ABSTRACT. Agentic generative artificial intelligence is beginning to transform not only how researchers and practitioners interact with data and analytical models, but also how complex decision problems may be formulated, solved, evaluated, and continuously improved. Building on the emerging vision of agentic architectures for evolutionary computation, this keynote explores how similar principles could shape the next stage of operations research, logistics, and business analytics.

The talk examines a potential shift from isolated models and manually configured algorithms toward collaborative multi-agent ecosystems capable of connecting data preparation, mathematical modelling, simulation, optimization, predictive analytics, and decision explanation. In logistics, such systems may support routing, scheduling, network design, supply-chain resilience, and real-time operational coordination. In business analytics, they may provide a closer integration of descriptive, predictive, and prescriptive approaches within a unified decision-making process.

At the same time, the transition toward increasingly autonomous analytical systems raises important questions concerning reliability, validation, explainability, governance, domain knowledge, and human oversight. Rather than replacing operations researchers, analysts, or decision-makers, agentic generative AI may serve as an orchestration layer that lowers technical barriers, accelerates experimentation, and facilitates access to advanced analytical methods.

The keynote concludes by outlining a research agenda for trustworthy, interoperable, and human-centred agentic decision-support systems across operations research, logistics, and business analytics.

ABSTRACT. Large Neighborhood Search (LNS) is a powerful improvement framework for vehicle routing problems, but its performance depends strongly on the way neighborhoods are destroyed and repaired. This paper proposes ML-LNS, a neural LNS framework for the Capacitated Vehicle Routing Problem (CVRP). Instead of constructing complete routes directly, the learned policy parameterizes Slack Induction by String Removals (SISR), a strong structured destroy operator, inside an iterative search process. At each iteration, the policy combines precomputed static graph embeddings, dynamic route-state descriptors, and global instance/search features to select the removal anchor, the SISR coefficient, the destroy size, and repair-order scores. The resulting partial solution is repaired by a deterministic insertion heuristic, while feasibility control, acceptance, and best-solution management remain classical. We evaluate ML-LNS on CVRPLIB Uchoa X instances against controlled LNS variants and published neural routing/search methods. Under matched initialization, repair backend, seeds, and time budgets, ML-LNS improves over unstructured removal baselines, while the unlearned SISR operator remains a strong reference. The ablation study shows that adaptive destroy size is the most consistent learned component, and that dynamic and global features provide instance-dependent gains. With multiple learned SISR-style candidates, ML-LNS achieves competitive quality–time behavior, indicating that learning can effectively adapt a strong structured neighborhood to the current solution state.

ABSTRACT. Most Electric Vehicle Routing Problem with Time Windows (EVRPTW) formulations assume deterministic travel times, energy consumption, and vehicle availability. Real delivery systems are less stable: traffic congestion, mechanical failure, and battery depletion can propagate through a route and produce late service. In this research, we develop computational experiments on a two-stage Multi-Objective Stochastic EVRPTW (MOS-EVRPTW) that jointly optimizes expected energy consumption, time-window violation, and operational resilience. First-stage routes are fixed before uncertainty is observed; second-stage uses a fixed simulation policy to translate disruption events (congestion, breakdown, battery-depletion) during execution into recourse costs. The solution method is a simheuristic, and candidate route sets are evolved with NSGA-II, while Monte Carlo simulation estimates expected objective values under sampled disruption scenarios. Energy consumption is computed with a simplified physics-based model adapted from Goeke and Schneider. We conducted computational experiments on Goeke EVRPTW instances with up to 100 customers to compare the stochastic Pareto-front knee point against a deterministic VRPy baseline re-evaluated under the same scenarios. The results indicate that considering multiple objectives under uncertainty produces consistent improvements in energy efficiency and operational resilience. However, time-window performance depends strongly on customer spatial distribution, time-window flexibility, and the severity of disruption risks.

ABSTRACT. Recent neural methods for multi-objective combinatorial optimization involve solving preference-specific subproblems with a single model and have achieved competitive performance. However, they still suffer from limited learning efficiency and insufficient exploration of the solution space. This paper conducts a theoretical analysis that reveals the equivalence between this single-model paradigm and an implicit Mixture-of-Experts architecture. Furthermore, we propose a Preference-Aware mixture-of-experts (PA-MoE) framework that learns preference-specific representations while explicitly modeling \emph{preference-instance} interactions. By integrating a sparsely activated expert module with an innovative preference-aware gating mechanism, PA-MoE enhances preference-conditioned representation learning while preserving parameter efficiency. Moreover, PA-MoE is generic and can be applied to three different neural MOCO solvers. Experimental results on three standard MOCO benchmarks show that PA-MoE is able to generate a Pareto front with higher diversity, achieving superior overall performance.

