node supplies have the ability to shift over time and the trafficker can recruit new victims. We model the follower's ability to change the supply in nodes to respond to the round's interdictions through a bilevel mixed integer linear optimization problem. We study how supply shifts to nearby victims, showing how interdiction impacts those that remain in the network. We utilize a column-and-constraint generation algorithm to highlight that globally optimal solutions capture interdiction plans that are not greedy.

ABSTRACT. We consider a system of shared electric vehicles that operates on an on-demand basis over a spatial network of multiple locations. We model this system as a closed queueing network and derive closed-form lower and upper bounds on the minimal fleet size and number of chargers at each location so that the fraction of demand unfulfilled and average waiting time for charging are below predetermined thresholds. More significantly, these dimensioning rules are highly interpretable with separable components that can be attributed to different drivers of capacity. We leverage the bounds to propose approximations of the optimal capacity that are asymptotically optimal and otherwise exhibits strong performance as validated by extensive numerical experiments. We show how the results we obtain can be used to examine the tradeoff between fleet size and charging infrastructure size and assess the impact of electrification.

ABSTRACT. The electrification of transportation represents a critical challenge in the global transition toward net-zero emissions, as the sector often accounts for more than one-quarter of national energy consumption. Achieving this transformation requires not only widespread adoption of electric vehicles (EVs) but also their seamless integration into interdependent infrastructure systems—specifically, the transportation-electricity nexus (TEN). This paper develops an optimal multi-modal transportation and electric power flow (OMTEPF) model to evaluate the benefits of coordinated, dynamic system operation. Building on hetero-functional graph theory, the framework enables joint optimization of five operational decisions in TEN management. The mesoscopic, dynamic model explicitly represents individual EVs and their state-of-charge trajectories, thereby extending beyond the prevailing literature’s focus on macroscopic traffic assignment. On the power system side, an IV-ACOPF formulation ensures globally optimal solutions to the electrical subproblems. Comparative analysis demonstrates the substantial value of coordinated TEN operation relative to the uncoordinated infrastructure management.

ABSTRACT. Public transport electrification is a key pillar of urban decarbonization, however the transition from diesel to electric bus fleets remains a complex long-term planning problem involving capital-intensive investments, operational constraints, and distributional concerns. This extended abstract presents a tri-objective, multi-period optimization framework for the strategic electrification of urban bus fleets. The model simultaneously minimizes total life-cycle costs, promotes a gradual and operationally feasible transition, and ensures an equitable spatial allocation of electric bus services across urban regions. Trade-offs between economic, environmental, and social objectives are explored using an ε-constraint algorithm that approximates the Pareto front. A real-world application to the nighttime bus network in Santiago, Chile, illustrates the relevance of explicitly accounting for equity and transition smoothness in electrification planning.

ABSTRACT. Within the context of a same-day stochastic pickup-and-delivery problem, we consider the combined use of in-store shoppers as crowdshippers and providing customers the option to specify a parcel locker for their delivery location. We develop a solution approach to determine the assignment of requests through a knapsack formulation that anticipates the appearance of future in-store crowdshippers. The parameters for this anticipatory policy are estimated using a machine learning method trained offline with data labeled by solving deterministic optimization problems corresponding to perfect information scenarios. We examine several problem settings in our computational experiments and observe potential synergy between in-store crowdshipping and parcel lockers.

ABSTRACT. Crowdsourced delivery platforms rely on independent couriers who retain autonomy, creating significant uncertainty in operational planning. Existing matching procedures typically rely on static or population-average acceptance models, which fail to capture heterogeneous and evolving courier preferences. We study a dynamic matching problem in which a platform sequentially assigns customer requests to couriers who may stochastically accept or reject offers based on private preferences. We propose a learning-based matching framework that estimates courier-specific acceptance probabilities online using logistic regression, and we integrate these estimates directly into the matching procedure. We develop initial synthetic data that incorporates population-level behavior to boost the performance of our learning approach. Computational experiments based on empirical courier and delivery data show that our approach substantially outperforms distance-based heuristics and average-courier benchmarks, rapidly converges to full-information performance, and remains robust across varying levels of courier heterogeneity, selectivity, market size, and time-varying preferences.

ABSTRACT. This paper investigates auction-based procurement mechanisms for emerging Freight as a Service (FaaS) systems, where shippers submit mode agnostic requests fulfilled by multimodal carriers. We propose a compact bidding language and formulate the winner determination problem as a network flow model aimed at minimizing total cost. To balance efficiency, incentive compatibility, and fairness, we design core selecting payment rules that minimize the deviation from Vickrey–Clarke–Groves (VCG) payments while encouraging truthful bidding. Carrier fairness is incorporated through a VCG based Gini coefficient, which quantifies the distribution of VCG payoffs among winners. Theoretically, we establish the NP hardness of the underlying problem and characterize the non emptiness of the core, demonstrating that VCG outcomes belong to the core only when transshipment is not allowed. Numerical experiments using real world intermodal data validate the proposed framework, compare payment rules in terms of incentives and fairness, and evaluate computational performance across different problem scales.

ABSTRACT. We study a harvest scheduling and routing problem motivated by banana production practices observed among small- to medium-sized producers in Turkey. An individual producer decides when to harvest bananas from independent growers and transports them to a ripening facility. Harvest timing decisions are critical, as banana yield evolves over time, and harvesting early or late results in food loss. We formulate this problem as a time-sensitive variant of the vehicle routing problem, where days are modeled as vehicles and revenue depends on harvest timing. The model jointly determines harvest schedules and routes while accounting for capacity constraints, multiple daily trips, and partial harvesting. Numerical experiments show that the proposed approach increases revenue by better aligning harvest times with crop maturity and significantly reduces transportation costs. In addition, the model yields substantial reductions in harvest-related food loss, highlighting its operational and sustainability benefits.

ABSTRACT. Urban logistics increasingly relies on heterogeneous fleets combining conventional and electric vehicles, which complicates large-scale vehicle routing due to vehicle-specific costs, energy consumption, and access restrictions. This paper studies large-scale heterogeneous vehicle routing problems (HVRPs) under four objective functions: travel distance, travel time, load- and speed-dependent energy consumption, and comprehensive operational costs. We propose a general-purpose solution framework that integrates dynamic decomposition, cluster-based routing, and a global set covering model, guided by an adaptive learning-based meta-model. The framework efficiently handles non-linear, vehicle-dependent cost structures and rich constraints such as zone restrictions. Computational experiments on large synthetic benchmarks and real-world urban instances demonstrate how different objectives significantly influence routing patterns, vehicle assignment, and the utilization of electric vehicles. The results provide insights into when electric vehicles are operationally advantageous and support informed fleet composition and investment decisions.

