Balancing convergence and diversity remains a key challenge in multi-objective optimization. This paper proposes SOM-MSMOEA, a novel multi-stage evolutionary algorithm that uses a self-organizing map (SOM) to guide the search process. The evolution is divided into three stages: early, middle, and late. In the early stage, a SOM-based uniformity enhancement strategy improves population spread to avoid premature concentration. In the middle stage, an elite-led global search combined with SOM-guided local search in the convergence subspace accelerates front approximation. In the late stage, SOM-driven local search in the diversity subspace refines solution distribution. SOM-MSMOEA exploits the topological structure of SOM to guide stage-specific search strategies, effectively coordinating convergence and diversity throughout the optimization process. Experimental results demonstrate its competitive performance across a wide range of problems.
This work focuses on the insertion and evaluation of a reinforcement learning algorithm based on the Multi-Armed Bandit method within a multi-agent framework for combinatorial optimization. The algorithm was integrated into the cooperation structure of the framework to manage the pool used for sharing solutions among agents and to reduce the occurrence of inefficient steps during the cooperation process, guiding agents toward the selection of better solutions. To enable and support this integration, the framework’s cooperation structure and the diversity of agents were also analyzed. The cooperation structure intermediates the communication among agents, while diversity is ensured by implementing agents with different behaviors in the framework. The experiments were conducted using the Vehicle Routing Problem with Time Windows as the test problem. The results of computational experiments showed that the cooperation structure of the framework and the diversity of agents are of paramount importance, enhancing the quality of the objective function values and runtime achieved so far in tests carried out with the framework. Additionally, the results showed better convergence in the search structure of agents when the Multi-Armed Bandit Pool was incorporated into their cooperation structure, achieving improved runtimes and, in certain cases, enhancing the objective function values as this method prevented agents from exploring inappropriate spaces during the search process.
This article presents the Energy-Regularized Bi-Objective Maximal Covering Location Problem (ERBOMCLP), an extension of the classical Maximal Covering Location Problem (MCLP) that incorporates spatial dispersion among selected facilities. The classical MCLP maximizes covered demand under a fixed facility budget and service radius, but may place facilities in spatial clusters when high-demand areas are concentrated. ERBOMCLP addresses this limitation by augmenting coverage maximization with a Riesz s-energy objective that penalizes short inter-facility distances. To solve the model, we develop a memetic Non-dominated Sorting Genetic Algorithm II (NSGA-II) with binary fixed-cardinality encoding, problem-specific crossover and mutation operators, coverage-oriented local search, archive management, and a set-convergence-based termination criterion for detecting approximation-boundary stabilization. Experiments on ERBOMCLP instances derived from a real-world MCLP dataset show that the main memetic variant improves selected-solution coverage and often gives competitive or lower Riesz s-energy, while the spread of the Pareto front approximation remains instance dependent. A controlled hypervolume experiment with a Riesz-nonincreasing local-search rule further improves standard dominated hypervolume on the tested instances, at the cost of substantially more objective evaluations. The study connects potential-theoretic dispersion modeling with facility-location decision support for settings where high demand coverage is primary and dispersion is a complementary proxy for geographic balance.
Travel short videos have become a major channel through which destinations are discovered, evaluated, and socially amplified. Predicting whether a video will become popular is very challenging since the video popularity is jointly determined by content quality, creator-level exposure, and the long-tailed dynamics of online interaction. This paper first develops a comprehensive popularity score from likes, comments, favorites, and shares, where principal component analysis determines data-driven indicator weights and the 80th percentile defines a dynamic popularity boundary. To model tourism-specific content, visual, audio, and textual representations are extracted using TimeSformer, Whisper, and a pre-trained language encoder, respectively. A bottleneck multimodal transformer is developed to constrain cross-modal communication through learnable bottleneck tokens, reducing redundant interactions among scenery frames, background audio, and textual metadata. Finally, a video category-aware mixture-of-experts predictor combines fused content features with traffic metadata to model heterogeneous diffusion mechanisms across tourism themes. Experiments on collected 2,274 real-world travel short videos from TikTok show that the proposed framework achieves 89.45
The island model is a common framework for parallel and distributed evolutionary algorithms. This framework that is closely connected to memetic computing, since it studies how multiple search processes with different operators exchange information, preserve diversity, and collectively improve candidate solutions. In this model, we have two or more islands, and every island contains a different population that evolves semi-isolated from the others, exchanging solutions among them periodically through a migration operator. This work performs a robust and extensive evaluation of this model from a memetic-computing perspective, emphasizing cooperation, heterogeneity, and knowledge transfer among search components. We evaluate implementations ranging from two to fifty islands. Furthermore, we evaluate this model using five different evolutionary algorithms: the genetic algorithm, the ant colony optimization, the particle swarm optimization, the differential evolution, and the CLONAL algorithm, where each island can be evolved by any of these five algorithms, allowing for homogeneous and heterogeneous configurations. This evaluation is carried out using the 2015 IEEE Congress of Evolutionary competition on learning-based real-parameter single objective optimization as test functions. We optimized 49 different co-evolutionary island models using the five chosen evolutionary algorithms, whereas each co-evolutionary algorithm was developed using a different number of islands. A statistical analysis of our experimental data demonstrated that the use of several islands is beneficial to the results of the evolutionary algorithms. Indeed, we were able to show that increasing the number of islands generally leads to solutions closer to the optimum, especially when islands combine complementary search behaviors.