ABSTRACT. Wind propulsion systems (WPS) are gaining attention as a means of reducing fuel consumption and emissions in commercial shipping. Their effectiveness depends strongly on encountered wind conditions, making weather routing a key factor in maximizing their value. Furthermore, routing decisions typically require balancing two conflicting objectives: minimizing voyage time and minimizing energy consumption. In response, we study the routing of a wind-propelled vessel which uses a WPS and conventional engine power for propulsion. We formulate a bi-objective mathematical model and represent the trade-off between the two objectives through Pareto-optimal solutions generated by solving a one-to-many shortest path problem over a time-space network. The model is applied to a case study in the North Sea and North Atlantic, using historical weather data to establish idealized performance bounds and forecast data to represent routing decisions made prior to departure. Forecast‑optimal routes are subsequently evaluated under realized weather conditions. Results show that deviations between the predicted and realized energy consumptions vary by region and season. Furthermore, a simplified economic analysis indicates that allowing longer voyage times may reduce total sailing costs, which include fuel, carbon and time-charter costs.

ABSTRACT. When designing a new schedule for a shipping liner service, an important parameter to consider is its reliability. High schedule reliability allows the shipping line to charge higher prices, as customers can keep less safety stock and reduce costs in their end-to-end supply chains. However, obtaining a realistic reliability estimate for a new service has not been readily available in the industry. Traditionally, designers would have to wait for vessels to sail many round trips, possibly taking several months, before collecting enough data to refine the schedule. We propose a method incorporating uncertainty in distances and port stay times, to quickly provide reliability estimates on terminal and service level, allowing designers to iterate on schedules before having initiated a single rotation with a vessel. We further propose a validation methodology based on retrospective schedule designs that enables measuring the goodness of fit of such simulation frameworks despite operational interventions in the observed data.

ABSTRACT. This study examines the use of a Construction Consolidation Center (CCC) to improve urban construction logistics. We develop a multi-period, multi-material optimization model with heterogeneous vehicles and multiple trips to jointly determine delivery schedules, vehicle assignments, and routing decisions for both suppliers and the CCC, subject to time windows and limited on-site storage space. The model also incorporates synchronization constraints on vehicle arrivals at construction sites to reflect restricted unloading capability. The objective is to minimize total logistics costs while accounting for the physical and temporal constraints of dense urban construction environments.

Preliminary computational experiments show that a CCC can improve logistics efficiency by consolidating deliveries, reducing supplier transportation fees, and alleviating on-site storage pressure. Sensitivity analyses indicate that the benefits of the CCC are especially pronounced when on-site storage space is limited or costly. Finally, we test several urban logistics policies and their impact on cost and delivery patterns.

ABSTRACT. With the growth of global maritime trade, tugboat scheduling in giant ports faces frequent dynamic disturbances and spatial and temporal complexity. Traditional heuristics suffer from low online computational efficiency, while existing deep reinforcement learning faces immense action spaces in large scale scheduling. Thus, this paper proposes a hierarchical decision frame-work with Reinforcement Learning (HDRL). First, a dual level collaborative mechanism, comprising upper-level task decision and lower-level heuristic resource assignment, is constructed. Then, the environment is modeled as a bipartite graph of tasks and tugboats, and a spatial distance bias is explicitly injected into the Cross-Attention calculation, and Proximal Policy Optimiza-tion (PPO) algorithm is used for training. Finally, experiments using data from the Port of Ningbo-Zhoushan validate the approach. Compared to man-ual dispatching, HDRL maintains the completion rate while reducing average empty travel distance and delay by 44.8% and 60.6%, respectively. Com-pared to NSGA-II, it achieves a superior balance in comprehensive rewards and delay control, reducing the daily online decision time to under 2 sec-onds.

ABSTRACT. In maritime hub-and-spoke networks, feeder services connect feeder ports with hub ports. In multi-hub feeder networks, shared feeder ports create scheduling flexibility but complicate task-to-hub assignment when vessel states and service times change over time. To address this issue, this paper studies a disruption-aware feeder transport scheduling problem with flexible task-to-hub assignment in a multi-hub network. A hierarchical rolling scheduling framework is proposed, in which the upper layer determines task-to-hub assignment and dynamic task re-assignment, while the lower layer solves vessel routing and scheduling subprob-lems for each hub. The lower-level subproblems are solved using adaptive large neighborhood search (ALNS), and upper-level decisions are iteratively updated using operational feedback. A disruption module is further introduced to simulate service-time uncertainty and trigger rescheduling over rolling periods. Experi-ments on a regional network modeled on the Guangdong–Hong Kong–Macao Greater Bay Area show that, within the tested setting, the proposed framework improves computational efficiency and reduces tardiness compared with one-time global optimization. Dynamic reassignment further reduces tardiness with limited changes in empty sailing, and the framework remains feasible across the tested disruption-budget levels.