ABSTRACT. This study investigates fleet utilization inefficiencies in periodic vehicle routing problems (PVRP). While fixed delivery schedules simplify planning, they often induce significant daily demand variability throughout the horizon. Consequently, fleet sizing decisions are driven by peak daily demand rather than average demand, resulting in persistent underutilization of fleet capacity.

To address this, we propose a peak shaving heuristic that smooths realized demand over the planning horizon. Rather than shifting delivery days, the heuristic reallocates excess load to supplementary visits while strictly adhering to customer inventory constraints. Computational experiments demonstrate that this approach reduces daily load variance, decreasing both the required fleet size and total service costs. These findings highlight the potential of demand management strategies to reduce fleet requirements and overall operational costs in periodic routing problems.

ABSTRACT. Column generation (CG) in branch-and-price often suffers from dual oscillations that produce erratic reduced costs and many redundant columns. Deep dual-optimal inequalities (DDOIs) can stabilize CG by shrinking the dual region while retaining at least one optimal dual solution, but existing structure-derived DDOIs are problem specific and often too weak. We propose learning-based DDOIs (L-DDOIs) that automatically select effective pairwise inequalities of the form p_i - p_j <= 0. A consistency-aware probabilistic set-labeling procedure extracts mutually consistent inequalities from sampled optimal dual solutions, and a classifier ranks candidate pairs; a graph-based post-processing step enforces acyclicity/transitivity and a size budget. On challenging CVRP benchmarks solved by RouteOpt, L-DDOIs deliver a geometric-mean 5.1x reduction in CG time (up to 55x) with only 0.53% average deterioration in the root-node duality gap.

ABSTRACT. Large-scale dynamic resource allocation problems often involve agents governed by nonlinear and continuous system dynamics, rendering classical optimization approaches computationally intractable. Column generation (CG) is a powerful framework for such problems, but its applicability hinges on efficiently solving pricing subproblems, which become challenging when pricing takes the form of a nonlinear optimal control problem with objectives that vary across iterations. We propose Learning to Price, a novel integration of CG and meta reinforcement learning (Meta-RL). We cast pricing as a parametric optimal control problem indexed by the dual prices of the restricted master problem and agent characteristics, and train a single Meta-RL policy that generalizes across dual and agent contexts to generate near-optimal reduced-cost trajectories. Computational experiments on a large-scale vaccine allocation problem with nonlinear epidemic dynamics exemplify how the Learning to Price framework enables CG to scale to nonlinear dynamic systems beyond the reach of existing approaches.

ABSTRACT. Some airlines use the Preferential Bidding System to construct pilot schedules, where pilots submit bids for various activities and schedules that lexicographically maximize pilot scores according to seniority are selected. Traditionally, the problem is solved sequentially: first using the bids of the most senior pilot, then the next most senior, ensuring that earlier scores are not decreased. We propose an alternative approach based on column generation that simultaneously considers all objectives. Our method relies on two key contributions: first, we demonstrate that bounded linear lexicographic programs admit primal–dual feasible bases that can be computed efficiently; second, we extend standard techniques for resource-constrained longest path problems to lexicographic objectives, which is essential for generating new columns. Numerical experiments on Air France instances with up to 150 pilots show that our approach achieves good quality schedules within practical operational times, outperforming traditional sequential methods.

ABSTRACT. Although Autonomous vehicles (AVs) can improve urban mobility, they may induce additional vehicle travel and exacerbate congestion, particularly in dense urban areas. This induced demand is spatially heterogeneous because the benefit of AV's reduced value of time varies with trip distance. We develop a spatial equilibrium framework that jointly models congestion, residential location choice, and endogenous commuting mode and departure-time decisions. Commuting cost is determined by travel distance and aggregate traffic, while housing cost varies with location and residential density. A case study based on the Chicago metropolitan area shows that AVs attract a substantial share of commuting trips, nearly 20% of which shifted from public transit. At the citywide level, AVs reduce commuting and housing costs by 1.76% and 0.23%, respectively, but increase average congestion time and total vehicle miles traveled. We further evaluate multiple policy interventions, such as congestion pricing, vehicle-specific fees or subsidies, and transit fare adjustments.

ABSTRACT. Autonomous mobility-on-demand (AMoD) systems can mitigate congestion through centralized coordination. Their performance, however, depends on routing algorithms that act proactively under travel demand uncertainty. Recent work demonstrates that balancing AMoD traffic across alternative routes and staggering trip departures over time can significantly improve system performance. Yet, existing methods rely on full demand foresight and remain offline. We address this gap by proposing an online framework for balanced and staggered routing with sequentially arriving requests. We develop a combinatorial-optimization–augmented machine learning controller trained via imitation learning to approximate high-quality offline solutions. At runtime, decisions are obtained by solving a shortest path problem on a time-expanded graph with learned arc weights that anticipate emerging traffic conditions. Numerical experiments demonstrate that the proposed approach outperforms greedy baselines and recovers a substantial fraction of offline performance, thereby reducing total travel time and delay in online settings.

ABSTRACT. Autonomous trucking (AT) has the potential to unlock substantial reductions in transportation costs and faster long-haul service. However, safety and technological constraints limit current deployments to specific highway corridors, resulting in a hybrid setting where autonomous trucks handle portions of long-haul movements while regular trucks provide first- and last-mile service. Under emerging Driver-as-a-Service models, logistics companies lease autonomous trucks from AT providers and operate them within provider-specific hub networks, creating additional planning challenges. We develop an optimization framework that jointly determines order paths and continuous-time AT routes across disjoint autonomous networks. Using data from 33,000 U.S. full-truckload orders, our results show that ATs can serve a substantial share of freight—up to 45% of shipments—and generate cost savings of up to 8%. However, the reduction in human drivers is comparatively modest (between 6% and 21%), indicating that AT adoption is more likely to restructure rather than eliminate human labor.

ABSTRACT. Transitioning heavy-duty freight to low-carbon energy requires coordinated expansion of renewable electricity, green hydrogen, and zero-emission vehicle (ZEV) fleets. We propose a stochastic multi-agent optimization framework linking long-term investment with short-term operations for renewable investors, conventional generators, freight distributors, and mixed battery-electric (BEV) and fuel-cell (FCEV) fleets under endogenous electricity, hydrogen, and freight prices. A Texas case study maps corridor-level freight flows and resulting energy infrastructure concentration. Relative to a centralized planner, the decentralized equilibrium invests less in PV and hydrogen yet maintains comparable freight service via price-guided operations. Sensitivity analyses show hydrogen’s role expands when conventional capacity is tight and renewable costs are low, while abundant conventional capacity shifts the system toward BEV charging. Results motivate joint planning of freight, renewable, and hydrogen infrastructure and targeted incentives to reduce coordination gaps.

