ABSTRACT. To reduce emissions, shipping companies are increasingly deploying dual-fuel vessels capable of utilizing multiple fuel types during voyages. These vessels enable compliance with emission regulations while minimizing operational costs through fuel-switching and speed optimization. Given the significant fuel price volatility across ports, decisions regarding fuel-switching, bunkering, and sailing speed must account for price uncertainties. This study develops a distributionally robust chance-constrained programming model based on the Wasserstein uncertainty set to minimize operational costs under uncertainty. The voyage between each pair of ports is subdivided into sub-legs to account for regional emission requirements or canal sections. This segmentation allows for the optimization of fuel usage proportions, sailing speeds, and bunkering strategies for each sub-segment. The model is validated using real-world data from COSCO Shipping. Numerical experiments demonstrate that the model can identify optimal solutions for real-world scale instances within practical computation times. Furthermore, the robust solutions significantly outperform those derived from traditional sample average approximation methods.

ABSTRACT. The aquaculture industry faces a major challenge in accommodating expected future growth while meeting ambitious greenhouse gas (GHG) emission reduction targets. To achieve this, a transition from fossil fuels to green alternatives is necessary -- though this shift may be costly and require new refueling infrastructure. In response, we study the Aquaculture Fleet Renewal Problem (AFRP), which involves optimizing the renewal of aquaculture service vessels to comply with GHG reduction goals. We propose a novel mixed integer programming model for the AFRP. Based on real data for a major Norwegian fish farming company, we explore the most cost-effective transition strategies up to 2055, focusing on the two most mature alternatives to fossil fuels: battery-powered electricity and hydrogen. The results show that electric vessels are preferred over hydrogen vessels when feasible. Furthermore, our analysis suggests that a more flexible transition pathway toward 2055 could achieve greater cumulative GHG reductions at the same cost compared to the targets set by the International Maritime Organization.

ABSTRACT. The decarbonization of the Norwegian deep-sea fishing fleet is crucial for achieving national greenhouse gas emission reduction goals. Although this fleet comprises only around 270 vessels and 15\% of the total active fishing fleet, it contributes to more than 75\% of the fishing sector's total emissions. One of the challenges in this transition is that these vessels often operate far from shore for extended periods. Alternative green fuels, such as hydrogen and ammonia, have significantly lower energy densities compared to marine gas oil (MGO), thereby limiting sailing ranges and potentially altering established sailing patterns--particularly when enlarging vessel fuel tanks is impractical and raises safety concerns. To address this limitation, this paper examines a maritime bunkering system that enables refueling at sea, ensuring uninterrupted fishing operations. The problem is formulated as a mixed-integer programming model using a time-space network representation to determine the optimal number of bunkering vessels, their routes, and synchronization with fishing vessels. A case study focusing on Norwegian deep-sea fishing operations in the Barents Sea, where the longest fishing trips occur, is considered to study refueling logistics. The computational results provide insights into the feasibility and efficiency of this ship-to-ship refueling strategy, highlighting trade-offs between operational costs and fishing operations.

ABSTRACT. In this paper, we study the problem of locating hydrogen fueling stations for zero-emission aquaculture vessels in Norway in Lofoten and Vesterålen. Transitioning these vessels to hydrogen fuel presents challenges due to their high energy demands and limited experience with hydrogen-driven well-boats. We formulate a location-routing problem considering the location of hydrogen fuelling stations and the routing of aquaculture vessels. The objective is to minimize the total costs while satisfying customer demand. The total costs consist of investment costs in fueling stations and fuel costs related to energy usage when sailing and servicing fish farming facilities. We compare an open location-routing problem and a closed-loop location-routing problem, as well as a single and a multi-trip problem. We further conduct a sensitivity analysis on various cost parameters. The computational results show that open location-routing problems lead to lower routing costs compared to their closed-loop counterparts. Among the cost parameters, the vessel costs have the highest impact on the solution structure, further highlighting the importance of vessel-redeployment in real-world problems.