ABSTRACT. Abstract. India's 12.5 million medium- and heavy-duty trucks are 2% of the national fleet but produce 45% of transport-sector greenhouse-gas emissions, 53% of particulate matter, and over 60% of on-road black carbon [1]. This study evaluates battery-electric trucks (BEVs) as a zero-emission replacement, using operational data from nine Container Corporation of India (CONCOR) routes out of Chennai. It has three parts: a stochastic cost model, machine-learning tools for fleet operations, and policy design. 1. Social Total Cost of Ownership (S-TCO) The model prices environmental and health externalities into each trip. Diesel externalities work out to ₹16.16/litre (CO₂ ₹10.56; NOx ₹5.00; PM2.5 ₹0.60); electricity externalities to ₹3.10/kWh (grid CO₂ ₹2.60; grid upgrade ₹0.50) [2]. The objective function is: S-TCO = E[C_cap + Sum_{t=1..T} (C_op,t + C_ext,t - S_oil,t)(1+r)^-t - V_res,T (1+r)^-T] where C_cap is capital cost (₹1.5 cr for an 800-1,000 kWh BEV truck [2]), C_op,t operating cost, C_ext,t externalities, S_oil,t avoided oil-import savings, V_res,T residual value, discounted for the thin secondary market [3], and r the social discount rate. 2. Machine-Learning Logistics Tools An LSTM State-of-Health model is trained and validated on public battery-degradation data, including LFP cells, the chemistry used in Indian trucks, aged at 43-46 C. Those pack temperatures come from our physics-based duty-cycle simulation of summer load; pairing the two gives route-level battery life. Two further tools are proposed rather than built out. A Proximal Policy Optimisation agent, set up as a Markov decision process, is shown in proof-of-concept form against the 20-40% of miles the sector runs empty on return legs, subject to EV range limits. A Federated Learning scheme would let fragmented fleets coordinate charging without sharing route data; the ICCT puts the annual saving from coordinated smart charging at ₹3.5-8.5 billion across twelve national e-truck corridors [4]. 3. Results and Policy Across the nine routes, BEVs save an average of about ₹6,463 per trip in social terms. The saving comes from the gap between diesel and electricity externalities. Short- and medium-haul routes under about 300 km, which can charge at the depot, are the near-term case. Long-haul is harder: capital cost is ₹1.5 cr against ₹45-50 lakh for diesel, and corridor charging is sparse [1]. Four policy measures follow. First, replace flat PM E-DRIVE subsidies with carbon credits paid against verified emissions reductions. Second, use concessional loans and partial-credit guarantees to protect Indian exporters under the EU Carbon Border Adjustment Mechanism (CBAM), in force from January 2026, given that about USD 8 billion of India's exports lie in CBAM-exposed sectors. Third, set up state-level public-private partnerships for renewable microgrids at freight depots. Fourth, fund retraining for the estimated 80% of the workforce without EV skills. NITI and RMI put the cumulative CO₂ avoided on this trajectory at roughly 3 Gt by 2050 [1].Keywords: Zero-emission trucks · Total cost of ownership · Freight electrification · Reinforcement learning · Carbon Border Adjustment Mechanism · Green economics

ABSTRACT. This paper addresses the critical challenge of optimizing hazardous healthcare waste (HHW) collection in Tunisia through a specialized case study of healthcare facilities in the Sfax metropolitan area. Formulated as a Capacitated Vehicle Routing Problem (CVRP), the study aims to minimize total transportation distance while strictly adhering to vehicle capacity and safety constraints. We introduce a synergistic Neuro-Heuristic Hybrid Framework designed to bridge the gap between high-dimensional representation learning and classical combinatorial optimization. While Deep Reinforcement Learning (DRL) offers rapid adaptability, its application in logistics is frequently hampered by inefficient exploration and high-dimensional search spaces. To mitigate this, we propose the Attention-based Vehicle Routing Policy Network (Att-VRPNet), an Actor-Critic architecture utilizing Multi-Head Self-Attention to endogenously extract spatial dependencies and relational node features. The framework incorporates a dual-stage enhancement: (1) Heuristic Guidance, which embeds Clarke and Wright Savings logic to guide early-stage exploration toward high-reward trajectories, and (2) Iterative Structural Refinement via a post-decisional 2-Opt layer. Validation on real-world HHW data demonstrates that Neuro-Hybrid 2Opt (NH-2Opt), reduced the optimality gap closely approximating the global optimum. Our results confirm that integrating attention-based representation learning with deterministic heuristics provides a scalable, and computationally efficient paradigm for the safe and sustainable management of healthcare waste logistics.