ABSTRACT. Hydrogen fuel cell trucks (HFCTs) have emerged as a technologically viable pathway for decarbonizing long‐haul freight transportation. The deployment of a strategic hydrogen refueling station (HRS) network is a prerequisite for large‐scale adoption, yet planning such infrastructure remains challenging due to uncertainty in future freight demand. This study develops a national‐scale stochastic optimization framework for identifying cost‐effective HRS locations under uncertain HFCT demand, supported by a hybrid Benders–singular value decomposition (BD–SVD) algorithm that enables computational efficiency on national scale. A reduced‐rank representation of the refueling-combination matrix is constructed using truncated SVD to significantly lower computational burden while preserving essential refueling characteristics. Numerical results on the simplified U.S. interstate system demonstrate major improvements in convergence speed compared with commercial solver and provide insights into early‐stage hydrogen infrastructure planning.

ABSTRACT. Freight transport holds a substantial share of industry emissions. Falling within the scope 3 emissions category, it is simultaneously one of the hardest sources to account for and one of the most impactful to decarbonise. To help businesses minimise the emissions of their freight transport networks, we leverage linear programming, network science and supervised machine learning to construct and mobilise an Emissions Efficiency Index (EEI) - a metric to assess and benchmark emissions efficiency relative to the minimum emissions achievable within existing transport infrastructure and without compromising site-level net balance. Through an extensive numerical analysis with Estée Lauder Companies, we offer preliminary evidence of the feasibility and practical application of the EEI as a reliable metric of freight transport emissions efficiency, with potential as a useful industry benchmark to guide corporate interventions towards net zero.

ABSTRACT. This paper addresses the computation of tight optimistic bounds for multi-stage stochastic optimization problems using information relaxation duality. We introduce a parametric class of penalty functions that are bilinear in decisions and transformed innovations of the underlying stochastic process. Our approach provides a generic computational framework for deriving such bounds without requiring explicit knowledge or approximation of the problem's value functions. We formulate a minimax problem to determine optimal penalty parameters within this class, yielding the tightest bound achievable with these penalties. Under regularity conditions, this minimax problem admits a dual reformulation as a stochastic program with expectation constraints. We propose a proximal bundle method to solve the minimax problem directly through iterative optimization. Computational experiments on six canonical operations research problems—including parallel machine scheduling, multi-item lot sizing, and stochastic knapsack—demonstrate gap reductions up to 40\% compared to existing information relaxation methods that require problem-specific value function approximations.

ABSTRACT. Performance evaluation frameworks for online stochastic dynamic programs, both in the online optimization and machine learning literature, are based on offline regret or omniscient bounds (also known as perfect-information-relaxations). We present novel tractable methods to tighten these bounds, allowing for superior evaluation of policies from any method. Our approach builds on the idea of penalized information-relaxation bounds (Brown, Smith, and Sun 2010), which penalize the additional information available to an anticipative decision-maker. Their approach relies on the enumeration of all system states, rendering it intractable at scale. We present novel approaches for the design, modeling and tractable computation of tighter regret bounds by capturing penalty functions implicitly rather than explicitly, avoiding the need for state enumeration, making our approach applicable for systems with uncountable state spaces. We demonstrate the success of our methodology on real-world instances from two broad problem classes modeled as time-space networks and block-diagonal mixed-integer programs.

ABSTRACT. In this paper, we introduce the Information Collection Problem where an agent must traverse a network to maximize information gathered before mission termination by hostile entities. Information is emitted stochastically over time and allows for multiple visits to the same location to maximize information gain. The agent's speed along each edge impacts both the rate of information collection and the risk of detection, motivating explicit speed control within the model. We develop a MILP formulation that approximates this nonlinear problem while preserving key features. To improve the MILP solutions, we propose two speed-enhancement heuristics that use nonlinear optimization techniques. Our methods produce high-quality solutions for large-scale instances within reasonable computation times. We also demonstrate how the ICP framework (i) generalizes several classical routing and other combinatorial problems, and (ii) can be adapted to a range of civilian and military information-gathering scenarios, highlighting its broad practical relevance.

ABSTRACT. Online grocery fulfillment requires retailers to make real-time order acceptance decisions under limited picking capacity. A key input to these decisions is an accurate estimate of the picking time and associated costs of incoming customer orders. While detailed routing- and layout-based heuristics can provide realistic picking-time estimates, they are computationally infeasible for use during checkout. We study data-driven picking-time estimation in a brick-and-mortar click-and-collect setting using real-world order data and a detailed supermarket layout. We construct a benchmark model that captures routing, item retrieval, and handling effort and translate picking times into fulfillment costs. Based on basket composition and layout-derived features, we develop predictive models that provide accurate picking-time estimates for many customer orders. At the same time, we identify basket characteristics for which prediction accuracy deteriorates. The results highlight both the potential and limitations of data-driven picking-time estimation for capacity-feasible and cost-aware order acceptance in online grocery retail.

ABSTRACT. Designing resource allocation policies that perform well across a wide range of demand conditions remains a central challenge in logistics. Fully dynamic approaches can adapt decisions to each realized scenario but are often too complex to implement at scale. Static allocations are easy to deploy yet may perform poorly when demand deviates from expectations. We study k-adaptable policies as an intermediate solution: the system pre-computes k allocation designs offline and then selects which one to operate after demand is realized. We characterize when an individual design is effective and how multiple k-designs can be generated to complement one another, yielding a tractable optimization procedure that performs well across different instance sizes and compositions. Experiments in parcel sorting and zone picking show that small-k designs outperform static allocations while capturing most benefits of fully dynamic policies, with the magnitude of these gains varying significantly with system structure, scale, and load conditions.

ABSTRACT. A hybrid robotic bin-picking system (HRBPS) featuring heterogeneous robots has been widely adopted in modern fulfillment centers. In an HRBPS, retrieval robots operate within storage aisles, retrieving or storing bins between storage layers and buffer tiers located at the bottom of racks, while transport robots move bins between buffer tiers and workstations to fulfill orders. To estimate performance and different operational policies in an HRBPS, we compare two scheduling policies (sequential vs. parallel) and three class-based storage policies (vertical, horizontal, and diagonal). We model the system as semi-open queuing networks and solve them using the approximate mean value analysis and matrix-geometric method. Simulation experiments validate the accuracy of the analytical models. Numerical experiments assess policy performance in terms of throughput, average order cycle time, expected queue length of external orders, and robot utilization rates.

