ABSTRACT. We address the problem of efficiently routing a fleet of electric vehicles (EVs), to serve customers spread across the territory within their time windows, with possible (also partial) recharges en-route at recharging stations (RSs). The objective is to minimize the total operational cost, given by the vehicle usage costs, recharge costs and drivers’ wages. We assume that the energy consumption depends on both the load and speed of the EV, where the latter is a decision variable that can vary between a minimum and a maximum value. Moreover, in this work, we assume that the RSs are equipped with different recharging technologies allowing faster recharges with higher costs. A Mixed Integer Linear Programming (MILP) model is formulated in which RSs can be used more than once in the routes without cloning them. Due to the problem complexity, a tailored matheuristic approach is developed. The method combines a Large Neighborhood Search framework with exact optimization components. Problem-specific removal strategies are employed to explore the solution space, while new solutions are generated by solving a restricted MILP defined on a dynamically selected subset of routing arcs. Computational experiments are carried out on a newly generated set of benchmark instances. The results demonstrate the effectiveness of the proposed approach and highlight the impact of incorporating realistic energy consumption and multiple recharging technologies. Further comparisons with existing methods for simplified problem variants confirm the competitiveness of the proposed solution strategy.

This research is within the project ``SUPporting Ecological tRanSitiOn iN mId-haul logistiCs (SUPERSONIC)'', funded by the MUR program `Bando relativo allo scorrimento delle graduatorie finali del bando Progetti di Ricerca di Rilevante Interesse Nazionale'' (PRIN) 2022 (code: 2022WXFN5B, CUP: I53C24002250006).

ABSTRACT. This paper focuses on the Electric Vehicle Routing Problem with Time Windows and Stochastic Waiting Times at Recharging Stations that is modeled through two-stage optimization. The routes, built through an Adaptive Large Neighborhood Search metaheuristic in the first stage, are corrected in the second stage according to the realizations of the stochastic waiting times. Unlike the current literature, costumers, who may not be served on time by an electric vehicle (EV) due to possible delays during recharges, may be handled by another. Numerical results on the benchmark instances prove that collaboration among EVs allows for additional improvements of up to 4.65\% with respect to the approach in which skipped customers are served by a new route, demonstrating the significant value of horizontal cooperation in managing uncertainty in EV routing.

ABSTRACT. This study proposes a heuristic approach for defining the routes of a fleet of homogeneous unmanned aerial vehicles to spray pesticides on a set of target points located within a crop field. The proposed approach aims to quickly generate high-quality feasible solutions by integrating a constructive procedure with an optimization phase based on mathematical programming. The numerical results demonstrate that the framework is highly efficient and maintains an average execution time of less than 10 seconds across various scenarios. The robustness of the framework suggests that it can serve not only as a standalone solver for large instances, but also as an efficient warm-start mechanism for exact MILP solvers. Finally, the proposed approach constitutes a contribution toward practical and scalable routing methodologies for precision agriculture, successfully reconciling mathematical rigor with the computational performance required in real-world field operations.

ABSTRACT. In this paper, we propose a hybrid optimization approach that combines Reinforcement Learning (RL) principles with the Variable Neighborhood Search (VNS) framework to address the Capacitated Vehicle Routing Problem (CVRP). Specifically, the Q-learning algorithm is utilized to dynamically guide the selection of neighborhood structures within the VNS. The reinforcement learning is based on improvements in solution quality, allowing the algorithm to learn which neighborhoods are most effective over time. The proposed Q-VNS method is evaluated on a benchmark set of CVRP instances. Computational results demonstrate that our approach consistently outperforms the classical VNS in terms of solution quality. On average, the proposed method achieves solutions with 1.5 times lower optimality gap compared to the baseline VNS, while maintaining comparable computation times, demonstrating the potential of reinforcement learning to enhance the performance of metaheuristics in combinatorial optimization problems.

ABSTRACT. Vehicle Routing Problems (VRPs) involve optimizing delivery routes from a depot to multiple customers. We address the Electric Vehicle Routing Problem with Road Junctions and Road Types (EVRP-RJ-RT), a variant that incorporates battery constraints and models road junctions and road types, which dictate speed limits. While the Clarke and Wright heuristic is a standard solution method that can be easily adapted to this variant, the inclusion of road junctions significantly increases the computational overhead of calculating the so-called savings values. In this work, we replace the traditional savings computation with a forward pass of a simple neural network, bypassing the junction-induced bottleneck. Additionally, we introduce a set of realistic benchmark instances modeled after electric delivery operations in various European cities.