ABSTRACT. Hours-of-Service (HoS) regulations limit drivers’ working and driving times to ensure road safety. Even though they play a critical role in realistic transport planning, they remain relatively underrepresented in the vehicle routing literature. The objective of this paper is twofold. First, an overview is provided of the existing literature that integrates HoS regulations into routing and scheduling models. Specifically, two problem categories are investigated: the Truck Driver Scheduling Problem (TDSP) and the Vehicle Routing and Truck Driver Scheduling Problem (VRTDSP). A total of 37 studies are analyzed, discussed and classified based on the regulatory framework used, HoS parameters, modeling assumptions and solution methods. Second, this theoretical perspective is supplemented by examining how HoS regulations influence route planning procedures, drawing on insights from interviews with route planners in different transportation companies. The findings reveal gaps in the modeling of operational uncertainty, decentralized planning structures, complex real-world constraints, and the integration of HoS regulations with emerging technologies such as electric vehicles.

ABSTRACT. Driving passenger or goods vehicles requires a professional card, obtained after completing mandatory training at accredited centers. This process involves scheduling both theoretical lectures and practical driving sessions under limited resources. A key challenge for these centers is to ensure that each group completes all components of the program. Groups assigned to practical sessions may miss some lectures; therefore, the decision-maker must ensure that all groups sharing the same classroom during a given period receive an equal total duration of theoretical lectures to avoid conflicts. This work proposes Variable Neighborhood Search (VNS) heuristics to generate timetables while optimizing resource utilization under regulatory and availability constraints, constituting the first application of VNS to this timetabling problem. The proposed approach comprises four VNS variants that employ local search strategies based on swap and shift neighborhood moves. Each variant incorporates multi-time-column and row-based perturbation mechanisms to escape local optima and enhance the exploration of the solution space. Benchmark experiments show that all VNS variants achieve optimal or best-known solutions. Moreover, the variant based on swap moves demonstrates superior performance on average.

ABSTRACT. Acquiring reliable travel demand data remains a critical challenge in mobility studies. Although methods have been developed to extrapolate a fully-sized mobility demand from external or sampled data, they mostly do not apply to cases where no data exist, such as synthetic cities and networks. This paper explores the extent to which origin-destination (OD) matrices can be derived solely from network topology indicators. Using open-source data from five heterogeneous French regions, road networks are modelled as graphs aggregated at the neighbourhood level, and a comprehensive set of single-zone and pairwise topological metrics is computed. Relationships between network structure and human mobility flows are then analysed through Spearman correlation, followed by the use of Histogram Gradient Boosting models for both flow magnitude classification and OD flow value regression.

Results reveal no single, city-independent correlation common to all regions, which is expected given the diverse urbanisation levels across the studied areas; correlations are nonetheless considerably stronger in regions presenting a contrasted urban-rural structure. Network topology identifies the sparse structure of OD matrices and highlights major demand hubs across all regions, but fails to fully discriminate finer variations in intermediate flow intensities. These findings suggest that whilst simple topological indicators are not sufficient to fully reconstruct OD matrices, they provide a valuable foundation for identifying the structure and principal hubs of travel demand, in existing as well as hypothetical or synthetic cities.

ABSTRACT. The present work proposes an inventory-control-driven methodology for designing an integrated distribution network with capacitated facilities and vehicle routing. The framework combines a common-frequency Economic Order Quantity policy, a capacitated facility-location stage based on an angular/radial sweep procedure, and Capacitated Vehicle Routing Problem route construction using the Clarke-Wright savings algorithm. The approach transforms customer-level demand into synchronized replenishment quantities, capacity-feasible service regions, candidate distribution-center locations, and operational delivery routes. The methodology was evaluated using a rescaled SJC benchmark instance with 402 customers and GIS-based coordinates. Results show that the common-frequency EOQ policy reduced inventory-management cost from $5,611,122.30 to $3,765,717.62. The capacitated clustering stage generated 20 distribution centers satisfying 570,774 kg of periodic demand under a 30,000 kg capacity limit. The routing stage produced 117 feasible routes with an estimated transportation cost of $14,741.37, supporting cost-efficient and operationally feasible distribution planning for medium-scale logistics systems requiring coordinated inventory and transportation decisions simultaneously.

ABSTRACT. The Capacitated Vehicle Routing Problem with Time Windows (CVRPTW) is a fundamental NP-hard optimization problem in logistics. Solving large-scale instances remains computationally challenging for exact solvers. This paper introduces a multilevel graph coarsening and refinement strategy that aggregates customers into meta-nodes based on a spatio-temporal distance metric. The reduced problem is solved using classical heuristics, a QUBO-based simulated annealing solver, and the D-Wave Leap hybrid BQM solver, then expanded back into the original space with arrival times recomputed and constraint violations recorded.