ABSTRACT. We consider the Meal Delivery Routing Problem (MDRP) faced by online food delivery platforms, where customer orders arrive dynamically throughout the day and must be assigned to crowd-sourced couriers. Each assigned order is then picked up by a courier from a merchant and delivered to the customer. The platform faces several challenges in finding good courier-to-order assignments that minimize click-to-door time. First, couriers differ in behavior and make their own rejection, routing, and idling decisions. Experienced couriers may opt to reject certain courier-to-order assignments or decide to wait near popular restaurants. Second, assignment decisions should carefully weigh the trade-off between known direct and uncertain future assignment costs. For example, sending all couriers away from the city center might temporarily minimize the direct distance traveled by all couriers, but this can be detrimental to any future assignments. Third, to ensure responsiveness of the platform for its users, decisions---often with thousands of couriers and orders---must be made in seconds. Fourth, to increase efficiency at the cost of computational resources, the platform can bundle multiple (similar) orders and assign them to one courier. Similarly, the platform can choose to postpone assignment of orders until a later more favorable moment. Finally, uncertainty arises from stochastic food preparation times, fluctuating travel speeds, and the spatio-temporal distribution of orders. To face the challenges of the MDRP, we propose a cost function approximation framework using a Graph Neural Network (GNN), to estimate the implicit future cost per courier-to-order assignment. These implicit future costs are then added to a static one-to-many assignment problem which is solved repeatedly. To test the performance of our framework, we create an open-source and modular digital twin of the MDRP. Our digital twin allows us to simulate the delivery of millions of online food delivery orders in seconds. We validate the accuracy of our digital twin on real-world data from Meituan. We train our GNN-based framework with Twin Delayed Deep Deterministic Policy Gradient (TD3). Our results show a reduction of 11.6% in average click-to-door order delivery time and a reduction of 2% in order assignment rejections compared to a myopic policy that disregards implicit future costs and only minimizes the total courier-to-order distance. Our GNN-based framework also learns varying courier behaviors and spatiotemporal order distribution, enabling real-time decision-making for online food delivery platforms and related applications, such as ride-sharing services.

ABSTRACT. Learning-based methods for combinatorial optimization problems, particularly vehicle routing problems (VRPs), have shown remarkable promise in recent years. Among these, reinforcement learning (RL) methods are especially compelling, as they do not rely on pre-existing solution datasets. Most RL approaches adopt a sequential, autoregressive construction paradigm, where solutions are built step by step through a series of local decisions. While these methods have achieved impressive results, they often fall short of matching the performance of handcrafted operations research (OR) algorithms, particularly on large-scale instances. One reason for this is their inability to explore as many solutions as traditional approaches within the same amount of time.

We challenge the sequential construction paradigm and introduce a novel iterative search framework, Neural Deconstruction Search (NDS). Rather than constructing solutions from scratch, NDS operates by deconstructing existing solutions using a learned neural policy and reconstructing them through a fast, deterministic greedy insertion heuristic. Our approach builds on the principles of large neighborhood search (LNS) and ruin-and-recreate strategies, with a critical innovation: a deep neural network (DNN), based on the Transformer architecture, that guides the deconstruction process.

Unlike other neural methods that take only the problem instance as input, our neural policy consumes both the instance and a candidate solution. It then sequentially selects customer visits (nodes) to remove, effectively shaping the neighborhood structure of the search. The reconstruction phase sequentially reinserts the removed nodes using a greedy insertion algorithm.

NDS is trained using reinforcement learning, with the reward signal defined as the change in objective value before and after the deconstruction and reconstruction step. This training scheme enables NDS to operate without access to ground-truth solutions, making it broadly applicable. In contrast to prior learning-based methods that rely on neural policies for solution construction, NDS emphasizes rapid solution generation. It achieves significantly higher throughput, processing approximately 120,000 solutions per second for the CVRP with 500 customers. This enables NDS to perform an extensive search while still benefiting from the guidance of a DNN.

We evaluate NDS on established test datasets in the machine learning literature as well as on newly generated, more complex instances with clustered customer positions. We consider three challenging VRPs: the capacitated vehicle routing problem (CVRP), the vehicle routing problem with time windows (VRPTW), and the prize-collecting VRP (PCVRP). Across various problem sizes, NDS outperforms or matches state-of-the-art OR approaches such as HGS and SISRs, as well as the best-performing learning-based construction methods. For example, on the CVRP with 500 customers, NDS outperforms HGS and SISRs with a gap of -0.2%. To the best of our knowledge, this makes NDS the first learning-based approach to achieve parity with handcrafted OR solvers across a range of VRP variants.