ABSTRACT. The open pickup and delivery problem with time window and optional cross-docking (OPDPTW-OCD) represents a critical challenge in middle-mile e-commerce logistics. This problem deals with multi-product aspect which requires precise synchronization between vehicles when eventually transshipping at cross-dock. To tackle this problem, we propose a hybrid genetic algorithm enhanced with reinforcement learning for dynamic parameter control. We compare different genetic algorithms including one using Taguchi method for robust parameter tuning, supervised machine learning and random parameter control. To cope with this static parameters settings, we use reinforcement learning in order to dynamically adjust crossover and mutation rates for each problem instance and for each generation. Experimental results show that guided parameter control based on reinforcement learning achieves better solution quality by reaching the best solutions for the most instances. In particular, the difference in performance is more visible when compared to genetic algorithm variants for problem instances with larger size.

ABSTRACT. The Physical Cell Identity (PCI) assignment problem is an NP-hard multi-objective optimization task in LTE and 5G networks. Network operators must assign integer identifiers to cells while simultaneously minimizing signal interference, deployment failures, and operational costs. To address this challenge, we propose a memetic algorithm based on the Non-dominated Sorting Biased Random-Key Genetic Algorithm (NS-BRKGA) framework. Our approach is distinguished by a specialized decoder that first applies a fast greedy repair to handle hard constraint violations and then executes a lexicographically guided local search. The improved assignment is re-encoded into the random-key chromosome so that local-search improvements are inherited by the evolutionary operators. We conducted an extensive empirical evaluation on a benchmark of 291 real-world instances, comparing our approach against five widely used multi-objective evolutionary algorithms, including NSGA-II. Under the same runtime budget, NS-BRKGA achieves substantially better non-dominated fronts according to two well-established performance indicators, while scaling well to highly constrained instances.

ABSTRACT. Renewable Energy Communities are key to the EU Clean Energy package, empower-ing citizens to jointly produce, share and valorize flexibility to support system securi-ty. We propose a bilevel optimization (BLO) model for an energy community in which a central entity (leader) maximizes profit by setting internal buying/selling elec-tricity prices, operating an energy storage system (ESS) and procuring energy from the wholesale market to maintain system balance. The leader’s decisions include price setting and ESS management, considering all energy transactions with the communi-ty (multiple consumers with different profiles) and the external market. At the lower level, consumers aim to minimize electricity costs by scheduling flexible and non-flexible loads, photovoltaic generation, stationary storage and electric vehicles, sub-ject to individual operational constraints and preferences. Energy sharing among consumers is allowed, prioritizing self-consumption. The community pursues two lexicographically ordered objectives: (i) minimizing the maximum individual cost (fairness), and (ii) minimizing total cost, ensuring no consumer incurs a higher cost than that obtained in the first stage. An optimistic BLO framework is adopted, requir-ing solving an additional optimization problem for each leader decision to guarantee the leader’s best outcome under the community’s optimal response. The BLO prob-lem includes bilinear terms in the upper-level objective, while the lower-level problems reduce to MILPs once prices are fixed. To address the problem, we develop a hybrid metaheuristic combining a genetic algorithm (GA) to explore the leader’s price decisions with an exact solver to handle the lower-level lexicographic optimization. Each GA individual encodes buying/selling price vectors and evolves through selection, crossover, mutation and repair (to ensure price feasibility) operators. For every candidate solution, three lower-level problems are solved using Gurobi, and the leader’s profit defines fitness. We also implemented a deterministic BLO algorithm. Although exact, it requires solving a mixed-integer nonlinear program at each iteration, in addition to the lower-level problems, which is highly sensitive to numerical issues. Computational results show that the hybrid metaheuristic achieves high-quality solutions, as assessed against upper bounds obtained from the deterministic approach.

ABSTRACT. Differential evolution remains among the most effective evolutionary algorithms for global optimization, yet its performance depends on balancing exploration and exploitation. This paper proposes Guided Differential Evolution (GDE), which alternates explicitly between an exploration mode using the classical DE/rand/1 scheme and a guided mode that incorporates historical population information via a sliding window of previous generations. The latter steers candidate solutions toward a population centroid, reducing fluctuant behavior and improving convergence stability without added complexity. Retaining the classical DE structure, GDE augments it with a guidance mechanism and deterministic phase scheduling. Experimental validation on the CEC'2017 benchmark suite confirms competitive performance across diverse function categories, demonstrating the effectiveness of the proposed phase alternation strategy. https://github.com/espartan0007/GDE