Comprehensive experiments on Solomon benchmark instances demonstrate that our method significantly reduces computation time for classical heuristics, with solution-quality effects depending strongly on the solver and instance family. For QUBO-based solvers, experiments across all 56 Solomon instances at N = 5 and N = 10 customers show that coarsening improves solution quality and, on clustered C-type instances, simultaneously reduces vehicle count and route duration with no feasibility loss.

On the D-Wave Leap hybrid backend, the same route distance improvement transfers directly, with approximately 26% improvement and p < 0.001, while post-coarsening feasibility exceeds the simulated annealing baseline. Coarsening effectiveness is strongly instance-structure dependent: C-type instances achieve 100% post-coarsening feasibility with measurable quality improvements, while narrow-window random R-type instances present structural constraints that limit achievable coarsening depth.

ABSTRACT. Industry 5.0 introduces a new industrial paradigm centered on the human, resilience and ethical integration of advanced technologies within manufacturing and supply chain systems. Despite growing interest in this paradigm, existing assessment frameworks remain limited by their reliance on assumptions of equal weighting and compensatory aggregation methods. Indeed, strong performance in some dimensions can mask critical weaknesses in others. These limitations reduce the reliability and interpretability of maturity assessments in human-centric industrial environments. To address these challenges, this paper proposes the Industry 5.0 Human Capability Framework (HCF), an optimization-based assessment framework designed to evaluate human-centric capabilities in industrial supply chains. The proposed framework integrates the BestWorst Method (BWM) to determine weightings for evaluation criteria, based on expertise, and Min-Max Objectives Programming to minimize compensatory effects when assessing maturity. Unlike traditional averaging approaches, the proposed optimization model ensures that weaknesses in critical dimensions are not masked by high scores in other areas. The assessment framework analyzes organizations along five key dimensions: Human–technology synergy, employee sustainability & wellness, continuous capability development, responsible technology governance, and operational human resilience. These dimensions are operationalized through 30 sub-dimensions identified through a systematic literature review and an expert validation process. The proposed framework has been empirically validated using data collected from 20 manufacturing companies in the automotive and aerospace sectors. The results demonstrate that the optimization-based assessment modified 42% of maturity classifications compared to traditional averaging methods, highlighting the importance of non-compensatory assessment approaches in the evaluation of Industry 5.0. Statistical validation confirmed acceptable internal consistency and strong interrelationships between the proposed dimensions. Finally, the HCF provides organizations with a structured and transparent decision-making tool to evaluate human-centric capabilities, identify performance gaps, and support strategic Industry 5.0 transformation initiatives.

ABSTRACT. The distribution of IT infrastructure resources across geographically heterogeneous university campuses constitutes a canonical computational logistics problem: a multi-constraint, multi-objective resource routing problem in which demand is stochastic, supply is budgeted, and the objective function couples efficiency (aggregate service level) with equity (distributional fairness). This paper repositions the classical IT infrastructure planning problem as a logistics network design problem and proposes an AI-calibrated decision support framework for its solution. The framework integrates three components: (1) a data-driven demand forecasting model using PLS-SEM coefficients (β₁ = .371, f² = .281) empirically validated on 723 students across 12 Moroccan public university campuses as the IT service elasticity function; (2) a multi-objective integer linear programme (MOILP) formalising the IT Infrastructure Allocation for Resilience Maximisation (ITARM) problem under budget, minimum provision, and access threshold constraints; and (3) a Hybrid PSO-Grey Wolf Optimiser (HPSO-GWO) solving the ITARM problem at operational scale. Computational experiments on the 12-campus Moroccan network demonstrate a 34.7% improvement in aggregate Digital Resilience Index over baseline allocation, a 48.9% reduction in inter-campus Gini inequality, and statistically significant superiority over standalone PSO, GWO, Genetic Algorithm, and Differential Evolution (all Wilcoxon p < 0.01). The framework advances computational logistics theory by establishing empirically grounded IT service demand functions for university network design, and generates an actionable, AI-supported infrastructure investment roadmap for Morocco's Maroc Digital 2030 strategy.