ABSTRACT. We formulate a vacant ride-sourcing or taxi driver's routing decision as an infinite-horizon semi-Markov decision process (SMDP) in a road network, where a driver decides which link to take at each node and transitions to a new node depending on the stochastic vehicle-passenger matching process. A driver's decision is based on observable and unobservable states. The modeler's job is to jointly estimate an average driver's parameterized utility function as well as the state transition function based on the sequence of observed states and actions. We conduct theoretical analyses to establish the existence and uniqueness of a fixed point solution to the Bellman equation for the SMDP. The expected fare, expected operating cost and number of intersections in the urban area are found to be significant predictors of routing decisions. Comparison with several base models suggest the advantage of considering multiple decision cycles and joint estimation of routing and matching parameters.

ABSTRACT. We introduce the Generalized Route Planning and Scheduling Problem with Time-dependent Travel Times on Road Networks, where each request offers pickup and dropoff clusters of candidate stops, each with its own time window. As a motivating application, we consider a generalized dial-a-ride problem on a real road network with time-varying speeds. We develop a three-phase pipeline. In preprocessing, we precompute closed-form piecewise-linear arrival/departure profiles between candidate stops by solving the time-dependent quickest-path problem, constructing the needed lower envelopes, and mapping each stop-to-stop subnetwork to a single logical arc. In optimization, we propose DP-ALNS-SA, an adaptive large neighborhood search with simulated annealing that embeds exact dynamic-programming route schedulers to optimally schedule generalized routes. We accelerate move evaluation via fast feasibility screening and exact cost updates without rescheduling from scratch. The resulting framework enables end-to-end runs on instances with up to 3,000 requests, far beyond the reach of commercial solvers.

ABSTRACT. The pickup and delivery problem involves the routing of vehicles to transport goods between pickup and delivery locations. Our work addresses two common difficulties: large numbers of vehicles and limited information regarding future demand. First, we formalize a new variant of the online pickup and delivery with no demand backlog and a distance minimization objective. We then show a network flow equivalency of the offline problem, under a constraint stating that all vehicles’ fixed starting and ending locations cover all nodes. We embed the network flow into polynomial-time online algorithms for the online pickup and delivery problem. We show that the online algorithms are asymptotically 2-competitive, whereas a myopic integer programming baseline is not c-competitive for any c>0. Computational experiments show that the online algorithm returns solutions within 0.5 percent of the integer programming benchmark at less than a thousandth of the runtime.

ABSTRACT. We study the operation of fast-charging stations over a short, finite horizon with surging demand, when many electric vehicle (EV) drivers anticipate elevated travel risk or future resource scarcity. In these events, drivers engage in precautionary charging, intentionally charging more energy than is strictly required. However, fast charging follows a nonlinear constant-current/constant-voltage (CC–CV) profile: once the state of charge exceeds approximately 80\%, charging power tapers sharply. This tapering substantially increases vehicle occupation time and reduces effective station capacity (i.e., vehicles served per hour). We propose to manage the above panic charging issues by developing the first single-station charging queueing model that jointly captures: (i) CC–CV charging dynamics at DC fast chargers; (ii) endogenous, congestion-driven precautionary demand over a finite surge horizon; and (iii) optimal control policies and managerial insights that translate analytical results into practically relevant levers for mitigating panic charging and restoring charging capacity when it is most needed.

ABSTRACT. Uncoordinated charging of EVs poses a significant threat to power infrastructure. Our central question is: what is the maximum potential of a holistically coordinated charging system at scale to mitigate feeder overloading risks?

We propose a mobility-aware charging coordination framework, which requires maintaining proper energy levels throughout the time horizon such that the EV has enough mileage to complete each trip. We developed a custom Alternative Direction Method of Multipliers (ADMM) approach for efficient distributed optimization of the extremely large problem on parallel clusters.

Using realistic EV trajectory data and feeder-level hosting capacity, we optimize the one-week charging schedule for around 2 million EVs across 1300+ feeders in full coordination. The fully coordinated scheme can almost eliminate distribution feeder upgrade needs thus saving billions of upgrade costs. Further, we identify suboptimally determined session demand targets ignorant of future mobility demand as a key barrier from achieving this full potential.

ABSTRACT. As the electric vehicle (EV) market expands, Charging Service Providers (CSPs) must decide whether to offer Alternating Current (AC) charging, which is slower and less costly, and/or Direct Current (DC) charging, which is faster but more expensive. Consumers differ in their valuation of charging speed, while AC and DC systems face distinct electricity tariffs and efficiency levels, resulting in different operating costs. Governments increasingly provide power-tariff subsidies to lower these costs and promote charging infrastructure. This study develops a Stackelberg framework in which the government acts as the leader maximizing social welfare, and the CSP acts as the follower maximizing profit under pricing and service constraints. We analytically characterize equilibrium outcomes for optimal pricing and service configurations under varying subsidy levels. The analysis identifies four market regimes: joint AC–DC offerings, exclusive AC or DC markets, and market exit. We extend our study to analyse the impact of competition.

ABSTRACT. Ride-hailing platforms face two important challenges: (i) there are significant spatial demand imbalances that require some repositioning (empty routing) of drivers; (ii) the control of driver supply is partially decentralized in that drivers strategically decide whether to join the network, and if so, whether and where to reposition when not serving riders. For such ride-hailing networks, we study the following question: Under decentralized repositioning, how effective are driver wage policies in achieving the optimal centralized performance benchmark?

We consider a stationary fluid model of a ride-hailing network in a game-theoretic framework with riders, drivers, and the platform. We characterize the steady-state system equilibrium under decentralized repositioning for various driver wage policies and show how the effectiveness of driver wage policies depends on the interplay of demand imbalances, wage flexibility, and the congestion-sensitivity and spatial relations of travel times.

ABSTRACT. We study pricing and matching design in ride-hailing platforms that integrate autonomous vehicles (AVs) alongside human drivers. Motivated by industry practice, we model AVs as a complementary capacity expansion rather than a perfect substitute for human drivers. Rider requests differ in service desirability, and supply exhibits heterogeneous willingness to serve, inducing imperfect substitutability between AVs and human drivers. We propose a certification-based equilibrium framework that assigns tiered scores to riders and vehicles, enabling assortative matching and tier-based pricing. Under mild conditions on market characteristics, we show that the platform’s profit maximization admits a tractable two-stage formulation and that certification-based matching is optimal. Our main result establishes the existence of a hybrid equilibrium in which introducing a positive AV fleet strictly improves rider welfare, driver welfare, and platform profit relative to a human-only benchmark. Numerical experiments further illustrate the win–win–win effects and the role of AVs as complementary capacity.