ABSTRACT. Combinatorial optimization algorithms often perform very differently across problem instances: an instance that is easy for one algorithm may be difficult for another. To predict which algorithm will perform best on each instance, we need large sets of problem instances that are both diverse, covering a wide range of scenarios, and discriminatory, highlighting performance differences among algorithms. Generating such labeled instances manually is costly, motivating automated methods that produce rich, pre-labeled instance sets. To address this, we explore a Novelty Search approach that optimizes a linear-weighted objective balancing feature-space novelty and portfolio-based discrimination. Our main contribution is a problem-agnostic descriptor learned by an autoencoder (AE) from raw instance data, which replaces handcrafted features and yields a compact, low-dimensional embedding suitable for measuring novelty. Using the linear ordering problem (LOP) as a case study, and a local search scheme under a portfolio of three neighborhood variants (insert, swap and inversion), we compare the AE-based descriptor against a carefully designed feature-based descriptor. Overall results show that AE-driven descriptors are practical, generalizable, and effective for generating large, pre-labeled instance sets that support algorithm selection beyond the LOP.

ABSTRACT. Ant Colony Optimization (ACO) is a powerful heuristic algorithm inspired by the foraging behavior of ants, widely used to solve complex combinatorial optimization problems. However, the performance of ACO is highly dependent on the careful calibration of key parameters, including the pheromone evaporation rate and pheromone increment rate. While much research has focused on optimizing individual parameters, fewer studies have explored the interactions between multiple parameters or the impact of colony grouping strategies on algorithm performance. This study aims to address this gap by examining the influence of pheromone-related parameters, initial ants' distributions, and colony group configurations on the efficiency of ACO in weighted networks. Our findings suggest that lower values of the evaporation rate and pheromone increment generally improve the success rate (SR) and reduce costs. The decreasing distribution typically yielded the best SR and lowest costs, while the increasing distribution underperformed. Additionally, as the number of groups increases, performance declines, especially with higher evaporation rate values. However, moderate levels can mitigate this decline, indicating that carefully optimized parameter and grouping configurations can enhance overall algorithm performance.

ABSTRACT. We propose a reference framework to apply the multiparent path relinking (MPR)methodology. MPR is an extension of path relinking (PR) that has been proposed and described in the literature but, to the best of our knowledge, it has never been applied so far. PR is a trajectory-based neighborhood search strategy that explores paths that traverse pairs of solutions. PR has been widely used for search intensification purposes within metaheuristic implementations. To show how MPR can be embedded in more than one setting, our proposal consists of both employing MPR as a post-processing step in a greedy randomized adaptive search procedure (GRASP) and as the combination method in a scatter search (SS). We use the power dominating set problem (PDSP) as our testing platform because its difficulty and structure enable the design and assessment of various strategies. The PDSP seeks to find the minimum placement of measurement devices in an electrical network to monitor the entire system. Our computational experiments are designed to identify the contribution of each element of our proposed MRP implementations.

ABSTRACT. Network preprocessing techniques can simplify combinatorial optimization problems without compromising solution quality. In the context of information diffusion, the Perfect Awareness Problem seeks the smallest set of initial spreaders to ensure complete awareness across a social network. This work introduces two novel reduction strategies for this problem: edge reduction and leaf blindness. The edge reduction strategy identifies and eliminates redundant edges based on domination relationships, while leaf blindness handles terminal nodes that do not contribute to the spreading process. These preprocessing methods are integrated within an Iterated Local Search framework that employs reservoir-based local search and perturbation mechanisms. The proposed reductions significantly decrease instance complexity, enabling more efficient exploration of the solution space across networks of varying sizes and densities. Experimental evaluation on 840 benchmark instances demonstrates the effectiveness of our approach, achieving 809 best-known values and discovering 33 new best solutions.

ABSTRACT. We propose a variant of the Construct, Merge, Solve and Adapt (CMSA) metaheuristic for solving the Least Cost Influence Maximization Problem (LCIP). This problem models influence diffusion in a social network. It seeks to determine payments to a selection of vertices to ensure a certain minimum portion of all vertices is influenced, i.e., they adopt a given idea, while minimizing the total cost. Influenced individuals may subsequently influence their neighbors, potentially triggering a cascading diffusion process. We develop several construction heuristics and introduce a strategy that combines them to derive bounded payment intervals, thereby restricting the search space. Furthermore, we present a Mixed Integer Linear Programming (MILP) formulation, which is employed both as a standalone exact approach and as the embedded solver within the CMSA framework. Computational experiments on benchmark instances derived from real-world social graphs with 30,398 to 154,908 vertices demonstrate that the proposed CMSA approach consistently outperforms both the standalone MILP model and the construction heuristics.