ABSTRACT. This paper addresses the Electric Vehicle Charging Scheduling Problem (EVSP) with uncertain arrival times in large-scale charging stations. The objective is to minimize total tardiness under stochastic arrivals modeled through Sample Average Approximation (SAA). We propose a Tabu Search with Q-learning and Adaptive Parameter Sampling (TS-QAPS), a reinforcement learning–based adaptive operator selection framework integrated within Tabu Search (TS). Unlike conventional Q-learning approaches, TS-QAPS dynamically adjusts the learning rate and discount factor online, thereby eliminating the need for manual hyperparameter tuning. The proposed method is compared with three Tabu Search variants: Classical Tabu Search (CTS), Multi-Armed Bandit-based Tabu Search (TS-MAB), and Q-learning-based Tabu Search (TS-QL). Experiments conducted on 126 stochastic instances with n=20--500 vehicles show that both Q-learning approaches outperform CTS and TS-MAB. TS-QAPS achieves the best overall performance, with a mean gap of 0.83% to the best-known solution and improvements of up to 21% over CTS. TS-QLearning ranks second with a mean gap of 1.15%, while TS-MAB deteriorates significantly on medium and large instances.

ABSTRACT. Maritime ports are under increasing pressure to reduce vessel emissions during berthing operations. While mobile Emission Capture System (ECS) barges have recently emerged as a promising green initiative, their operational deployment introduces a complex scheduling challenge that has received little attention in the literature. This study investigates the ECS barge routing problem, which aims to dynamically allocate and route a fleet of ECS barges to serve berthed vessels while minimizing total generalized operational cost.

The proposed optimization model explicitly considers operational constraints including vessel-specific service time windows, battery state-of-charge limitations, pollutant tank capacity restrictions, depot visits for recharging and tank cleaning, and service continuity requirements. For vessels with extended berthing durations, multiple barges may be sequentially assigned to ensure uninterrupted emission capture throughout the entire berthing period.

Due to the combinatorial complexity of the proposed optimization problem and the dynamic nature of vessel arrivals and service requests, a Multi-Agent Reinforcement Learning (MARL) framework is developed in which ECS barges act as cooperative decision-making agents. A Shapley value-based reward allocation mechanism is introduced to address the multi-agent credit assignment problem and improve cooperative learning. Furthermore, a K-means clustering strategy is integrated to partition vessels according to their spatial locations and berthing time windows, thereby reducing the effective action space and accelerating policy convergence.

Computational experiments are conducted on problem instances of varying scales and compared against exact optimization using Gurobi and a Genetic Algorithm benchmark. The results demonstrate that the proposed MARL framework achieves high-quality routing solutions while significantly improving computational scalability for large-scale port operations. The study provides the first decision-support framework for ECS barge deployment and offers practical insights into the operational management of mobile emission capture systems in sustainable ports.

ABSTRACT. Many decision-support tasks in computational logistics depend on supervised prediction on tabular data, such as estimating travel times or forecasting demand. For a long time, gradient-boosted decision trees (GBDTs) like XGBoost have been the strongest models for this kind of structured data, but they need heavy hyperparameter tuning. Recently, tabular foundation models like TabPFN [1, 2] have challenged this approach by using a transformer pretrained offline on synthetic datasets. At inference time, TabPFN does not update its weights and predicts new rows in a single forward pass by looking at the training rows in its context window [3]. Our main goal is to determine whether we should use this new foundation model instead of classical supervised machine learning methods for transportation data. Real transportation datasets routinely contain large numbers of records, which exceed TabPFN's native row limits and force us to subsample. This creates a tension between the two approaches. In this paper, we evaluate the accuracy and runtime trade-offs between TabPFN and GBDTs on transportation data to compare their performance. Our preliminary results show that tabular foundation models can match or outperform well-tuned supervised machine learning methods in this setup, helping to decide whether eliminating the tuning effort is worth the scale limitations for practical transportation services.

ABSTRACT. We study a Continuous Interval Selection Problem on identical machines with multiple availability intervals, motivated by waste bin collection scheduling in the construction logistics sector. We propose an exact MILP formulation inspired by Osanno et al. (2025) and a Kernel Search matheuristic and evaluate them on generated instances with up to 300 tasks. The Kernel Search emerges as the central contribution: it matches the MILP optimum on small instances, equals or outperforms it for larger instances with more than 70 tasks, and remains the only viable approach for instances with 200 tasks and more, where the MILP finds no feasible solution within one hour.

ABSTRACT. The Gaza Strip has experienced repeated cycles of armed conflict, resulting in widespread destruction of infrastructure, the economy, and civilian life. Despite the extensive level of devastation, systematic and institutionalized methods for managing conflict-related waste and supporting coordinated reconstruction are limited. This paper presents 14 context-specific strategies, organized within an integrated post-war debris logistics and circular reconstruction supply chain framework for Gaza, based on a comprehensive review of disaster waste literature and qualitative interviews with 16 key stakeholders. By filling a crucial knowledge and operational logistics gap in post-war recovery efforts, these strategies offer practical and adaptable solutions to support sustainable rebuilding in war-affected areas. While focused on Gaza, the insights apply to other conflict zones, especially in the Middle East, through three main pillars: hazard mitigation, reverse logistics optimization, and circular reconstruction supply chain integration. This research sets the foundation for future policy efforts in post-conflict environmental and infrastructure recovery.

