With the increase of orders and riders, how to allocate orders to riders appropriately and complete delivery timely are huge challenges in on-demand food delivery (OFD). This work focuses on the multi-objective on-demand food delivery scheduling problems (OFDSP) with minimizing the total delay time of all orders and the total traveling distance of all riders simultaneously. First, both order-based mathematical model and a location-based mathematical model are established for the concerned problems, respectively Then, three basic algorithms (differential evolution (DE), particle swarm optimization (PSO), and artificial bee colony (ABC)) are improved to solve OFDSP. Combining order and rider information, a double string encoding method is designed to represent candidate solutions. A multi-constraint satisfying initialization strategy is proposed to avoid infeasible solutions. Seven problem-specific knowledge-based local search operators (KLSOs) are developed to speed up the convergence of the algorithms. Third, two reinforcement learning (RL) algorithms with two Q-table design strategies are constructed, yielding a total of four RL-based strategies for local search operator selection during iterations. Finally, 30 instances from real scenarios are solved to verify the performance of the proposed algorithms. The two developed mathematical models are verified and compared. The effectiveness of the improvement strategies is validated in the respective algorithm groups. Furthermore, the algorithms with the best performance in the respective groups are compared to the two advanced OFD algorithms. The results and comparisons show that the ABC_LS_Q1 is most competitive among all compared algorithms.

Ethylene serves as a cornerstone feedstock within the petrochemical industry, rendering the operational optimization of cracking furnaces imperative for maximizing yield and efficiency. Nevertheless, coking effects frequently impair heat transfer efficiency, demanding effective mitigation strategies. However, conventional optimizers often struggle to balance conflicting objectives within such complex landscapes, leading to suboptimal convergence. Therefore, this paper proposes the Nonlinear Search and Dynamic Leader Selection Integrated Multi-objective Optimization (NSDMO) algorithm. Established upon the Multi-objective Harris Hawks Optimization (MOHHO) framework, the NSDMO integrates a nonlinear escape energy mechanism with an exponential decay profile to balance early global exploration with late local convergence. Furthermore, an adaptive jump intensity strategy is introduced to facilitate the transition from aggressive searching to refining the local exploitation. Then, a pruning strategy based on comprehensive metrics combining convergence and diversity indicators is further utilized to enhance the distribution of the solution set. Extensive benchmarking against five prominent algorithms across the ZDT, the WFG, and the DTLZ suites substantiates the superior efficacy of the NSDMO. When applied in the optimization of an ethylene cracking furnace, the NSDMO increases the ethylene yield by 6.73% and simultaneously achieves a slight improvement of 0.047% in the propylene yield.

The flexible job shop scheduling problem (FJSP) traditionally emphasizes machine flexibility, while worker heterogeneity is often simplified or neglected. In practical manufacturing systems, worker skill levels affect both processing efficiency and task feasibility, leading to the skilled-worker constrained flexible job shop scheduling problem (SWC-FJSP). This article investigates the SWC-FJSP with multi-level worker skills and skill-dependent processing efficiency, aiming to minimize the makespan. A mixed-integer linear programming model is first formulated to characterize the coupled operation–machine–worker assignment decisions. Then, a branching deep Q-network enhanced memetic algorithm, termed MABDQN, is developed as a learning-assisted evolutionary framework. MABDQN integrates an adaptive hybrid initialization strategy to generate high-quality and diverse populations, a BDQN-driven operator orchestration mechanism to select operator category and execution intensity according to the evolving search state, and a multi-stage adaptive restart strategy guided by a stage adaptive reward function to alleviate premature convergence. Extensive experiments on a newly constructed test set with different instance scales show that MABDQN outperforms several state-of-the-art algorithms in solution quality and robustness. Ablation studies further verify the contribution of the main components and demonstrate the effectiveness of state conditioned operator orchestration for complex SWC-FJSP instances.

Addressing the dynamic flexible job shop scheduling (DFJSS) problem to generate a high-quality rescheduling scheme has gained increasing significance in various applications in recent years, such as the intelligent manufacturing system for aluminum profiles. In the intelligent manufacturing system for aluminum profiles, random events such as machine faults and variable processing times significantly affect production completion times. However, existing methods for solving DFJSS face challenges in generating high-quality rescheduling schemes in real time. To address these issues, a search strategy that combines graph neural networks (GNNs) and neighborhood search features is employed in the Monte Carlo tree search (MCTS) algorithm (GNS-MCTS) for real-time solving of the DFJSS. The search strategy in the GNS-MCTS aims to build a model that analyzes the scheduling disjunctive graph and predicts the searching probability, guiding the MCTS to important locations in the solution space rather than searching unimportant areas and wasting computing time. Specifically, the search strategy of GNS-MCTS incorporates encoders that fuse neighborhood search features and graph embedding vectors (GEVs) to improve the quality of the searching probability and make decisions about optimizing the scheduling disjunctive graph by breaking low-importance node pairs and reallocating low-importance nodes to other machines. Experimental results indicate that GNS-MCTS surpasses baseline algorithms across various problem sizes and real-time constraints, significantly enhancing computational efficiency and solution quality for the DFJSS. Further ablation studies and analysis reveal the impact of neighborhood search features on the enhancement of the learning-based search strategy when combined with neighborhood search for real-time DFJSS solutions.