ABSTRACT. The Fair Max-Min Diversity (FMMD) problem extends the classical Max-Min Diversity Problem by incorporating group fairness constraints, aiming to select a diverse subset while ensuring balanced representation across predefined categories. The FMMD problem is an NP-hard combinatorial optimization problem, and existing studies have mainly focused on approximation approaches.

In this work, we first propose an exact solution approach based on a binary search scheme combined with a node-packing formulation. Compared with standard integer linear programming models, this approach is significantly more efficient and is able to solve small- and medium-sized instances to optimality, thus providing a strong baseline for evaluating heuristic methods.

To tackle large-scale instances, we investigate and compare two metaheuristic paradigms: the memory-less Greedy Randomized Adaptive Search Procedure (GRASP) and the memory-based Probabilistic Tabu Search (PTS). Within the PTS framework, we develop two variants with different memory integration strategies: PTS1, which incorporates short- and long-term memory during the improvement phase to avoid cycling and encourage diversification; and PTS2, which uses information from elite solutions to guide the construction phase. Our results show that incorporating memory mechanisms leads to substantial improvements over the memory-less approach in terms of solution quality and robustness, with particularly strong gains observed when memory is applied during the improvement phase.

Building on these findings, we introduce an Advanced-PTS algorithm that enhances PTS1 by integrating strategic oscillation and memory-based perturbation. Strategic oscillation enables controlled exploration of feasible and infeasible regions through a relax-and-repair mechanism applied to fairness constraints, while memory-based perturbation leverages frequency information to diversify the search and escape local optima. Extensive computational experiments on synthetic benchmarks and real-world datasets show that Advanced-PTS consistently outperforms GRASP and basic PTS variants, and achieves solution quality that is competitive with exact methods on instances that are beyond the reach of optimal solvers.

ABSTRACT. In the so-called “mobile broadband slice” of a wireless 5G network, heterogeneous content streaming services delivered over radio technologies must reach a vast number of mobile users across wide geographic areas. These services need to be fast, resilient, and with high-quality performance, while minimizing energy consumption in a shared, dynamic, and stochastic network environment. To this end, we propose an integrated approach to manage two underlying optimization problems at the operational management level: user assignment to network access points (APs) and path-finding between an access point and a target service node in the 5G network. The former is modeled as a constrained plant location problem and solved using a tailored iterated local search-based procedure (ILS); the latter is solved using an intelligent path-finding algorithm based on A*. Specifically, the ILS explores neighboring feasible solutions to the CPLP by jointly considering user–AP assignment decisions and the resulting effects on the capacity and utilization of the links connecting APs to the server node. Both the ILS and A* solutions are embedded in a simulation model that mirrors the stochastic and dynamically changing conditions of the network, along with additional features such as bursty user arrivals, failure-prone nodes, and service retrials for rejected users. We discuss neighborhood structure, intensification/ diversification, and related tuning efforts for the ILS parameters. Preliminary results show the central role of a proper metaheuristic within a framework effectively combining combinatorial optimization and discrete-event simulation. This amounts to successfully supporting decision-making in applications spanning from remote healthcare management and intelligent transportation services for citizens to manufacturing, assembly, warehousing systems, and last-mile logistics for businesses.

ABSTRACT. Given a set of data points in a multidimensional space, we consider the problem of selecting a subset of cardinality k whose variance is minimum. We refer to this problem as the minimum variance k-subset problem. Such a problem arises in applications where only a limited number of data points can be selected, and internal homogeneity is paramount. In this paper, we propose a nearest neighbors-based heuristic for the problem as well as a GRASP method to scape from local optima. Our computational experiments indicate that the proposed GRASP method consistently attains the best known solutions across a broad collection of benchmark datasets and subset sizes. The nearest neighbors heuristic also demonstrates competitive performance while requiring negligible computational time. Together, the results show that high quality and computationally efficient solutions can be obtained for the minimum variance $k$-subset problem in practice.