ABSTRACT. Energy-intensive industrial systems require proactive tools capable of monitoring electricity imbalance and supporting operational decision-making. In steam-based production environments, electricity gen- eration and consumption are tightly coupled through process conditions, steam-network dynamics, and equipment operation, making energy-balance management a challenging task. While recent studies have successfully applied machine learning to energy forecasting, most focus on predicting a single target, such as electricity demand or consumption, and provide limited support for balance-oriented decision making. This paper proposes an interpretable machine-learning framework for predictive energy-balance monitoring and decision support in industrial energy systems. Unlike direct balance-estimation approaches, the pro- posed framework separately models electricity production and electricity- consumption behavior before computing the resulting energy balance. The framework integrates industrial time-series preprocessing, feature engineering, predictive modeling, SHAP-based explainability, deficit-risk assessment, and scenario-based evaluation of corrective actions within a unified Energy Management System architecture. The proposed approach is evaluated using real industrial time-series data collected from the JFC4 production environment of OCP Group. Ex- perimental results show that Random Forest achieved an R2 of 0.898 for electricity-production prediction, while Linear Regression achieved an R2 of 0.910 for electricity-consumption-proxy estimation. The result- ing predictive energy-balance framework achieved an R2 of 0.891 on the held-out test set. Compared with a direct balance-estimation baseline evaluated under the same test protocol, the proposed dual-model frame- work achieved lower error and stronger deficit-detection performance, improving recall from 0.67 to 0.78 and F1-score from 0.79 to 0.87. SHAP analysis identified high-pressure steam admission to the turbo-alternator as the dominant driver of electricity production and balance behavior. Scenario analysis further indicated that combined corrective assumptions can reduce predicted deficit occurrence by approximately 31%. The results demonstrate that combining energy-balance decomposition, explainable machine learning, and operational risk assessment can pro- vide valuable decision-support capabilities for industrial energy manage- ment. The proposed framework establishes a practical bridge between industrial monitoring systems and future optimization-based EMS solu- tions.

ABSTRACT. This paper addresses the integrated Berth Allocation and Quay Crane Assignment Problem (BACAP) under realistic operational constraints. A mixed-integer optimization formulation is proposed to jointly optimize vessel berthing, crane allocation, and crane scheduling while considering berth capacity limitations, vessel arrival times, noncrossing constraints, and crane reconfiguration operations. The objective function simultaneously minimizes vessel stay time and crane reconfigurations in order to improve operational efficiency, schedule stability, and sustainability. To evaluate the impact of integration, four planning architectures are investigated, including sequential, partially integrated, and fully integrated approaches. The proposed formulations are solved using the CPSAT solver and evaluated on synthetic benchmark instances of different sizes and congestion levels. Computational results show that integrated planning improves solution quality compared with sequential approaches, particularly in terms of vessel stay time and crane reconfigurations. The study also highlights the computational scalability challenges associated with fully integrated exact optimization models.

ABSTRACT. Reducing road-related emissions in freight transport requires incentive mechanisms that improve environmental performance without imposing excessive economic pressure on shippers, operators, or public authorities. This paper proposes a data-driven decision-support framework for sustainable intermodal freight transport. Real freight-demand flows are combined with road, rail, and maritime infrastructure layers to generate direct-road, road-rail-road, and road-marine-road candidate routes. Each route is evaluated using cost, travel time, emissions, and policy-adjusted generalized cost. The problem is represented as a leader-followers setting in which a regulator sets rail subsidies, marine subsidies, and road penalties, while freight flows respond through minimum generalized-cost assignment. The framework evaluates fixed policies, group-based mechanisms, stress-scenario robustness, and metaheuristic feasibility of emission-reduction targets. Experiments on 26,790 candidate routes show that rail alternatives form the main intermodal option space, whereas marine alternatives remain spatially selective. Road penalties and combined mechanisms produce stronger modal-shift responses than isolated subsidies, but they reveal a trade-off between public subsidy expenditure and road-penalty burden. Under the tested constraints, the 1% CO2 reduction target is feasible, with a best feasible reduction of about 1.33%; higher requested targets remain infeasible although the best achieved reductions stay close to 1.34%. A comparison with random search, NSGA-III, and MOPSO confirms that this frontier is stable across the tested procedures.