Neural architecture search (NAS) has demonstrated competitive performance on various application domains, including image classification, object detection and semantic segmentation. However, many existing methods either ignore resource constraints or rely on manually designed penalties and post-hoc pruning, resulting in suboptimal performance on resource-constrained devices. In particular, static constraint handling may prematurely discard promising but over-parameterized candidates, whereas decoupled post-hoc pruning often incurs additional search cost and may undermine the performance of the searched architectures. To address these issues, we propose PCAFP-ENAS, a Progressive Constraint and Adaptive Filter Pruning framework for efficient evolutionary NAS. PCAFP-ENAS adopts a MobileNetV3-inspired lightweight search space with variable-length encoding, enabling joint optimization of network depth and width. A progressive constraint mechanism imposes time-dependent penalties on model size and FLOPs with a length-aware correction term that preserves complex candidates early and strengthens constraint pressure as evolution proceeds. Moreover, an adaptive filter pruning operator is embedded into the evolutionary loop and performs L1-norm–guided one-shot structured pruning on selected high-potential architectures, yielding compact architectures without costly iterative fine-tuning. Experiments on CIFAR-10/100 show that PCAFP-ENAS achieves competitive or superior accuracy with fewer Params and lower FLOPs than handcrafted and NAS baselines, while requiring only a few GPU days. Results on a real-world osteoporosis X-ray dataset further demonstrate competitive accuracy with smaller models, highlighting the potential of the proposed framework for resource-constrained devices.
This paper proposes a surrogate-assisted evolutionary algorithm (SAEA) for expensive, high-dimensional problems, named RDDSP-SAEA (Reduced-Dimension Decision Space Partition−SAEA). The proposed algorithm is based on the DSP-SAEA framework and expands on the original method by incorporating a Multi-Decision Space Partitioning (MDSP) approach for global search and an Enhanced Surrogate Pool (ESP) for local search. In the local search, the ESP strategy integrates Principal Component Analysis (PCA) and Partial Least Squares (PLS) regression techniques to generate more reliable surrogates in lower-dimensional spaces. In the MDSP global search, the original decision space and additional lower-dimensional projections are partitioned and clustered. Radial Basis Function (RBF) surrogates are trained on the various regions within these decision spaces, and only the projection that yields the best models is selected for further global exploration. The algorithm is evaluated on test problems ranging from 30 to 1000 variables and compared against classical and recent competitive algorithms, as well as on a real-world engineering problem. The main contribution of this study is the proposal of an improved DSP-SAEA algorithm for optimizing expensive problems, which outperforms other state-of-the-art SAEAs, particularly for test cases with 300, 500, and 1000 variables.
Deep convolutional neural networks (CNNs) are highly effective for a wide range of computer vision tasks, especially image classification. However, creating an optimal CNN architecture is a complex, labor-intensive process that demands substantial manual effort and expert domain knowledge. Neural Architecture Search (NAS) automates this task, with evolutionary algorithms (EAs) offering strong potential but often struggling in multimodal search spaces. Neural Architecture Search (NAS) aims to automate this process, with evolutionary algorithms (EAs) showing strong potential; however, conventional EAs often struggle in multimodal search spaces due to limited exploration–exploitation balance. The paper introduces a new memetic differential evolution (DE) algorithm with a niching strategy and a Gaussian distribution-based local search to effectively explore and exploit the CNN architectural search space for image classification tasks, named MDE/NS. The proposed method is evaluated on standard image benchmarks, including MNIST and its variants, Fashion-MNIST, and SVHN, and compared with competitive CNN models, with additional validation on NAS-Bench-101 and NAS-Bench-201 confirming its effectiveness under standardized NAS benchmarks. A case study on COVID-19 chest X-ray images further demonstrates the adaptability of the evolved architectures to real-world classification tasks.