ABSTRACT. This study presents a comparative evaluation of four metaheuristic algorithms: Cuckoo Search (CS), Dwarf Mongoose Optimization (DMO), Butterfly Optimization Algorithm (BOA), and the proposed Butterfly Optimisation Algorithm 2 (BOA2) for multi-objective routing optimisation in Mobile Ad Hoc Networks (MANETs). Simulations were conducted in MATLAB under varying node densities (20, 50, 100, 200, 500, and 1000 nodes). Performance was evaluated using Packet Delivery Ratio (PDR), throughput, end-to-end delay, and standard deviation to assess robustness and stability. Results show that network density significantly influences algorithm performance. At 20 nodes, BOA and BOA2 achieved the highest PDR (37.33% and 37.38%) and throughput (30.66% and 30.70%), along with low delay values, indicating efficient routing in less congested environments. As node density increased, performance differences became more pronounced. CS experienced substantial degradation in both PDR and throughput, demonstrating limited scalability. DMO showed moderate improvements over CS but did not maintain consistent superiority at higher densities. Although BOA remained competitive at 50 nodes, its performance declined at 100 nodes. BOA2 consistently achieved the best overall performance at higher densities, recording the highest PDR (18.65%) and throughput (15.32) at 100 nodes while maintaining competitive delay. Performance recovery was observed at higher densities (200 and 500 nodes), attributed to improved connectivity and route diversity, and BOA2 consistently achieved the highest PDR and throughput. Under ultra-dense conditions (1000 nodes), overall performance declined due to congestion and routing overhead; however, BOA2 maintained comparatively stable performance across all metrics. These results confirm that adaptive enhancements in BOA2 improve exploration–exploitation balance, scalability, and routing reliability, making it the most effective algorithm for dense MANET environments.

ABSTRACT. Real-world decision problems often involve uncertainty: travel times fluctuate, customer demands are stochastic, processing times vary, and financial returns are unpredictable. Deterministic optimization models commonly approximate this uncertainty using expected values, which simplifies the problem and enables tractable analysis. However, solutions obtained in this way may not perform well once variability is explicitly considered. Simheuristics address this limitation by integrating metaheuristic optimization with simulation. The procedure typically begins with a deterministic approximation based on expected values to guide the initial search, where a metaheuristic explores the solution space and generates promising candidate solutions. The main difference is the use of Monte Carlo simulation to explicitly account for uncertainty when evaluating these solutions. Through repeated sampling, the simulation estimates performance measures such as expected objective value, variance, and the probability of constraint violations. To manage computational effort, a short simulation is embedded during solution construction to provide preliminary evaluations, followed by a more extensive simulation applied to a pool of the best candidates identified. Recent extensions expand the methodology’s flexibility: fuzzy simheuristics use fuzzy logic when probabilities are hard to define, and sim-learnheuristics incorporate machine learning to handle dynamic, evolving environments. These approaches have been successfully applied in transportation, logistics, production scheduling, healthcare, and finance, consistently outperforming deterministic solutions when evaluated under real-world uncertainty.

ABSTRACT. Decision-making becomes particularly challenging when dealing with NP-hard optimization problems. While metaheuristics are traditionally designed for deterministic environments, real-world systems are rarely static or predictable. To address this limitation, simheuristics emerged to improve solutions when variables present stochastic behavior, while learnheuristics were developed for scenarios where dynamic variables can be predicted using machine learning (ML) models. However, many contemporary problems involve both random uncertainty and predictable dynamic patterns. The sim-learnheuristic framework integrates simulation, metaheuristics, and ML, providing a unified approach to handle these challenges by combining the strengths of each method.

The proposed methodology follows a three-stage process. In the first stage, the problem is simplified into a deterministic version by replacing random variables with their expected values; a metaheuristic then generates new solutions using ML to manage dynamic inputs, which are then evaluated by the simulation module to identify their performance under stochastic scenarios. During the second stage, the best solutions are subject to intensive simulation runs to improve the precision of their performance estimates. Finally, the third stage utilizes these results to conduct a risk and reliability analysis, providing probabilistic insights that detail the behavior and stability of each solution.

ABSTRACT. We propose a framework that combines biased-randomized heuristics, agile optimization, and transformer-based deep learning for real-time combinatorial problem solving. Biased-randomized heuristics introduce controlled stochasticity into classical deterministic rules, generating thousands of high-quality solutions while preserving heuristic logic. The resulting solution set is used to train a transformer model that acts as a deep heuristic, capturing complex decision patterns across thousands of variables. Once trained, the transformer provides fast solutions to new instances and adapts efficiently to changing inputs. To further improve exploration, we incorporate biased randomization into the transformer's decoding process, allowing multiple solution trajectories to be generated in parallel. This enables fast exploration and selection of high-quality solutions within milliseconds, even for large instances. The proposed approach integrates heuristic reasoning, data-driven learning, and parallel probabilistic search into a unified framework for dynamic, real-time optimization.