Current cooperative multi-task assignment models do not adequately capture electronic signal-guided imaging reconnaissance, where UAVs can execute tasks from feasible stand-off positions rather than by reaching exact target coordinates. This mechanism changes the evaluation of route length, task timing, energy consumption, and task feasibility. Accordingly, a heterogeneous cooperative multi-UAV reconnaissance task assignment model is formulated with actual flight distance as the optimization objective and payload- and maneuver-dependent energy consumption explicitly considered. Meanwhile, EIR–EOR precedence, task time windows, and heterogeneous task-resource compatibility substantially narrow the feasible solution space. To address this constrained search structure, an adaptive dual-group variable neighborhood search algorithm is proposed, combining diversification- and exploitation-oriented neighborhoods with a timing conflict repair strategy that restores EIR–EOR temporal feasibility. Evaluations across eight instances demonstrate that ADGVNS outperforms five baseline algorithms in diversity and convergence, particularly in large-scale scenarios (instance 6), reducing actual flight distances by up to 13.07% and computational costs by up to 82.64% compared to baseline algorithms.

This work investigates the space-limited aircraft assembly scheduling problem (SAASP) based on real-world cases. A computational model, minimizing the makespan, is developed to formulate the complex operational relationships and space constraints of the workstations in SAASP. To address SAASP, we propose a reinforcement and statistical learning-assisted evolutionary algorithm (RSLEA). First, hierarchical encoding and queue decoding methods are designed to capture the intricate sequencing and workstation space constraints. Second, a statistical learning strategy is implemented to accelerate the convergence of early exploration. Next, a reinforcement learning strategy is introduced to control the sampling size of the statistical learning process. Additionally, another reinforcement evolutionary learning strategy is developed for exploitation. Three crossover operators are employed, with their computational resources adaptively allocated by the agent. To validate the performance of the proposed algorithm, RSLEA is tested on 520 instances and three real-world cases from a Chinese aircraft manufacturing factory. These cases are the assembly workstations of wings, partial bodies, and tails of aircraft, which are critical and complex. The RSLEA is compared with state-of-the-art algorithms and obtains better performance on 89% instances, and reduces the average gap to the baseline by 2.72%. Moreover, RSLEA achieves the best solution and overall performance in real-world cases.

Collaborative robots (cobots) are increasingly used to help human workers perform assembly tasks or complete assembly tasks themselves in assembly lines. The ergonomic risks of human workers are a key factor influencing assembly line efficiency. Therefore, this study investigates the mixed-product-model assembly line balancing and sequencing problem (ALBSP) with cobots, considering ergonomic risks in cases where human workers and cobots can operate different tasks in parallel. A mixed-integer programming model is formulated to optimize the makespan and ergonomic risks; this model can solve small-scale instances optimally using the CPLEX solver. A Q-learning-based multiobjective coevolutionary algorithm (QMOCEA) is then developed to handle large-scale instances. This algorithm adopts five vectors for encoding: the task assignment vector handles the task allocation subproblem, the worker allocation vector handles the worker allocation subproblem, the cobot allocation vector handles the cobot allocation subproblem, the process alternative selection vector handles the process alternative selection subproblem, and the product model sequencing vector handles the product model sequencing subproblem. Additionally, this algorithm uses knowledge-based decoding and initialization to obtain high-quality initial solutions. A parameter self-update strategy is proposed to adjust algorithm parameters dynamically. Comparative analysis demonstrates that the proposed method outperforms the original version and exhibits promising performance in comparison with benchmark methods, achieving the highest average hypervolume (HV) ratio of 0.805 and the lowest inverted generational distance (IGD) of 0.028 across 22 instance groups.