ABSTRACT. In maritime logistics, the representative container vessel stowage planning problem (RCSPP) is a complex yet crucial combinatorial optimization challenge [1]. Solving realistic instances requires navigating large search spaces subject to numerous operational constraints. Search-based heuristics, including large neighborhood search (LNS), have been employed to find feasible or near-optimal solutions in reasonable time limits. However, the RCSPP involves a large and diverse set of competing constraints and objectives. As a result, LNS approaches have a large number of modifiers (i.e., destructors and constructors) that make modifier selection challenging, as the likelihood of choosing a promising modifier is low. This underscores the need for intelligent selection strategies to drive the search process more effectively.

In this work in progress, we propose a machine learning (ML) model that selects modifiers in an LNS search framework to solve instances of the RCSPP. The search process is formulated as a Markov decision process, where the state captures characteristics of the current solution, actions correspond to modifier selections, stochastic transitions reflect the effects of modifiers, and rewards are based on objective value and feasibility satisfaction. Additionally, we employ an actor-critic neural architecture with graph-based feature embeddings to exploit the structural properties of the vessel and voyage. The proposed ML-assisted search framework will be evaluated on real-world problem instances. Its performance will be benchmarked against general and problem-specific baselines to assess gains in solution quality and computational efficiency.

References: [1] A. Sivertsen, L. Reinhardt, and RM Jensen. “A representative model and benchmark suite for the container stowage planning problem”. In: Transportation Research Part E: Logistics and Transportation Review (Accepted) (2025).

ABSTRACT. Ocean freight pricing has become increasingly challenging and volatile due to global economic uncertainties and a growing number of supply chain disruptions. Traditional pricing methods based on historical data and averages have limited effectiveness, particularly when it comes to revenue maximization. Our research proposes a dynamic pricing optimization framework that integrates structural time series modeling with contextual bandit learning to support the setting of short-term (spot) and long-term (contract) prices in ocean container shipping.

The study integrates a variety of data sources obtained from a large ocean-freight company, including operational data about vessel utilization, customer data related with website search behavior and booking conversion, and price data about spot and time contracts, at the port-pair level on the trade route between the Mediterranean, the Middle East, and India. We apply Bayesian structural time series models to dynamically update price, operational and market information available prior to the spot price decisions. Subsequently, a contextual bandit is proposed to analyze the effectiveness of revenue-maximizing price strategies.

The research is still ongoing. We aim to validate our pricing model through with historical operational data across selected major port pairs. We also compare optimized with realized revenues over the research period. We expect an enhanced revenue, higher booking conversion ratio, better price responsiveness to market changes and an increased vessel utilization.

The immediate gains from a decision maker perspective is that the proposed framework relies on actual, internally available data, which eliminates the dependency on aggregate market indicators, e.g., GDP, consumer price index, and interest rates, which often fail immediately capture market signals. Moreover, this research provides practitioners in the ocean shipping industry with near real-time, data-driven support for key pricing decisions in a complex domain.

ABSTRACT. The Mobile Production Vehicle Routing Problem (MoP-VRP) is a new variant of the Vehicle Routing Problem in which production takes place on vehicles en route to customers. The objective is to determine both the production schedule and routing sequence for the vehicles such that the total travel cost and service delays are minimized. Due to the high complexity of simultaneous production and routing planning, CPLEX can only solve instances with 15 customers and a few with 20 customers to optimality. This work presents an enhanced model for this problem and a branch-and-price algorithm to find optimal solutions for larger instances. The algorithm incorporates specially designed effective dominance rules and different heuristic pricing methods to efficiently find the columns with negative reduced costs for the pricing problem. We have tested the model and the B&P on benchmark instances and have successfully solved many medium-size instances (with up to 50 customers) that have not been solved to optimality in the literature.

ABSTRACT. As transportation systems become more complex, the Multi-Objective Vehicle Routing Problem (MO-VRP) has emerged as a key research area in logistics, focusing on finding trade-offs between cost, delivery time, and environmental impact. For example, Karimi et al. (2024) developed an ALNS-based metaheuristic for a three-objective model, while Hou et al. (2025) proposed a hybrid NSGA-II approach to solve a multi-objective vehicle routing problem, ensuring feasible and constraint-respecting routing solutions across conflicting goals such as cost, time, and service quality.