We evaluate the methodology on the team orienteering problem. Compared with exact MILP solutions, deterministic greedy inference produces an average optimality gap of 5.11%. In contrast, probabilistic decoding strategies substantially improve performance: geometric biased randomization with moderate exploration achieves an average gap of 0.41%, consistently producing near-optimal solutions across all instances. These results are obtained with an average inference time of 0.06 seconds per instance using parallel multi-trajectory decoding, whereas exact optimization requires approximately 18 seconds on average and exhibits significant runtime variability, showing that the proposed framework achieves near-optimal performance while maintaining stable real-time execution speed.

ABSTRACT. Cutting stock problems (CSPs) are challenging combinatorial optimization problems that arise in many industries, including steel, wood, and textiles, when stock materials need to be cut into smaller pieces. In this paper, we address the two-dimensional flexible stock $x$ cutting stock problem (2D-FS$x$CSP), where $x \in \{\text{W}, \text{H}, \text{S} \}$, representing “width”, “height”, and “size”, respectively. In this setting, large rectangular plates with variable width, variable height, or both variable width and height are cut to satisfy a given set of demands for rectangular item types with fixed orientation using two-stage, non-exact guillotine cuts. We propose a metaheuristic framework that focuses exploration on the most critical decisions by applying a split-and-sort (S\&S) strategy through a sequence of strips. We illustrate this framework using genetic and ant colony optimization algorithms. To evaluate performance and assess the impact of the S\&S strategy, we tested the proposed approaches on randomly generated instances.

ABSTRACT. Wireless Local Area Network (WLAN) planning requires op- timizing Access Point (AP) placement while accounting for complex sig- nal propagation effects. In this work, we propose a heuristic approach termed Carousel Directional Relocation for Access Point positioning (CDR- AP), a search strategy that combines cyclic relocation with a directional improvement mechanism. Starting from a feasible configuration, CDR- AP iteratively refines AP positions through directional local search while dynamically adjusting the spatial step size to balance global exploration and local intensification. The heuristic is specifically designed to reduce the number of signal-propagation evaluations and radiomap construction, which represent the primary computational bottleneck. Experimen- tal results were performed across 20 scenarios, showing that CDR-AP achieves near-zero optimality gaps compared to MIP baselines.

ABSTRACT. Chemotherapy treatments are personalized, long, painful, and stressful for patients. To reduce disruptions to patients’ daily lives and alleviate stress, home chemotherapy has been introduced. This system also helps free hospital beds for shorter treatments; however, it results in significant logistical complexity. Due to the complexity of drug dosages and injection schedules, and to facilitate patients’ daily lives, patient visits are preferably scheduled in a time-consistent manner; that is, each patient should receive treatment within a similar time slot across all visits. To model this system, we formulate a bilevel bi-objective optimization problem. The leader level assigns a unique visit time slot to each patient for every treatment day over a multi-week planning horizon. The follower level, defined on a daily basis, consists of vehicle routing problems with soft time windows. At both levels, the objective functions minimize travel time, earliness, and tardiness, thereby capturing the trade-offs between operational efficiency and service quality. To tackle this bilevel bi-objective problem, we develop a hybrid solution framework that integrates a Linear Programming (LP) formulation with the Non-dominated Sorting Genetic Algorithm II (NSGA-II). The LP model optimally determines the time-window assignments at the leader level, whereas NSGA-II is employed to efficiently solve the daily follower routing problems. NSGA-II defines a projection of patients onto a plane, which is used to generate one giant tour per day. Each tour is then subsequently split into feasible routes using the Beasley splitting algorithm. The proposed framework is evaluated on synthetic instances of various sizes.

ABSTRACT. Despite emerging concurrently with exact methods, metaheuristics research has never found a comparable theoretical foundation. This absence has resulted in a "horse race" mentality, where algorithmic performance is prioritized over the generation of robust scientific knowledge. Consequently, while the field has produced highly effective optimization tools for large-scale applications, it suffers from a scarcity of verifiable scientific insights. We contend that the advancement of metaheuristics research requires a transition toward a rigorous empirical paradigm. Our recent meta-analysis of adaptive large neighborhood search (ALNS) indicates that many established "truths," such as the purported benefits of adaptiveness, rely only on anecdotal evidence and demonstrate significantly less impact than traditionally assumed. To establish metaheuristics as a mature scientific discipline, the community must adopt standardized experimental protocols to ensure comparability, open data repositories to facilitate transparency, and modular algorithmic components to isolate specific mechanisms. Ultimately, the field must undergo a cultural shift, prioritizing the search for robust and reproducible scientific understanding over incremental performance gains.

ABSTRACT. The Automated Guided Vehicle (AGV) Scheduling Problem with Battery Constraints (ASP-BC) involves scheduling a set of jobs with given processing times and battery consumptions on identical AGVs, while minimizing the makespan.