This study addresses the path planning challenges for multiple UAVs engaged in cooperative operations and multi-task point coverage by introducing a multi-objective optimization model, referred to as the UAV path planning model (UAV-PPM). The model’s objective is to minimize path length, suppress altitude fluctuations, enhance trajectory smoothness, and reduce flight risk in complex environments. Traditional optimization methods often exhibit insufficient convergence and limited solution diversity under high-dimensional, multi-objective, and dynamic conditions. To overcome these limitations, we have developed a multi-objective intelligent optimization algorithm, the wise wayfinding algorithm (WWA), which integrates mechanisms from non-dominated sorting genetic algorithm II (NSGA-II) and multi-objective particle swarm optimization (MOPSO). From a memetic computing perspective, WWA incorporates three cooperative strategies: (1) the oscillation diversity strategy, which employs difference operators and Lévy flights to facilitate dynamic propagation and local refinement; (2) the adaptive compensation strategy, which fuses multi-source information with weight adjustment to enhance individual learning; and (3) the jet convergence strategy, which balances local convergence with global exploration while incorporating non-elite information to maintain population diversity. Experimental results demonstrate that WWA exhibits favorable convergence and robust solution distribution on standard benchmark functions (ZDT, DTLZ, UF). When applied to the UAV-PPM problem, WWA outperforms NSGA-II, MOPSO, NSGA-III, and multi-objective rime-ice (MORIME), achieving improvements of 22.96
Integrating trucks and drones into package delivery offers a promising path toward a future logistics system that is more efficient and sustainable than current methods. However, coordinating trucks and drones under uncertain traffic conditions, particularly variable travel times, remains a critical challenge, with limited research addressing this dynamic variant. The most challenging part of the problem is to make real-time decisions (i.e., whether to accept the newly arrived service requests or not) during the execution of the routes. To address this gap, this study proposes a novel genetic programming algorithm that is Hybrid Crossover Genetic Programming featuring two distinct crossover operators to tackle the dynamic vehicle routing problem with drone and time window constraints. The algorithm dynamically adjusts schedules and routes by incorporating real-time data, enabling it to adapt to fluctuating traffic conditions and delivery demands. Whenever a new request arrives, it tries to re-generate new routes to include the new request by the heuristic. It accepts the new request if successful and rejects otherwise. Extensive experiments were conducted to evaluate the algorithm’s performance against the existing methods, demonstrating its ability to find high-quality solutions efficiently. Experiments demonstrate that the proposed algorithm is highly effective, achieving fast convergence and consistently outperforming the existing methods.
The ε constraint method is an important technique for handling constraints in constrained optimization problems. The proper selection of ε is highly dependent on the population distribution in each generation, making it exceedingly difficult to derive an analytical expression for ε . Consequently, this nonlinear dependency inevitably compromises the performance of conventional ε -constraint methods. To tackle this issue, we incorporate the concept of variable universes of discourse into a fuzzy logic controller. In this paper, we propose a variable-universe fuzzy logic controller (C-VUFC) to adaptively determine ε . By exploiting fuzzy if-then rules and membership functions, the proposed controller explicitly models the inherently nonlinear relationship between ε and the population distribution. Furthermore, we develop a domain adaptation mechanism that dynamically adjusts the universes of discourse based on historical constraint violation trends, thereby improving the adaptability of ε selection. Therefore, the proposed approach outperforms existing methods. Two novel algorithms are designed by integrating the proposed controller. The proposed algorithms are compared with state-of-the-art algorithms and the results demonstrate their effectiveness.
With the rise of social media, multimodal sentiment analysis has gained increasing attention due to the rich and diverse ways users express emotions through text and images. However, existing methods often ignore fine-grained sentiment cues and semantic misalignment across modalities, limiting this task effectiveness. To address these challenges, we propose DualScope, a novel model that combines a global-local fusion strategy with bidirectional image-text generation for semantically consistent data augmentation. Furthermore, we introduce both label contrastive learning and data contrastive learning to align heterogeneous modalities and enhance model robustness. The fusion module integrates global and local features in a progressive manner under global semantic guidance. Extensive experiments on two widely used datasets, MVSA-Single and MVSA-Multiple, demonstrate that the proposed method achieves superior performance, confirming its effectiveness in handling fine-grained semantics and cross-modal inconsistency.