ABSTRACT. This paper studies centralized MoD service design problems that jointly optimize strategic (facility location, fleet sizing) and tactical (trip pricing, fleet allocation) decisions. We formulate a canonical model as a mixed-integer nonconvex optimization program. The model explicitly incorporates demand-supply interactions in a network setting, where customer demand is governed by a nonconvex demand function and service supply is modeled by a closed queueing network. To solve this problem, we introduce a spatial branch-and-cut methodology which involves (i) reformulation of revenue maximization into a convex problem, (ii) generalized Benders decomposition, (iii) McCormick envelopes, and (iv) a spatial branch-and-bound framework. We also analytically deduce variable bounds by a dual decompsition method, which enables significantly accelerated convergence of our algorithm. We apply the methodology to three centralized MoD cases: UAM, UAM location (with faciltiy location), and car-sharing. Extensive experiments demonstrate generalizability of our model and consistent scability of our algorithm over state-of-the-art benchmarks.

ABSTRACT. We consider shipping operations in middle-mile logistics. Middle-miles refers to the transportation of shipments within the logistic service provider's network of distribution centers and requires handling complex decisions regarding the truck dispatching trucks and order routing. To handle shipments, providers implement regular transportation connections in the network. However, demand appears dynamically, potentially causing bottlenecks in the available transportation plan. To overcome this, providers need to route orders in an anticipatory manner and book additional connections at the spot market in case insufficient capacities cannot be avoided. We model the problem as a sequential decision process focusing on the operational control of dynamic shipments, given a tactical service network design solution, and under uncertain demand and spot market rates. We develop a cost function approximation to nudge a shortest-path inspired policy towards travel-time minimal routes favoring consolidation possibilities. Computational experiments on real-world data show superior performance compared to a rolling-horizon approach.

ABSTRACT. The trucking industry faces mounting pressure from driver shortages and increasingly strict sustainability requirements. These challenges are particularly pronounced in dynamic full truckload operations, where real-time assignment decisions affect both near-term performance and future system capacity. This paper studies a dynamic driver–order assignment problem with heterogeneous fleets, detailed regulatory constraints, and sequentially arriving transportation requests. We propose a dynamic, multi-objective decision framework that balances operational efficiency, CO2 emissions, and driver satisfaction by controlling workload deviations from preferred working hours. Assignment decisions are generated using a constraint programming model embedded within a rolling-horizon framework to ensure real-time feasibility. To capture the long-term consequences of myopic decisions, driver workload balance and time-window slack are incorporated as cost-function approximations within a Markov decision process. The parameters are calibrated using Bayesian optimization. Computational experiments show that the proposed driver-aware policies outperform current practice, reducing emissions, improving workload balance, and increasing robustness under demand uncertainty.

ABSTRACT. Retail companies such as Target offer faster and faster delivery options. Customers order goods and expect them to be delivered the same day in a selected time window. This leads to a change in operations: Instead of two clearly separated phases, a capture phase where orders are collected and a subsequent delivery phase, both phases now merge into one. Consequently, delivery operations take place while new orders are placed and vehicles are dispatched dynamically. Routes must be constructed such that they allow for efficient fulfillment of existing orders while also keeping the fleet flexible to deliver potentially incoming future orders on time. We propose a cost function approximation with a state-dependent constraint limiting the route duration in the decision space. This constraint is learned via reinforcement learning and allows for a thorough search of the decision space via mixed-integer programming while keeping the fleet flexible for future operations.

ABSTRACT. Scheduled and on-demand jobs are often managed separately in transportation and logistics systems. Motivated by the advancement of the gig economy and vehicular technologies, we study a dual-delivery system where the two types of jobs can be coordinated and performed using the same vehicle. In this system, the driver decides when and how on-demand jobs should be added to the delivery of scheduled jobs to maximize the expected revenue rate. We develop a zoning-based coordination policy that captures the synergies between these two types of jobs and characterizes its revenue performance. We analyze the structure of the optimal zoning policy and derive an analytical upper bound on the maximum revenue rate achievable under any coordination policy, shedding light on the limit of coordination. Using this bound and a benchmark deep reinforcement learning policy, we show that the simple switching policy is effective at reaping the benefits of coordination.

ABSTRACT. Crowdsourced delivery leverages the unused transport capacity of vehicles already on the road to perform parcel deliveries. While it can potentially tackle challenges in urban delivery, it also introduces unique planning challenges, notably the operator’s lack of direct control over driver availability and acceptance behavior. Compensation, now, also is a decision of the operator, which affects drivers’ behavior. In addition, the operator can offer bundles of tasks, which further shape drivers’ responses and increase the problem’s complexity. Therefore, we study a setting in which tasks and drivers arrive dynamically and stochastically, explicitly model driver- and offer-dependent acceptance probabilities, and jointly determine bundles and compensation. We formulate the problem as a Markov Decision Process and solve it using Value Function Approximation within an Approximate Dynamic Programming framework. Preliminary results show clear advantages over benchmark policies.

ABSTRACT. Crowdshipping platforms rely on independent gig workers whose participation depends critically on fair compensation. We develop an equity-oriented framework for last-mile delivery that balances operational costs with workload equity among heterogeneous couriers. Our bi-objective optimization model minimizes routing costs while equalizing adjusted profits, a novel metric capturing both capacity utilization and earnings. We propose a custom heuristic combining NSGA-II with multi-directional local search to approximate Pareto-optimal solutions. We demonstrate that modest cost increases (2.5%) yield substantial equity improvements (up to 65%). We show that minimizing workforce size benefits all stakeholders and quantify opportunity costs for high-performing workers. The framework remains effective under vehicle shortages, providing actionable insights for sustainable gig-economy platforms.

ABSTRACT. On-demand meal delivery platforms can shape demand through the restaurant rankings shown in their apps, but ranking changes also affect customer choice and, in turn, courier availability and congestion from the perspective of the order-to-capacity ratio. We study the joint optimization of dynamic rankings and courier routing under endogenous customer choice, contractual revenue protections for restaurants, and fairness in workload allocation. We develop a multi-stage optimization framework and a learning-to-optimize algorithm that uses offline-trained machine learning models to predict whether a candidate solution will improve the objective and by how much, enabling dozens of parallel evaluations without commercial solvers and supporting real-time decision-making at scale. Computational results show that, relative to fixed rankings, flexible rankings increase fulfilled orders by at least 9.9%. Incorporating congestion effects yields considerably higher delivery efficiency and a more balanced spatial distribution of congestion.

ABSTRACT. Determining marginal fulfillment costs (MFC) is crucial for effective decision-making in online grocery retail, a sector struggling with small profit margins and arduous service requirements. Paramount to improving operational efficiency, e-grocers need accurate real-time MFC estimations to optimize their service offers and prices for online customers. Traditional methods for estimating MFC are either too slow for online decision-making or inaccurate. This paper introduces a novel machine learning (ML) approach that provides fast and accurate MFC estimations with the help of carefully engineered features that can capture complex routing dynamics. Experiments with real-world data demonstrate the superiority of the proposed approach over state-of-the-art MFC estimation methods. Our analysis of more than 2000 predictors, from which 20 are curated, reveals critical insights into the use of network-level, neighborhood-based, and node-level features in capturing complex VRP dynamics.