The presented method combines mathematical modeling with heuristic optimization in a two-phase framework to address the multi-objective nature of modern vehicle routing problems, particularly in large-scale or high-density transportation networks. A key novelty of this approach lies in the integration of a Gini-based composite score for solution space reduction and the development of a multi-objective extension of the classical Clarke and Wright savings algorithm. In the first phase, the Gini score is used to filter out less-promising network components, improving computational efficiency while preserving high-quality solution regions. In the second phase, classical Clarke and Wright savings algorithm is enhanced by incorporating a composite utility function that simultaneously accounts for cost, delivery time, and environmental emissions. This extension transforms the original single-objective model into a multi-objective framework, enabling the evaluation of routing alternatives across conflicting goals. The combination of mathematical structure and heuristic flexibility allows for the generation of Pareto-efficient solutions tailored to the diverse priorities of sustainable logistics. Preliminary experiments on large scale logistics network dataset demonstrate the method’s effectiveness in maintaining high solution quality and fast computation while balancing these objectives. Its adaptability to various logistical contexts highlights its strong potential for real-world use in sustainable transportation planning.

References [1] Karimi, A., Zhang, L., Fackrell, M., & Thompson, R. (2024). A multi objective two-echelon vehicle routing problem with multiple delivery options. Transportation ResearchProcedia, 79, 377–384. [2] Hou, Y., Shen, Y., Han, H., Wu, Y., & Huang, Y. (2025). Adaptive constrained multi- objective differential evolution algorithm for vehicle routing problem considering crowdsourcing delivery. Applied Soft Computing, 169, 112517.

ABSTRACT. The proposed work investigates the effects of hub congestion on the design of hub networks with service level and environmental objectives. We model hub congestion by incorporating delays on processing time resulting from the high utilisation of some hubs capacities (Alumur et al., 2018). The formulated problem extends the modelling assumption of the green p-hub centre routing problem with strategic and operational decisions presented in Ibnoulouafi et al. (2023). It follows that distinct vehicles can arrive at their assigned hub nodes, at different times, for the processing of their carried demand flows. Furthermore, we assume that located hubs can be opened at different capacity levels each represented by the maximum hub processing capacity, and processing rate of inbound demand flows at the hub. Therefore, the interaction between the amount, arrival time, and processing rate of demand flows that needs to be processed at a hub, and the capacity of the hub are the main factors driving congestion. The single-allocation p-hub centre allocation problem was proven to be NP-Hard in Ernst et al. (2009) even with pre-established hub facilities. Further, solving large-scale instances for the mono-objective variant of the green p-hub centre routing problem cannot be accomplished in a timely manner using exact methods. Evidently, incorporating hub congestion effects into the multi-objective problem raises additional challenges towards the discovery of efficient solutions. In this paper, we propose a hybrid heuristic combining a population-based and a single-solution based meta-heuristic. Respectively, the non-dominated sorting genetic algorithm-II (NSGA-II) and the simulated annealing algorithm (SAA) are the two chosen meta-heuristics. Our aim is investigate the efficiency of Pareto frontier approximations achieved by hybrid method sharing the traits of both, the exploration-search and the exploitation-search algorithms.

ABSTRACT. Equity in Vehicle Routing Problems (VRPs) has gained significant attention in both theoretical and practical contexts. While the literature mainly focuses on classic VRPs, where the goal is typically to minimize the total routing cost to serve all customers and equity is measured in terms of workload distribution (e.g., route cost, time, distance, or number of customers served) (Matl et al., 2018), less attention has been given to equity in the context of Vehicle Routing Problems with Profits (VRPPs) (Archetti et al., 2014). In these problems, the primary objective is to maximize overall profit, with the possibility that some customers may not be served. In this setting, it becomes more relevant to ensure a fair distribution of profits among drivers rather than balancing workload.

The trade-off between equitable profit distribution among drivers and overall profit in VRPPs remains relatively unexplored despite its practical significance. Motivated by this gap, the aim of this work is to identify key properties and propose appropriate strategies for profit equity within this specific class of problems. More specifically, we explore two relevant strategies to incorporate profit equity into the optimization process: first, by introducing lower bound constraints that guarantee a minimum profit level for each driver, ensuring no driver is disproportionately disadvantaged; and second, by reformulating the objective function to a max-min criterion that seeks to maximize the profit of the least profitable driver’s route, thus promoting a balanced profit allocation across all routes.