As a solution method, we propose a Multi-Neighborhood Search (MNS) composed of the Move, Swap, and 2-Opt neighborhoods, capable of handling two variants from the literature that assume, respectively, fixed and flexible recharge policies. It uses an implicit solution representation and a fast-forward greedy decoder.

Preliminary results show the effectiveness and speed of our MNS, which can find nearly optimal solutions on benchmark instances, achieving, for the two variants, average gaps of 1.23% and 0.34% from the best-known solutions within 10 seconds, a significantly shorter computing time than the ones reported in the literature.

ABSTRACT. The Train Dispatching Problem (TDP) involves rescheduling and rerouting delayed trains on a railway network, while minimizing delay-related costs. In this work, we study the TDP introduced for the DISPLIB2025 Computational Competition. We propose a metaheuristic approach based on a Multi-Neighborhood Search embedded within a Simulated Annealing framework. The method is designed with problem-specific search components, such as greedy constructive heuristics and tailored neighborhood structures.

The proposed search method was tested on the full DISPLIB2025 benchmark set of 112 instances derived from real-world applications. It successfully found feasible solutions for 111 instances and achieved 19 best known solutions. The method ranked seventh in the competition ranking. Although it was outperformed by specialized methods on some instance classes, it achieved competitive performance on other classes. These results demonstrate that a carefully designed metaheuristic can achieve feasible solutions and, in some cases, high-quality ones, for large-scale train dispatching instances.

ABSTRACT. Energy-aware scheduling has become a central challenge in modern manufacturing environments. As industries increasingly aim to reduce operational costs and carbon emissions, aligning production activities with electricity tariffs is increasingly important. The integration of battery energy storage system (BESS) further enhances the potential to reduce energy purchase costs; however, incorporating such systems introduces complex interdependencies between job scheduling and BESS management decisions, significantly complicating the problem structure. In this work, we study the problem of minimizing the energy cost of executing a set of jobs on a single machine within a fixed time horizon, where electricity prices follow a Time-of-Use (TOU) tariff. In addition, we consider the presence of a BESS that can be charged from the grid and discharged during high-price periods to reduce overall energy costs. To efficiently address this problem, we propose and analyze two matheuristic algorithm variants designed to effectively coordinate production scheduling and BESS usage decisions.

ABSTRACT. This paper investigates whether problem-agnostic decomposition can effectively compete with problem-specific strategies in integrated lot-sizing and scheduling. We propose a structured adaptation of the generic Construct, Merge, Solve and Adapt (Gen-CMSA-BIP) matheuristic for Mixed-Integer Linear Programming (MILP), explicitly separating binary and continuous decision spaces and introducing a slack-based feasibility mechanism to cope with rigid capacity constraints. This modification replaces traditional constraint propagation mechanisms from purely binary domains, enabling the framework to handle large-scale capacitated formulations without embedding problem-specific logic. The resulting approach is benchmarked against the tailored Relax-and-Fix-and-Optimize (RFO) decomposition framework. Computational experiments reveal a consistent and topology-dependent performance pattern. In the GLSP-PL, characterized by localized discrete setup states, the adapted Gen-CMSA-BIP systematically outperformed RFO by efficiently optimizing restricted capacity configurations. Conversely, in the CLSP-HFS environment, where sequence feasibility induces strong topological interdependencies, RFO dominated, as agnostic sampling struggled to preserve sequence connectivity when rounding relative precedence variables. These results provide empirical evidence that the effectiveness of decomposition strategies is strongly conditioned by formulation topology: while problem-agnostic sampling is highly competitive in localized-state models, precedence-based sequence structures require structural guidance from problem-specific decompositions. This insight motivates the development of hybrid matheuristics that explicitly account for formulation topology.

ABSTRACT. Automated Guided Vehicles (AGVs) are essential components of modern automated manufacturing and logistics systems. Efficient scheduling of AGV operations is critical for minimizing delays and maximizing system throughput, especially when battery limitations and charging requirements are considered. This work focuses on AGV scheduling with flexible charging operations, where vehicles may perform partial recharges and charging activities involve setup times. The problem is formulated as a mixed-integer linear program, and a logic-based Benders decomposition algorithm is used to efficiently solve large-scale instances. By explicitly integrating energy constraints and realistic charging processes, the approach improves the practicality and performance of AGV scheduling in complex Industry 4.0 environments.