ABSTRACT. We study methods to approximate road network distances on dense urban maps considering downstream optimization tasks such as route planning. In many applications, distances are obtained from a routing engine on demand, creating millions to billions of origin–destination queries and making distance retrieval a major computational bottleneck. We treat road distances as values of an arbitrary, possibly asymmetric oracle and build surrogate models that approximate these distances while preserving routing quality. We train neural networks by direct regression and decision-focused learning methods using grid-representation based network-level features. We consider the asymmetric traveling salesperson problem in a real last-mile service area in Montréal, using the Open Source Routing Machine as oracle. Our learned surrogates are more than 23 times faster than direct queries to the routing engine and also improve the solution quality by more than 92% compared to Euclidean baseline.

ABSTRACT. We develop a Markovian traffic equilibrium model for ride-hailing, where ride-hailing vehicles make sequential link choices in a network, whether empty or hired, to maximize random utilities in an infinite-horizon, discounted-return semi-Markovian process. Competitions among empty vehicles for passengers and traffic congestion due to road usage are both modeled on links. We define the equilibrium as the solution to a fixed point problem, and prove the existence of the fixed point solution. We conduct computational experiments in networks with varying sizes to explore the convergence of the fixed point iteration and the method of successive averages, and derive policy and managerial insights. We contribute to the state of the art by developing a new equilibrium model that accounts for both traffic congestion and vehicle competition for passengers in a sequential decision making framework to account for drivers' forward-looking capabilities.

ABSTRACT. We develop a generalized traffic equilibrium model that considers ride-hailing services provided by Transportation Network Companies (TNCs, e.g., Uber) and accounts for customer waiting. The generalized equilibrium model integrates three interacting sub-models, including TNCs' decision on dispatching ride-hailing vehicles, travelers' choice of which trip mode to use, and queueing dynamics that explicitly capture customers' waiting costs. The TNCs’ choices and customers’ choices together form a generalized noncooperative game, coupled with the queueing system of customer waiting. We provide the general conditions under which the customers' waiting cost mapping needs to satisfy to exist an equilibrium solution, and formulate the waiting cost function as a complementarity problem for computational tractability. Numerical experiments using the Sioux-Fall network show that compared with our method, existing methods tend to underestimate the waiting cost of ride-hailing customers. Thus, they overestimate the mode share of ride-hailing travelers and the vehicle miles traveled in the system.

ABSTRACT. Cruising is the search for parking when a driver finds no vacancy near the desired parking location, frequently identified along downtown curbside parking. Despite the uncertainty and frustration associated with this unpleasant experience, cruising for parking is widely preferred over off-site parking options due to its huge price advantage. A significant volume of cruising vehicles not only congests urban roads, but also creates adverse social outcomes such as air pollution, fuel waste, and increased chances of accidents. The ability to quantify these outcomes is largely based on an accurate characterization of the cruising traffic and/or the expected cruising delay. This work provides a Markov queueing perspective on the modeling and estimation of expected cruising delay, enabled through capturing the stochastic dynamics between parking availability and cruising traffic. With queueing reformulation, computational approximation, and performance analysis, we provide results that support a lightweight yet accurate computational approach for cruising delay estimation.

ABSTRACT. We study the design of high-power charging networks for heavy-duty electric vehicles serving long-haul highway freight. We model freight corridor as a time-expanded network and formulate a principal–agent bilevel program in which a charging network operator chooses hub locations, charger capacities, and time-of-day prices while truck fleets route and charge to maximize their own profit. To make the design problem tractable, we use a value-function reformulation and develop two complementary solution approaches: an augmented Lagrangian method and a neural-network surrogate with heuristic-guided search. A stylized case study based on the proposed I–10 Los Angeles–El Paso pilot shows how time-varying tariffs shape optimal charger deployment, congestion, and waiting times. We also compare designs from the two methods and find small revenue gaps but differences in capacity and delay. Our results highlight trade-offs between revenue, service quality, and investment and provide a template for planning corridor-wide HDEV charging under realistic pricing schemes.

ABSTRACT. We consider the Electric Vehicle Routing Problem, where customers are served by an electrified fleet that recharges at charging stations. The set-partitioning formulation for the EVRP has infinitely many path-based variables due to continuous charging decisions. We develop a column generation algorithm, where the pricing problem is decomposed into two phases: (i) generating subpaths between two consecutive charging stations, and (ii) combining subpaths into paths. We formalize subpath-based domination properties to establish the finite convergence and exactness of the column generation algorithm. We design domination criteria and prove that the resulting two-phase column generation algorithm can solve several EVRP variants, including time windows, nonlinear charging functions and relaxation tightening strategies (e.g., ng-routes, subset-row inequalities). Real-world experiments show that the two-phase algorithm outperforms path-based label-setting benchmarks and can scale to large problem instances.

ABSTRACT. This study considers a set of routes used by public transportation vehicles and dedicated distribution fleets in a general network. We aim to optimally locate alternative fuel refueling stations in the network to serve these dedicated routes. Deviations from prescribed routes for refueling purposes are allowed. Unlike most related literature, our approach considers all points in the network as candidate refueling station locations. We derive coverage constraints for any candidate location to serve a given route. Then we develop an exact algorithm to establish a finite dominating set (FDS) of candidate locations guaranteed to include an optimal solution to the problem. This set can be used in a mathematical model to minimize the number of stations required to cover all flows in the network. Numerical experiments on realistic networks are presented to illustrate the proposed methodology and to demonstrate its scalability and sensitivity to changes in parameter values.

ABSTRACT. Metro train operations are becoming more challenging due to overcrowding and unpredictable irregular passenger demand. To avoid passenger dissatisfaction, metro operators employ various operational strategies to meet the passenger demand adequately. This paper integrates metro timetabling with three operational strategies to improve passenger services. We propose three optimization models for timetable planning during both peak and off-peak hours. These models integrate operational strategies such as rolling-stock assignment, short-turning, and skip-stop to increase the number of services with limited trains on a bidirectional metro line. We perform experiments on these three mathematical models considering both time invariant and time dependent cases of demand. The paper also provides detailed calculations for train services, running times, and station dwell times. The proposed models are then implemented on a simplified Santiago metro line 1.