Our study uses the Team Orienteering Problem with Time Windows (TOPTW), a well-known and representative VRPP variant, as a test-bed problem. We solve it using a branch-and-price algorithm tailored to integrate the proposed profit equity strategies. Numerical experiments are conducted on state-of-the-art benchmark instances to evaluate the impact of the proposed strategies on key performance indicators, including overall profit and the distribution of profits among drivers. We will present and analyze the computational results, highlighting the tradeoffs between profit maximization and equity, and offering insights into when and how equity constraints lead to meaningful changes in routing decisions and profit allocation.

References: Archetti, C., Speranza, M. G., & Vigo, D. (2014). Chapter 10: Vehicle routing problems with profits. In Vehicle routing: Problems, methods, and applications, second edition (pp. 273-297). Society for Industrial and Applied Mathematics. Matl, P., Hartl, R. F., & Vidal, T. (2018). Workload equity in vehicle routing problems: A survey and analysis. Transportation Science, 52(2), 239-260.

ABSTRACT. This paper introduces a novel Truck Appointment System (TAS) model for container terminals that integrates human aspects and operational priorities into the slot allocation process. A rule-based scoring mechanism evaluates hauliers based on their qualitative feedback provided after their visits to the container terminal. This feedback is processed using a text mining algorithm, enabling the system to quantify driver satisfaction levels and use them in appointment rescheduling and prioritization. The resulting score supports strategic slot allocation, balancing terminal constraints with haulier incentive. A mathematical model is developed to optimize the assignment of trucks to time slots, considering yard stacking efficiency, container urgency, vessel departing time and hauler scores. Experimental results based on realistic data suggest that the system can optimize slot-truck allocation, while reducing rehandling operations, and mitigating hauliers' resistance to using TAS.

ABSTRACT. In this paper we use a fleet renewal model to study which alternative fuels are adopted in the Norwegian fishing fleet to achieve target reductions in CO2 emissions. The objective is to minimize the overall costs consisting of investment costs in new and retrofitted vessels as well as operational costs covering fuel/energy, CO2 taxation, maintenance and periodic regeneration. We compare five alternative fuels for both a medium carbon price and a high carbon price. Our results show in both instances that e-MGO should play a major in role the transition to a zero-emission fishing fleet.

ABSTRACT. We consider the computational complexity of stowage planning with a specific view of Roll-on/Roll-off (RoRo) vessels. Stowage planning is the process of determining where cargo is loaded on a vessel. A key part of a stowage plan is to avoid shifts of cargo, i.e., unnecessary cargo movements. We focus on the slot-based RoRo variant where the deck of a vessel has a predefined set of equally sized positions and each position can be occupied by a cargo item. We show that it is \NP-hard to decide if there is a slot-based RoRo plan with no shifts. We also show how to generate slot-based RoRo instances requiring shifts. Finally, we present a greedy linear time algorithm always generating a plan with no shifts for a special RoRo case.

ABSTRACT. The tonne-mile metric serves as a measure to proxy demand for commodity transport in tramp shipping, yet predicting it is challenging due to complex trade route interdependencies and a dynamic geopolitical landscape. This study proposes a methodology for predicting tonne-miles based on trade flow modelling. Leveraging country-level export and import data, we frame tonne-mile estimation as a variant of the Minimum-Cost Flow Problem (MCFP) tailored towards tramp shipping and construct a time-expanded flow network that integrates multiple datasets to capture trade relationships. To evaluate the methodology, we perform a series of experiments across four key Clean Petroleum Product (CPP) commodities, demonstrating the ability of the model to reconstruct tonne-miles over a four-year historical period with a 2.1% mean absolute percentage error (MAPE). Finally, we showcase our method’s utility as a decision-support tool by forecasting future tonne-miles under several market scenarios utilising projected country-level export and import data. The framework can be used to estimate tonne-miles per commodity segment both regularly and during disruptions, such as geopolitical tensions.

ABSTRACT. The presented bulk shipping problem can be characterized as a many-to-many multicommodity pickup and delivery problem involving a heterogeneous fleet with compartments. The cargo sizes are flexible with a given tolerance and the cargo can have multiple origins and destinations and can be split. The freight rate is optimized, calculated as the operational cost divided by the weight of the shipped goods. It is important to note that the charter rates vary depending on the regions where the routes start and end, which is a novel aspect of this problem. Arc flow and path flow Mixed-Integer Linear Fractional Programming (MILFP) models were introduced. Since the objective function is nonlinear, two approaches address this issue: the parametric algorithm and the reformulation-linearization method. The results indicate that the path flow model combined with the parametric algorithm yields good solutions in both quality and time.