ABSTRACT. Organizations increasingly recognize that effective shiftwork scheduling must balance operational needs with employee health and well-being. These considerations can contribute with the prevention of future negative health effects on workers due to shift work. Accordingly, this research presents a bi-objective NSGAII approach for shiftwork scheduling problem which in-corporates worker health and well-being conditions, including factors such as fatigue, workload pressure, and recovery times after night shifts. This method generates shift plans designed to maintain service levels, workload balance, while supporting sustainable working conditions following rec-ommendations around circadian rhythm conservation of workers. As a re-sult, competitive values for both objectives are obtained compared to exact approach method and improve computational times for large scale instances significantly. Thus, the proposed approach offers a practical, data-driven tool for organizations seeking to enhance productivity, reduce turnover, and promote a healthier work environment.

ABSTRACT. Many combinatorial optimization problems are naturally formulated as higher-order unconstrained binary models with cubic and higher-degree interactions. These are often reduced to QUBO to use existing solvers and quantum hardware, but such reductions add auxiliary variables and penalties that increase size and distort the energy landscape. We present a solver that operates directly on PUBO models, avoiding reductions, preserving model fidelity, and achieving superior performance, especially for higher-degree interactions.

ABSTRACT. Strategic Oscillation Strategies (S0S) in metaheuristic optimization dynamically alternate between exploration and exploitation to enhance search performance. By periodically adjusting parameters or search behaviors, these strategies help algorithms escape local optima and improve convergence. Variant algorithms of SOS are proposed for the Set Covering Problem (SCP) which is a well-known NP-hard optimization problem with wide applications and certain other well-known problems can be formulated as SCP's. Our results will be presented at the MIC2026 conference.

ABSTRACT. Ant Colony Optimization for Continuous Domains (ACO-CD) is a well-established approach for continuous optimization problems. ACO-CD employs a solution archive and explores the search space by sampling new solutions based on a Gaussian distribution. However, exclusive reliance on Gaussian sampling often results in slow convergence, requiring considerable computational effort to obtain high-quality solutions. To overcome this limitation, this study proposes a method that incorporates log-normal distribution sampling guided by information entropy, which reflects the current state of the search process. Specifically, log-normal sampling is applied during the early search phase, characterized by high solution diversity, to focus the search around promising regions. In contrast, during later stages with low diversity, Gaussian sampling is employed to refine the search and enhance solution quality. Numerical experiments on benchmark functions demonstrate that the proposed method outperforms the conventional ACO-CD in most cases, achieving superior performance with reduced computation time and minimizes the number of evaluations functions.

ABSTRACT. We address an extended Agile Sprint Planning Problem in which user stories are assigned to sequential sprints under capacity and AND/OR precedence constraints while maximizing time-discounted business value. The model incorporates pairwise affinity gains and sprint activation costs, leading to a mixed-integer quadratic formulation. To ensure scalability, we develop a column generation approach where the pricing problem is reformulated as a QUBO, enabling the use of quantum-inspired heuristics within a temperature-guided hybrid scheme that combines quantum sampling, classical metaheuristics, and exact verification. The framework provides a structured integration of hybrid quantum classical optimization into decomposition-based release planning.

ABSTRACT. We address a multi-product inventory and routing problem involving river transportation by tugboats pulling barges, where exact models struggle to find quality solutions for long planning horizons. Two decomposition strategies are presented: vertical and horizontal. Vertical decomposition iteratively solves the problem using a rolling horizon, while horizontal decomposition separates decision layers and solves the routing layer heuristically. Experiments show that both strategies improve performance compared to the exact model, with horizontal decomposition excelling on the largest instances.

ABSTRACT. We propose Quantum-Reinforced Dual-Space Differential Evolution (QR-DSDE), a novel metaheuristic that synergistically integrates five complementary enhancement mechanisms: quantum tunnelling with landscape-adaptive strength, policy-gradient strategy selection, PCA-based dual-space coevolution, long-term memory archives, and hierarchical population stratification. While each mechanism has been explored independently in recent literature, their principled combination within a unified DE framework remains unexplored. Theoretical analysis establishes convergence guarantees while maintaining O(NTDE) complexity, matching classical DE. Comprehensive experiments on all 12 CEC~2022 single-objective benchmark functions (D=30, 30 independent runs) show that QR-DSDE achieves an average rank of 1.58, placing first or second on 10 of 12 functions, and outperforms nine state-of-the-art DE variants, including iLSHADE-RSP, jSO, L-SHADE, NL-SHADE-RSP, LSHADE-RSP, L-NTADE, SHADE, jDE, and PSO. Wilcoxon signed-rank tests confirm statistical superiority (p<0.05) over all competitors on at least 10 of 12 functions, with decisive margins on multimodal, hybrid, and composition landscapes.