ABSTRACT. In major urban networks, rail systems represent the most efficient mode for high-capacity travel, and are typically fortified by feeder services. We consider a multimodal transport network consisting of a single bidirectional metro line reinforced by feeder buses serving its stations. We extend the commonly used FIFO passenger boarding rule and propose a new boarding policy tailored to general multimodal settings. Based on this policy, we formulate a demand-driven timetabling problem that optimizes train and feeder dispatching time, and trains capacities. We propose three families of valid inequalities for the problem, and develop an exact cut-generation algorithm enhanced with lower-bounding and warm-start techniques. We conduct extensive computational experiments and demonstrate the effectiveness of our solution methodologies, and show that our model and algorithm outperform benchmark metro-only dynamic programming methods. We further compare the optimized timetables with fixed-frequency timetables using maximum train formations, and show significant reductions in overall system cost.

ABSTRACT. This paper addresses tactical resource planning for railway operations in sea-rail container ports by formulating a new joint optimization problem that integrates train-to-track assignment and crane capacity allocation decisions. We first develop a mixed-integer programming model based on discretized time and space with the objective of minimizing total train delay, and then reformulate it as a set covering model using Dantzig-Wolfe decomposition. Building on this, we propose a tailored branch-and-price algorithm with a hybrid pricing approach, an advanced branching strategy, and several acceleration techniques. Extensive computational experiments demonstrate that our algorithm significantly outperforms a commercial optimization solver, achieving optimal solutions or optimality gaps below 2% for large instances, compared to gaps of up to 19% for the commercial solver. Furthermore, the computational results also show that our integrated planning approach reduces total train delay by up to 26% compared to sequential and rule-based heuristics, highlighting its practical effectiveness.

ABSTRACT. We propose a new surrogate model based optimization framework for network design and management problems, which makes use of a neural architecture that is atypical of the standard applications in literature. We justify our choice of architecture by showing it possesses desirable approximation properties on discrete domains. We also analyze the optimality bounds that one can generate using our model depending on the performance of the embedded neural network. We show how we can obtain easy convex formulations of our neural model to make embedding the surrogate into optimization models easier. We design a complicated crowdshipping network design problem to test our proposed framework on using real world data; and showcase the advantage our model has over conventional ReLU-based surrogate models in terms of computational efficiency, optimization performance and data efficiency.

ABSTRACT. The rapid growth of e-commerce has intensified pressure on urban logistics systems, motivating scalable and resilient network design methods. Within the Physical Internet paradigm, access hubs serve as first-tier transshipment nodes enabling efficient first- and last-mile operations. However, dense metropolitan environments generate numerous feasible hub locations, rendering direct location–allocation optimization computationally intractable. This paper proposes a scalable set-cover–based optimization framework for pre-designing access hub networks that explicitly allows joint coverage of multiple demand zones. The model captures dominant cost and capacity drivers through modular hub sizing and continuous approximation of courier routing costs, while ensuring resiliency via minimum multi-hub coverage requirements. A large-scale case study of Metropolitan Atlanta, involving over 166,000 candidate sites, demonstrates that the proposed approach reduces the candidate set by nearly two orders of magnitude while remaining computationally efficient, providing compact, high-quality inputs for downstream urban logistics optimization models.

ABSTRACT. We address fulfillment center location planning with complex assignment constraints arising from contractual, regulatory, and capacity limitations. We cast the problem as constrained k-means clustering: demand points are partitioned into k clusters to minimize total distance while enforcing bounds on points per cluster, capacity limits on aggregated volumes, and must-link/cannot-link rules. Because pure integer programming fails at scale and prior BLP-kmeans variants handle only one constraint family, we extend BLP-kmeans to treat all constraints simultaneously. The approach couples exact assignment subproblems with k-means iterations, adds multi-constraint model-size reductions, and exploits constraint structure by shrinking must-link components and unifying pairwise constraints. We build benchmarks by augmenting capacitated datasets with pairwise constraints and evaluate on instances up to 10,000+ points and ~100 clusters. Results show up to 50% runtime reductions without loss in solution quality or feasibility.

ABSTRACT. Truck platooning, enabled by connected and automated vehicle technologies, offers significant fuel and labor savings. In fragmented trucking industries, cross-fleet platoons are often needed but are unstable due to uneven platoon benefit distribution. Existing methods, benefit redistribution mechanisms, focus on individual trucks and fail to balance behavioral stability with system efficiency. We propose a benefit redistribution mechanism within a single-leader, multiple-follower Stackelberg–generalized Nash game framework. The government, as leader, allocates benefits based on truck roles and schedules to maximize system utility, while fleet operators, as followers, adjust trucks’ schedules and roles to maximize their utility. We prove the existence of a Stackelberg–generalized Nash equilibrium, ensuring behavioral stability. The bi-level model combines continuous upper-level and discrete lower-level variables in a structure not solvable by existing exact algorithm. We develop an exact algorithm using relaxation techniques and valid inequalities demonstrate through numerical experiments that our mechanism efficiently improves system utility.

ABSTRACT. Advanced air mobility operations rely on dedicated takeoff and landing infrastructure, since all services require initial/final ground access to/from the vertiports. These characteristics call for a deeper investigation into the integration and collaboration between AAM and other ground transport modes to transport people to these AAM hubs. Existing studies do not adequately capture multimodal collaboration among different operators in this setting, such as determining subsidies between AAM providers and ride-hail operators for first/last mile connectivity. We adapt a mobility hub model between a platform (leader) and coalitions of travelers and mobility providers (followers) as a stochastic assignment game modeled as a bilevel problem with perturbed utility route choice lower level. By simplifying the dimensionality of the upper level problem, we adopt scalable algorithms for solving the problem to optimality and apply it to evaluate AAM designs for serving airport access in NYC.

ABSTRACT. Electric vertical takeoff and landing (eVTOL) vehicles are enabling Advanced Air Mobility (AAM) systems that rely on corridor networks with directional and capacity-constrained flows. This paper develops a tractable optimization framework for AAM operations that jointly determines vehicle dispatching and routing, four-dimensional flight trajectories, and corridor flow directionality. We formulate an integer optimization model on a time–space–lane network that exploits a flight-level subpath structure to decouple routing from trajectory dynamics. The model is solved via column generation, decomposing dispatching and flow decisions in a master problem from trajectory generation in a pricing problem, which is addressed using a tailored backward label-setting algorithm. The approach scales to realistic instances with up to 50 vertiports, 2,000 corridor conjunctions, and hundreds of trip requests. Results show substantial gains in operating profit and demand coverage relative to benchmark approaches.