ABSTRACT. The Pod Repositioning Problem (PRP) in Robotic Mobile Fulfillment Systems (RMFS) involves selecting optimal storage locations for pods returning from pick stations. This work presents an improved solution method that integrates Adaptive Large Neighborhood Search (ALNS) with Deep Reinforcement Learning (DRL). A DRL agent dynamically selects destroy and repair operators and adjusts key parameters such as destruction degree and acceptance thresholds during the search. Specialized heuristics for both operators are designed to reflect PRP-specific characteristics, including pod usage frequency and movement costs. Computational results show that this DRL-guided ALNS outperforms traditional approaches such as cheapest-place, fixed-place, binary integer programming, and static heuristics. The method demonstrates strong solution quality and adaptability across different warehouse configurations, illustrating the benefit of learning-driven control within combinatorial optimization for warehouse systems.

ABSTRACT. Modern warehouse operations increasingly rely on coordinated teams of human pickers and Autonomous Mobile Robots (AMRs), raising new challenges in task assignment, scheduling, and system-wide efficiency. This project presents a hybrid optimization and learning framework that models swarm collaboration between pickers and AMRs, where multiple agents work in parallel under central task assignment, enabling synchronized and high-throughput fulfillment. The framework is designed with dual objectives: to minimize the overall order completion time (makespan) and to maximize the minimum satisfaction of human pickers' preferences based on their individual task-related sensitivities.

We formulate a Mixed-Integer Linear Programming (MILP) model that assigns items to pickers and AMRs and sequences their actions to minimize overall order completion time. The model integrates human-centric considerations by embedding individual picker preferences, capturing sensitivities to travel distance, item weight, volume, and shelf level, directly into the decision process. These preferences are derived from empirically obtained data and influence how tasks are allocated and ordered, supporting personalized workflows.

The exact optimization approach performs effectively on small to medium-scale problem instances. However, real-world warehouse settings involve much larger task volumes and resource teams. To address scalability, we train a reinforcement learning (RL) model using solutions from the MILP model on smaller instances. The trained RL policy is then used to extrapolate decision logic to larger settings, where it can support real-time, adaptive coordination among agents [4].

The contribution of this work is twofold. (1) A human-aware optimization approach that centrally coordinates (human) pickers and AMRs, integrating individual preferences into task assignment and scheduling, and (2) a scalable learning-based framework to extend optimal swarm strategies to practical warehouse environments. The research draws from advancements in e-commerce logistics [1], picker ergonomics [2], human–robot interaction [3], and AI-based decision-making [4].

References:

Boysen, N., de Koster, R., & Weidinger, F. (2019). Warehousing in the e-commerce era: A survey. European Journal of Operational Research, 277(2), 396–411.

Gabel, D., & Grosse, E. H. (2022). Human factors in manual order picking systems: A review. Computers & Industrial Engineering, 165, 107932.

Bortolini, M., Ferrari, E., Gamberi, M., et al. (2020). Advanced human–robot collaboration for industrial logistics: A review. International Journal of Production Research, 58(15), 4680–4707.

Lin, Y. H., Chen, T. L., & Shih, Y. C. (2021). A deep reinforcement learning-based approach for robotic task assignment in warehouse systems. Robotics and Computer-Integrated Manufacturing, 70, 102123.

ABSTRACT. This study investigates the storage location optimization in an automated storage and retrieval system (AS/RS). We introduce an optimization approach based on the Deep Q-Network (DQN) algorithm to enhance warehouse task efficiency and minimize stacker travel during storage and retrieval. To accelerate the algorithm training process, we integrate a prioritized experience replay mechanism. Furthermore, we decouple action selection from value estimation within the DQN framework to address the issue of value overestimation. The proposed model is evaluated against three heuristic methods. The experimental results demonstrate that our approach significantly outperforms these baselines.