Band selection (BS) is a key method in hyperspectral image (HSI) classification that helps to reduce the computational burden and improve the class separability. However, with the emerging of uncrewed aerial vehicle (UAV)-borne HSI datasets, their attributes, such as high spatial and spectral resolution as well as large-scale samples pose serious challenges to the existing BS methods, making them inefficient. In addition, the efficient utilization of the prior knowledge from the data collected by fixed UAV-borne sensors in different regions is often easily overlooked. In view of these issues, this article proposes a neural network-assisted particle swarm optimization (PSO) algorithm for cross-domain BS of UAV-borne HSIs. First, a knowledge learning strategy is designed for the source domain, which applies a neural network model to learn the useful prior knowledge in labeled source domain data. Then, a network-assisted PSO algorithm is introduced to search for the optimal subset of bands in the target domain under the guidance of the valid prior knowledge captured from the source domain by the network model. Moreover, a similarity-based grouping strategy is designed to group similar bands and then select bands from each group with the aims of reducing the redundant information in the subset of bands. Finally, experimental results on three common UAV-borne HSI datasets show that our proposed method can efficiently handle UAV-borne HSI data with large samples, as it is able to find a subset of bands with higher quality compared to several state-of-the-art BS methods.

Mass casualty incidents caused by unpredictable and devastating disasters result in significant property loss and pose serious threats to human life. In such crises, the ethical priority of emergency responders is to save as many lives as possible, despite limited medical resources. Effective delivery of emergency medical services is crucial in managing life-threatening events. The severe shortage of medical support during these disasters requires each medical professional to treat multiple patients simultaneously. Additionally, professionals should also promote the overall effectiveness of treatment by preventing "tying" themselves to one patient. This research introduces a scheduling model that integrates multitasking into patient treatment plans. The model aims to minimize time-related objectives to ensure a rapid medical response in time-critical scenarios, allowing all patients to benefit from more efficient prioritization schemes. To proactively evaluate the proposed model and treatment plans without incurring actual risks, a comprehensive solution framework has been developed. This framework includes asymptotically optimal heuristics, a well-designed branch-and-bound algorithm for exact solutions, and an improved memetic algorithm. The results demonstrate the effectiveness and robustness of the framework across various problem scales. By analyzing the structures of various solutions, managers gain valuable insights that inform policy decisions and guide emergency response planning.

Multimodal freight transportation is essential for enhancing logistics efficiency and reducing costs, with route optimization as its core component. However, uncertainties are prevalent in multimodal systems, posing challenges for the construction and validation of simulation models. Furthermore, the involvement of multiple stakeholders introduces conflicting optimization objectives. Effectively balancing these objectives to maximize overall benefits has become a critical issue that requires resolution. For the multimodal transport route optimization problem in hybrid uncertainty environments (e.g., there are uncertainties in transportation demand, transportation time, and transfer time), this study constructs a multiobjective optimization model based on fuzzy numbers. The complexity of the model is reduced by introducing the chance-constrained programming theory. To solve the model, a data-driven multi-objective evolutionary algorithm is designed, integrating Monte Carlo simulation with surrogate models to effectively reduce the computational cost of uncertainty estimation. Furthermore, a constraint prioritization strategy is developed to handle multiple conflict objectives, and complex constraints efficiently. Simulation results demonstrate that the proposed algorithm exhibits excellent performance across networks of varying scales, providing robust decision support for multimodal transportation decision-making.

Despite significant efforts to manually design high-performance evolutionary algorithms, their adaptability remains limited due to the dynamic and ever-evolving nature of real-world problems. The "no free lunch" theorem highlights that no single algorithm performs optimally across all problems. While online adaptation methods have been proposed, they often suffer from inefficiency, weak convergence, and limited generalization on constrained optimization problems (COPs). To address these challenges, we introduce a novel framework for automated component configuration in Differential Evolution (DE) algorithm to address COPs, powered by Deep Reinforcement Learning (DRL). Specifically, we propose SuperDE, a foundation model that dynamically configures DE's evolutionary components based on real-time evolution. Trained offline through meta-learning across a wide variety of COPs, SuperDE is capable of recommending optimal per-generation configurations for unseen problems in a zero-shot manner. Utilizing a Double Deep Q-Network (DDQN), SuperDE adapts its configuration strategies in response to the evolving population states during optimization. Experimental results demonstrate that SuperDE significantly outperforms existing state-of-the-art algorithms on benchmark test suites, achieving superior generalization and optimization performance.