ABSTRACT. L-shaped decomposition is a widely adopted strategy for solving large-scale two-stage stochastic programming problems, where a relaxed version of the original problem is repeatedly solved by incorporating Benders feasibility and optimality cuts. A key drawback of this approach is the need to generate a Benders cut for each individual scenario, which significantly increases the computational complexity. To address this issue, in this work, we propose a novel method which integrates the concept of the refinement chain of scenarios into L-shaped decomposition. In this approach, at each level of the refinement chain, the full set of scenarios is partitioned into subgroups, and the optimality cuts are generated based on the corresponding group subproblems. Theoretical relationships between optimality cuts at consecutive levels of the refinement chain are derived. The proposed approach is evaluated on a two-stage stochastic multicommodity network design problem under a mean-risk framework. Numerical results show promising computational performance improvements.

ABSTRACT. In this paper, we propose a novel ML-based approach for solving two-stage stochastic programs (2SP) by learning the objective coefficients of a surrogate master problem. We design two training losses: an imitation-based loss and an objective-based loss, corresponding to supervised and self-supervised learning, respectively. Both losses require differentiating the first-stage solution with respect to the input. The objective-based loss is particularly attractive because it circumvents the need to generate labels for first-stage decisions. However, it requires embedding the second-stage problem as an additional layer in the neural network, and differentiating the second-stage objective with respect to the first-stage solution in the backward pass. We show that, under mild conditions, this gradient can be obtained “for free.” Finally, we validate our approach on a real-world mining problem, achieving near-optimal solutions significantly faster than the extensive-form approach and outperforming a field-specific heuristic in profit.

ABSTRACT. SCGraph (Supply Chain Graph) is a dependency-free Python toolkit that provides fully offline, accurate multimodal routing and network-based distance estimation across global road, rail, and maritime networks. SCGraph leverages optimized adjacency list data structures and customized implementations of Dijkstra and A* algorithms, designed for sparse and large-scale transportation graphs. Accompanying SCGraph, SCGraph Data delivers global geospatial graphs that include worldwide highways, railways, and maritime networks. Based on our numerical results, SCGraph calculates the shortest path distances 10^5× faster than other Python toolkits such as OSMnx with NetworkX and 86× faster than Google Maps API, while remaining within 4% of Google Maps distances. By removing external dependencies, API costs, and the need for approximate circuity adjustments, SCGraph enables efficient and realistic modeling of logistics optimization problems, including vehicle routing, facility location, contextual shortest paths, and network design. It provides a reliable way to obtain accurate distances and paths at no cost.

ABSTRACT. Many industrial sites depend on diesel-powered light-duty trucks for transporting workers and small equipment, leading to significant greenhouse gas (GHG) emissions. Electrifying these trucks and planning their charging infrastructure are therefore essential. The operational schedules at the industrial site limit charging to designated parking times. Additionally, uncertainties in driving and parking behaviors greatly influence the number and types of chargers needed. We developed a decision-dependent, robust model for planning charging infrastructure with mixed charger types. The model is formulated as a mixed-integer linear programming (MILP) that optimizes the infrastructure by selecting multiple charger types and locations while determining individual charging schedules based on the chosen infrastructure and drivers’ behaviors, including range anxiety and charging abandonments. A case study was conducted at an open-pit mining site with 200 trucks involving more than 1.5 million integer variables.

ABSTRACT. Electric vehicle (EV) sharing is promoted as a low-carbon alternative, but its performance depends on how fleets are sized, charged, and operated. EV batteries degrade at rates that depend on charging intensity, depth-of-discharge cycles, and thermal conditions, creating long-term costs through service quality and battery replacement. Existing EV-sharing models optimize fleet, parking, and charging under uncertain demand, but typically ignore degradation and its life-cycle CO₂ implications. In stochastic optimization, robust satisficing (RS) has emerged as a target-based framework for planning under distributional ambiguity (Long et al. 2023). RS aligns naturally with cost, service, and CO₂ regulations, which are often expressed as performance targets and caps. We develop a two-stage stochastic model for EV sharing that embeds a degradation-aware time-space–state-of-charge recourse problem, prices battery wear inside profit, and imposes an RS objective on system-wide CO₂ emissions, quantifying how robust CO₂ requirements reshape optimal designs and their economic costs.

ABSTRACT. Truck electrification faces several challenges, including high purchase costs and insufficient charger coverage. Existing studies often overlook the joint effects of charging infrastructure planning and vehicle subsidies, or fail to utilize real-world freight tour data to reveal true charging demand. To address these challenges, this study develops a bi-level optimization framework for coordinating government investment in charging facilities and vehicle subsidies to promote electric truck adoption. Our framework models the government (upper level) to optimize charger placement and vehicle subsidies, minimizing public spending while meeting emission targets, while profit-driven carriers (lower level) make tour-based fleet electrification decisions. To efficiently solve this problem, we design a learning-assisted variable neighborhood search with a multi-task Gaussian process. Results demonstrate that effective decarbonization requires coordinated infrastructure-subsidy policies that are suited to actual freight operations, with strategic placement of chargers at high-impact nodes enabling meaningful emission reductions which is beyond what subsidies alone can achieve.

ABSTRACT. Climate variability increasingly disrupts transportation systems, yet conventional models often treat weather as a static input rather than a dynamic system driver. This study applies the Hybrid Dynamical Systems Thinking Approach (HDSTA), integrating Object-Process Methodology with ML to deliver both predictive accuracy and causal transparency. By mapping system interactions explicitly, HDSTA enables decision-makers to understand not only what models predict, but why. We examine climate-adaptive travel behavior across two contrasting cases: Israel's Highway 2 interurban bus corridor and New York City's Citi Bike network. Two novel metrics emerge: the Weather Resilience Transportation Index (WRTI), quantifying station-level vulnerability through dropout probability and relative ridership decline; and the Weather Resilience Infrastructure Factor (WRIF), measuring infrastructure's capacity to maintain ridership during adverse conditions. Results reveal context-dependent responses—extreme heat increased Highway 2 bus ridership while protected bike lanes in NYC strengthened resilience. This explainable approach enables targeted, equitable adaptation strategies grounded in systemic understanding.

ABSTRACT. On-Demand Multimodal Transit Systems (ODMTS) combine fixed-route buses with on-demand taxis to balance system operating costs and passenger convenience. Existing ODMTS studies have focused on latent demand modeling but have not accounted for uncertainty in passenger demand and travel times- two critical factors that strongly influence real-world performance. In this paper we develop a two-stage multimodal distributional robust optimization (DRO) framework for ODMTS design that explicitly incorporates both sources of uncertainty. Our ambiguity set captures modal uncertainty through probability variation and within-mode variability using mean absolute deviation. We propose a cutting-plane algorithm tailored to this setting, allowing us to obtain tractable solutions to bilinear uncertainty cases that cannot be solved with off-the-shelf solvers. Computational experiments on the Ann Arbor-Ypsilanti region in Michigan demonstrate the benefits of the DRO framework over deterministic approaches, highlighting its ability to deliver more resilient and reliable multimodal transit designs under uncertainty.