ABSTRACT. This study explores the optimization of shuttle path scheduling in a cross-aisle transport system of an automated storage and retrieval system. Shuttles queue at designated positions, awaiting the allowance to execute their transportation tasks in accordance to a path reservation list generated by a traffic controller. The scheduling problem is addressed by introducing an intelligent A*-based approach that minimizes the overall task execution time by optimizing the order of the dispatching of the entries of the path reservation list. The proposed algorithm is capable of efficiently managing resource conflicts and temporal constraints. The evaluation is conducted on static and independent test sets that represent snapshot views of the system at fixed scheduling horizons.

ABSTRACT. Climate policy increasingly recognizes the necessity to remove atmospheric CO₂ alongside emission reduction. So-called Carbon Dioxide Removal (CDR) methods use chemical, biological, and technical processes to extract atmospheric CO₂ and store it in various mediums. Ocean Alkalinity Enhancement (OAE) is one promising CDR method that strengthens the ocean's natural CO₂ uptake by adding alkaline minerals such as processed limestone to the seawater. While techno-economic and ecological aspects of OAE have been addressed in previous research, supply chain optimization for the processing and marine distribution of limestone remains insufficiently explored. In this contribution, we present a mixed-integer linear optimization model for the supply chain design of such an OAE system. The model includes decisions on sourcing limestone, processing it into reactive form, transporting it, and allocating ship-based distribution areas in the ocean, while also accounting for emissions generated along the supply chain. A Norwegian case study illustrates the trade-offs between cost-efficiency, carbon removal capability, and system scalability. Our results quantify the potentials of integrating OAE into long-term climate strategies from an infrastructure and operations research perspective (Lindland et al, 2025).

Reference: Lindland, M., Wigand, E. A., Fagerholt, K., Meisel, F., & Herlicka, L. (2025). Supply chain optimization for Ocean Alkalinity Enhancement: A Norwegian case study. International Journal of Greenhouse Gas Control, 145, 104395. https://doi.org/10.1016/j.ijggc.2025.104395

ABSTRACT. The service network design problem (SNDP) is a critical tactical decision for logistics service providers (LSPs) in freight transport. Despite being a tactical problem, solving it effectively requires careful modelling of short-term operations, especially when considering aspects such as uncertainty and complex interactions between transport system players. To address this, we propose a simulation model that can estimate the operational costs for the LSP. Moreover, using the outcome of this model, we explore the use of machine learning techniques to estimate the predicted costs. Given the inherently combinatorial nature of the SNDP and the large size of real-world instances, we designed a metaheuristic algorithm to solve it, embedding the learning model to evaluate candidate solutions efficiently. Our final goal is to develop a fast and effective solution method for complex versions of the SNDP.

We focus on an SNDP setting involving multimodal freight transport, where road travel times are uncertain, and delays can cause missed connections, so dynamic real-time replanning decisions are required. To evaluate the performance of candidate plans under such conditions, we developed a discrete-event agent-based simulation model that captures those dynamics in detail. Then, to efficiently use the simulation insights in the optimization process, we fit a regression model based on data from a large set of simulated feasible solutions. This regression model estimates the expected operational cost of new solutions without requiring full simulation, enabling fast cost evaluations within a metaheuristic framework. We embed this surrogate model into a simulated annealing (SA) algorithm designed to solve the SNDP. Two variants are compared: one using the surrogate model (SA-R) and another using full simulation for every solution (SA-S).

Experiments were conducted on benchmark instances of varying sizes (number of transport requests). As expected, SA-S produced the most accurate solutions, since it directly evaluates each candidate using simulation. However, its CPU time grows rapidly with instance size, reaching 1800 seconds for the largest cases. In contrast, SA-R achieved high-quality solutions in under one second. The fast runtime of SA-R highlights its potential, especially for large-scale instances. However, there is still space for improvement in the methodology, since the solution gap is up to 10% relative to SA-S. A deeper analysis of the generated solutions reveals that SA-R tends to produce more conservative plans, accepting fewer requests to minimize delay penalties, which in turn reduces revenue. This behaviour indicates a bias in the regression model’s cost estimation, likely due to the limited set of solution features used in the prediction. These findings highlight opportunities to improve prediction accuracy. Future research will focus on developing more sophisticated machine learning models that incorporate additional solution characteristics and enhance the precision of cost approximations, aiming to close the solution gap while preserving computational efficiency.