ABSTRACT. Generative artificial intelligence can summarize logistics dis- ruptions and propose preliminary maritime responses, but a fluent rec- ommendation is useful only when it satisfies executable constraints. This paper develops a proof-of-concept framework for auditing AI-generated maritime logistics recommendations before they are treated as imple- mentable decisions. A logistics hallucination is defined as a recommen- dation that appears coherent in natural language or structured form but violates at least one encoded logistics constraint. The framework combines a compact knowledge graph, formal constraint auditing, and an exact optimization-repair model. The optimization component is in- tentionally standard: a capacity-constrained flow-allocation model with eligibility restrictions and an unserved-flow penalty route. The contribu- tion is the integration of semantic verification and optimization repair as a safety layer for generated maritime decisions. The framework is demonstrated on a public aggregate Strait of Hormuz stress case using 2025 country-level oil/product flows, public disruption indicators, and alternative-route capacity estimates. Under a 95% restricted-Hormuz sce- nario, the model allocates 0.99 mb/d to residual Hormuz capacity, 5.00 mb/d to the Saudi route, 0.70 mb/d to UAE ADCOP, and leaves 13.18 mb/d unserved in the model. The results are not a geopolitical forecast or a berth-level validation; they show how encoded feasibility checks can expose capacity-infeasible recommendations, identify unsupported LLM claims, and convert recommendations into auditable allocations

ABSTRACT. Demand forecasting at the SKU-store level is operationally consequential but methodologically difficult: most retail products show sparse, intermittent demand on which strong gradient-boosted baselines such as LightGBM al-ready perform well, yet degrade systematically on cold-start products, promotional periods, and anomaly windows. We propose selective LLM-assisted forecast revision composed with statistical calibration as a deployable architecture for sparse-demand forecasting, and characterize its behavior across an intervention policy space. The method invokes a large language model only on forecasts flagged by observable difficulty triggers, then applies a single multiplicative shrinkage for residual bias. On M5 California foods (5,748 product-stores, 160,944 test observations) across three pre-registered policies, aggregate accuracy peaks at the Moderate policy (−3.50% MAE) with an in-verted-U shape while worst-case per-cell degradation rises monotonically with aggressiveness. The Conservative configuration achieves a deployable balance (−3.32% MAE, worst-case bounded at 1.25%) and is recommended for deployment.

ABSTRACT. Administrative delays in accrediting professional driver training centers arise primarily from the mismatch between resource capacity and scheduling demands in provisional annual plans. To minimize submission revisions and accelerate center accreditation, the current dual-manual system for formulating and auditing submitted plans must be replaced by a fully automated validation mechanism. In this study, we introduce an Integer Linear Programming (ILP) model for tactical capacity planning in training centers. This model serves as a pre-evaluation tool to overcome the aforementioned limitations. The ILP formulation relies on time-indexed variables to synchronize discrete training session timelines with the continuous utilization of shared, heterogeneous resources. We validate the ILP model under two distinct demand scenarios. First, the model operates under unconstrained demand, allowing full flexibility to schedule the number of drivers for each training type. Second, the model is subjected to regulatory bounds, where the scheduled demand must respect both minimum thresholds and maximum capacities. Computational results highlight that vehicle availability acts as the primary binding constraint, whereas mandatory demand bounds across distinct training categories reduce the center's total revenue.

ABSTRACT. This paper investigates a single-truck-single-drone collaborative routing problem for the pickup of fresh agricultural products in mountainous areas. Existing models largely rely on trucks as the only mobile energy hubs for drones, resulting in a tightly coupled routing framework for trucks and drones. This design severely limits the operational range of drones and lacks the spatial flexibility needed to accommodate the rapid quality decay of perishable goods during collection. To overcome these spatial constraints, this paper develops a single-truck-single-drone collaborative routing model under a decentralized battery-swapping strategy. Deploying cost-effective energy relay nodes directly at farm pickup points, we establish a dual replenishment network integrating dynamic battery swaps on mobile trucks with fixed swaps at farm locations. This fundamentally decouples the drone's spatial reliance on the truck. Experimental results show that, facilitated by the continuous multi-visit capabilities and the decentralized battery-swapping mechanism, the routing network topology can dynamically adapt under extreme asymmetric payloads.

ABSTRACT. Container retrieval operations in terminal yards play a critical role in maritime transport. In practical vessel loading operations, terminals follow vessel loading plans that specify the required container type or group at each loading step, requiring terminal operators to dynamically select one container from multiple feasible yard containers accordingly. In this paper, we formulate this yard-side container selection process as a Markov decision process for online decision making, where a learned policy sequentially chooses feasible containers based on the current yard configuration and the vessel loading requirements. To effectively capture the structural information of container selection process, we adopt an attention-based encoder--decoder architecture to represent the yard state and loading requirements, while a feasibility mask is incorporated to restrict the action space to valid container choices at each step. Computational experiments show that the proposed approach achieves competitive performance across scenarios and demonstrates promising zero-shot generalization to larger unseen instances.