Dynamic flexible job shop scheduling (DFJSS) problem is an important scenario in intelligent manufacturing domain with the requirement of real-time decision-making under complex constraints. Existing approaches struggle to handle dynamic environments and heterogeneous job-resource relationships effectively while keeping optimum scheduling performance simultaneously. To address this challenge, a heterogeneous graph reinforcement learning framework combining question-aware neighborhood aggregation and interoption prompt attention (QIHGRL) is presented to address the DFJSS problem with new job insertions and variable processing times to minimize total tardiness. The optimization objectives are transformed into attention-guided signals by the question-aware neighborhood aggregation module to improve feature representation in the heterogeneous graph encoding stage of the QIHGRL. The competition and collaboration among scheduling actions are explicitly modeled by the interoption prompt attention layer in the policy optimization phase of the QIHGRL via adjusting action selection weights through a multihead attention mechanism to balance exploration and exploitation of the candidates in the population of the QIHGRL. The experimental results testified that the performance and efficiency of the QIHGRL outperforms that of the state of the arts algorithms.

With the rapid digital transformation of manufacturing industries in recent years, effective production scheduling has become imperative for maintaining competitiveness and addressing operational challenges. Modern manufacturing involves a complex interplay of elements, such as equipment, raw materials, and operators, making effective scheduling crucial for timely and specification-compliant production. However, achieving accurate and efficient scheduling remains a formidable challenge. Deep Reinforcement Learning (DRL), which integrates Deep Learning (DL) and Reinforcement Learning (RL), has emerged as a powerful tool for tackling these complex scheduling problems. A considerable number of related researches have been published in the past five years, yet there exists no review literature for them. Therefore, this paper reviews 167 studies published between Jan. 2019 and July 2024 that utilized DRL methods to solve production scheduling problems. We classify these studies based on types of scheduling problems they address and summarize their approaches to state representation, action space, reward mechanisms, and training algorithms. Additionally, we explore the integration of the DRL methods with optimization algorithms. Finally, we analyze the current research trends and future directions for the application of DRL in production scheduling from both problem-centric and method-centric perspectives.

With the advancement of intelligence in Active Distribution Networks (ADNs), effective fault recovery methods have become increasingly crucial. In this study, a reconfiguration method combining the immune mechanism and Northern Goshawk Optimization algorithm (NGO) is proposed, aimed at swiftly restoring power post-fault, maximizing the recovery of lost power areas within ADN, and minimizing losses. Firstly, an identification model within the immune mechanism is crafted to precisely match failures in ADNs. Then, the successful matched failure types can be used to restore power supply by the direct invocation of the recovery strategy from the library of strategies. Secondly, the immune response of ADNs is modeled, known as the reconfiguration model. For faults beyond the recovery strategy library, NGO is leveraged to address distribution network failures, with restoration solutions integrated into the library. Additionally, a reversed learning approach and stochastic variation strategy enhance the robustness of algorithm, preventing it from converging to suboptimal solutions. Finally, through simulation experiments, it is demonstrated that the recovery scheme obtained using the algorithm can be used to recover failure as well as reduce network losses in an effective manner. When similar or identical faults recur, ADN failure recovery becomes swift and efficient.

With the rapid advancements in artificial intelligence and control technologies in recent years, uncrewed systems have become increasingly prevalent across various fields. Path planning, a critical technology enabling autonomy in these systems, remains a challenging and active area of research. This review provides a comprehensive overview of the fundamentals of path planning and deep reinforcement learning (DRL), laying the foundation for understanding the potential and limitations of DRL in uncrewed system applications. It systematically reviews DRL methodologies and examines their applications across uncrewed aerial vehicles (UAVs), uncrewed ground vehicles (UGVs), uncrewed surface vehicles (USVs), and heterogeneous platforms, highlighting representative algorithms, real-world deployment scenarios, and diverse operational environments. The review also identifies major challenges in DRL-driven path planning, including real-time adaptability, robustness in complex environments, and scalability across domains. In addressing these challenges, it offers valuable insights and outlines future research directions, emphasizing the need for enhanced efficiency, safety, and generalization to meet the demands of next-generation autonomous systems.

Combinatorial optimization (CO) concerns decision-making over discrete solution spaces under constraints and lies at the core of many complex systems across science, engineering, and society. Yet current approaches to solving CO problems still rely heavily on expert-designed solvers and manual tuning. Large language models (LLMs), trained on massive corpora, offer a promising new avenue for enhancing CO solution methods with stronger capabilities for generalization and adaptation in complex or previously unseen scenarios. Some existing surveys provide high-level taxonomies without examining solver-level methods, while others focus narrowly on specific perspectives. To fill this gap, we present a methodology-oriented synthesis of recent LLM-assisted CO work, covering three mainstream solution paradigms: mathematical programming, heuristics, and end-to-end solvers. We first develop a hierarchical taxonomy that groups existing methods by solution paradigm and, within each paradigm, further classifies them by the workflow stage and the interaction pattern through which LLMs engage with the solver. We then formalize the workflow of each paradigm, clarify the functional role of LLMs, and analyze the core techniques in key modules. We further summarize the problem domains studied in the literature and discuss representative applications. Finally, we identify current limitations and outline future directions in architecture, methodology, and problem scope.