ABSTRACT. The express delivery industry has witnessed significant growth over the past two decades. Air cargo hub location and service network design are two crucial decisions in express delivery operations. Traditionally, the hub location problem and service network design problem are solved sequentially, resulting in not-optimal or even infeasible solutions in the operations of express companies. This paper explores an air service hub network design problem (ASHNDP) that simultaneously addresses hub location and service network design. Specifically, the problem involves determining the optimal locations of hubs and assigning non-hub nodes to these hubs, while also deciding the number of cargo aircraft to purchase and the number of aircraft to operate within the hub network.

We first introduce a compact arc-based mixed-integer programming formulation for the problem, which is strengthened by valid inequalities. To solve large-scale instances, we reformulate it into a route-based formulation by Dantzig-Wolfe decomposition. We propose a Branch-and-bound-based exact approach which combines Benders decomposition and column generation. At each node of the branch-and-bound tree, we solve a mixed-integer linear programming using Benders decomposition, where the subproblem is solved by column generation. Through a tailored branching scheme, this approach can deliver an optimal solution to the ASHNDP. This approach is further computationally enhanced by two acceleration strategies: the lower bound lifting inequalities and primal heuristics for finding integer solutions.

Computational experiments using data instances from two famous datasets in hub location literature show that our solution can deliver optimal or near-optimal solutions for small-scale instances, and good feasible solutions for real-world scale instances. We show that our solution method outperforms CPLEX significantly. We conduct sensitivity analysis to observe the effects of changes in various parameters on the network. We also show the value of integrated optimization by comparing our approach with a method that address hub location and service network design sequentially.

ABSTRACT. We develop a non-cooperative game theory model to study strategic alliances among shipping companies and the resulting formation of transhipment networks. These networks are crucial for inventory sharing, improving market access, and optimising routes. Using a strategic network formation approach, we investigate how the structure of these alliances impacts firm efficiency. Our findings identify a specific point at which regular network structures achieve optimal efficiency. We also demonstrate that under certain conditions, hub-and-spoke networks are more efficient than regular networks. Furthermore, we find that networks with "structural holes", i.e. having gaps in the network, are less stable but foster greater cohesion within the shipping alliance. The results reveal a fundamental trade-off between the efficiency and stability of an alliance, highlighting the influence of network structure on the performance of transhipment networks.

ABSTRACT. We focus on a type of liner shipping network design problem (LSNDP) arising in cases where cargoes are to be transported between a major hub port and several regional ports. Because of the proximity of many of these local ports in some particular region, it makes sense to send long ship services from the major port, using big vessels, and then later transship these quantities, into smaller vessels to distribute to the regional ports, called feeder ports in this context. Such a type of network is called a feeder network.

For designing a feeder network, a combination of what are called mother and daughter routes are used. Mother routes are services that originate and end at the hub port while visiting a subset of other ports in between and transfer demands to daughter routes through some of these feeder ports, called transshipment ports. Daughter routes are ship services that start and receive their initial load from a mother route at some feeder port, called a transshipment port, and then serve a number of other feeder ports. This problem is called the Feeder Network Design Problem (FNDP). Additionally, we also consider a maximum time limit a container associated with a cargo can be in transit from the time of its pickup to the time of its delivery. We make the assumption that each daughter route have their weekly departure at the same time at a transshipment port, at exactly the weekly departure time of the associated mother route leaves from there. We also consider speed decisions on each port to port travel along the services, and a discretized version of the fuel costs which are realistically taken to be cubicly proportional to the speed.

We develop a matheuristic framework to tackle this complex problem. In the heuristic part , we handle the variables which decide how to build the network. For this, we use a solution representation which captures all of the edges in the network. In addition, a data structure is used to keep track of the rejection of the optional demands. The rest of the decisions are optimized by a mathematical model, which uses a penalized objective based on the network designed by the heuristic.

We start with a randomly generated solution and iteratively explore potential new and promising solutions by modifying specific mother routes and associated daughter routes or by merging different mother routes. After getting each possible network, the mathematical model identifies the parts of the solution the variables that need to be changed, and finds the optimal values for these. The heuristic operators make use of the output from the previous iteration of the math model to find and makes a selection of a few mother route networks to modify. The mathematical model of the matheuristic is shown to be efficient. This when combined with an adaptive heuristic selection for the different type of operators with different size of neighborhoods results in a promising solutions approach for this very complex logistics problem.