Mild cognitive impairment (MCI) functions as a critical intermediary stage between normal aging and Alzheimer’s disease (AD), highlighting the importance of accurate progression prediction for early intervention and disease management. Current research predominantly focuses on static cross-sectional imaging for binary classification, which fails to capture the temporal dynamics and longitudinal correlations essential for precise severity assessment and progression prediction. To address these limitations, this study proposes a dynamic time-series inference learning framework that models patients’ longitudinal disease trajectories as integrated training units. By utilizing structural magnetic resonance imaging (sMRI) data with flexible time intervals, the proposed network quantifies disease severity and predicts progression patterns in MCI patients over a 36-month period. Firstly, the dynamic time series inference learning module is proposed to explore complex correlations and dynamic characteristics within time-series data through continuous inference learning from known to unknown images. Secondly, a dual-track disease adaptation strategy is employed to simulate forward and reverse disease pathways for dynamic data augmentation, enhancing the robustness of small-sample and incomplete datasets. Finally, a joint task optimization strategy is introduced to handle various missing-data scenarios through three complementary learning tasks. Interpretability analysis and uncertainty quantification demonstrate that the model offers clinicians insights into MCI progression mechanisms while quantifying prediction confidence. Following extensive training, the proposed network attains exceptional accuracy relying solely on single-visit sMRI images. Moreover, the prediction accuracy shows a marked improvement as additional follow-up data are incorporated, ultimately achieving an impressive 90.7% accuracy and 87.1% area under the curve (AUC), demonstrating state-of-the-art performance.

Solving constrained multiobjective optimization problems (CMOPs) is highly challenging, especially when constraint functions are unknown or unquantifiable, yielding only binary outcomes (feasible or infeasible). Such cases significantly limit the performance of many advanced constrained multiobjective evolutionary algorithms (CMOEAs), particularly those relying on $\varepsilon $ -based methods. To address these challenges, this article proposes DRMCMO, a novel CMOEA framework based on the detection region method (DRM). In DRMCMO, detection regions dynamically monitor feasible solutions to enhance convergence. By using the existing feasible solutions as a reference, the detection regions are dynamically adjusted to help the population escape local optima. When the degree of constraint violation (CV) is unavailable, DRMCMO provides effective search guidance by constructing adaptive objective-space detection regions around discovered feasible solutions. We have modified three existing test suites to serve as benchmark problems for CMOPs with binary feasibility constraints and conducted comprehensive comparative experiments with the state-of-the-art algorithms on these test suites and real-world problems. The results demonstrate that DRMCMO is highly competitive against other state-of-the-art algorithms. Given the limited research on CMOPs with binary constraints (CMOP/BC), our study provides a new perspective for advancing this field.

Lot streaming, which combines lot splitting with overlapping operations, is used to reduce processing time and increase productivity in current manufacturing systems. This study focuses on the hybrid lot-streaming flow shop scheduling problem (HLFSP) and related subproblems, including lot planning and production scheduling. A mixed integer linear programming (MILP) model is proposed for the HLFSP to minimize the makespan of production processes. A critical path-based bi-population cooperative coevolutionary algorithm (CP-BCCEA) is proposed to simultaneously optimize sublot sizes, machine assignments, and sublot sequencing. The concept of critical jobs is defined by calculating the critical path of lot-streaming scheduling. The effectiveness and efficiency of CP-BCCEA are tested by experiments that compare it with state-of-the-art methods over a defined set of test instances.

To address the challenges of real-time performance, convergence efficiency, and reduced raw-data exposure in distributed online optimization, this paper proposes a distributed online sequential learning algorithm. Specifically, the algorithm integrates a discrete-time zero-gradient-sum (ZGS) consensus mechanism with online sequential learning, adapting to strongly connected weight-balanced directed networks. The algorithm consists of an initialization phase for parameter consensus via local training and ZGS iteration, and a sequential learning phase that leverages auxiliary variables and the Woodbury formula for incremental learning and global convergence. Furthermore, rigorous theoretical analysis based on algebraic graph theory and convex optimization proves that the node parameters of the algorithm converge exponentially to the corresponding batch-wise global optimum. Finally, the effectiveness of the proposed algorithm is demonstrated through experiments across six datasets, covering both online regression and time-series prediction tasks.