Against the backdrop of rising global supply chain volatility, logistics planning in high-risk regions demands robust optimization tools. This paper introduces a novel hybrid memetic computing framework for the Threat-Aware Container Truck Routing Problem (CTRP). The framework integrates three synergistic components: (i) a flocking-inspired mechanism that maintains population diversity through separation, alignment, and cohesion behaviors for global exploration; (ii) an echolocation-inspired frequency-modulated search that dynamically adjusts step sizes for local exploitation; and (iii) a threat-aware evasion operator that functions as a domain-knowledge meme, proactively repelling solutions from hazardous zones via distance-weighted velocity adjustment. Unlike penalty-based methods that evaluate risk after solution generation, this operator operates ex ante, fundamentally altering the search trajectory toward safer regions of the solution space. The multi-objective CTRP model incorporates static spatial threats—including environmental hazards such as flooding and drought-prone areas alongside security risks—with operational cost, distance, and threat exposure as competing objectives. The framework's effectiveness is validated through extensive experiments on modified CVRP benchmarks, large-scale Set X (100–1000 customers) and ultra-large Set XL (1,327–10,000 customers) instances, and a real-world East African Community case study with 16 threat zones. Comparative analysis against five established metaheuristics (ALNS, HGA, HADAD, Hybrid Cell-Wave, MA-PSO) and the commercial solver Gurobi 11.0 demonstrates the framework's superiority across all scales. Statistical validation over 30 independent runs confirms robustness. The proposed framework establishes a validated algorithmic foundation for proactive threat-aware logistics, extensible to dynamic and probabilistic threat environments.
Within the Industry 4.0 paradigm, manufacturers encounter mounting pressures to deliver personalized products while accelerating supply chain responsiveness. This necessitates dynamic adaptation to rapidly evolving production schedules to ensure product quality without compromising manufacturing efficiency. Unpredictable dynamic disruptions within production environments, however, significantly compromise scheduling stability and overall productivity. To solve these challenges with a focus on makespan minimization, this study proposes a Deep Reinforcement Learning (DRL) framework for solving the Dynamic Flexible Job Shop Scheduling Problem (DFJSP) under stochastic job insertions and machine breakdowns. By formulating the DFJSP as a Markov Decision Process (MDP), the DRL agent dynamically determines optimal actions at each scheduling point based on real-time system states, assigning pending operations to available machines. To effectively encapsulate complex shop-floor dynamics, a Heterogeneous Graph Transformer (HGT)-based representation approach extracts and augments salient features of jobs and machines, enabling the agent to make end-to-end decisions. Furthermore, a tailored reward function enhances learning efficiency and solution quality, while a softmax action selection strategy effectively balances the exploration–exploitation (EE) trade-off. Notably, the Proximal Policy Optimization (PPO) algorithm is employed for agent training, greatly enhancing the efficacy of policy optimization. Simulation-based experimental evaluation demonstrates that the proposed HGT-based Reinforcement Learning (HGTRL) methodology effectively solves the DFJSP. Comparative analysis confirms that HGTRL achieves superior scheduling performance: For small-scale instances, HGTRL maintains an average gap of less than 16
The rapid advancement of generative AI (GenAI) has introduced transformative methodologies to the financial sector, enabling the creation of novel data and solutions to longstanding challenges such as data scarcity, privacy, and domain adaptation. Despite the proliferation of GenAI research in finance, there is a lack of comprehensive surveys that systematically review the core generative techniques and the unique research problems posed by different financial data modalities. This paper addresses this gap by providing an in-depth overview of foundational GenAI methodologies, including Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Normalizing Flow, Diffusion Models, and Large Language Models (LLMs), as well as their adaptations for financial tasks. We categorize and analyze research challenges according to data modalities: textual, time series, tabular, and graph data. For each modality, we introduce representative tasks such as classification, forecasting, question answering, and synthetic data generation, and further discuss current limitations and future research directions. This survey aims to serve as a technical reference for researchers and practitioners seeking to understand and advance GenAI techniques in the financial domain.
The evolution of financial technology has been marked by successive waves of innovation, from rule-based systems to the widespread adoption of machine learning methods, including deep learning approaches, which have delivered remarkable performance in a range of financial tasks. Recently, the emergence of generative AI (GenAI) has marked a new era in financial evolution, fundamentally reshaping the industry by enabling the creation of data, content, and solutions. Unlike conventional discriminative AI models that primarily focus on prediction and classification, generative models—such as Variational Autoencoders, Generative Adversarial Networks, Normalizing Flow, Diffusion Models, and Large Language Models—offer broader capabilities, enhanced user interaction, and improved interpretability. These advances are driving a new wave of innovative applications across diverse financial domains. Despite the growing literature on GenAI, there remains a lack of comprehensive surveys dedicated to its applications and sectoral impact in finance. This paper systematically reviews the deployment of GenAI techniques across key financial sectors. In particular, we highlight representative applications in securities, investment, banking, accounting, and regulatory compliance, illustrating how GenAI is transforming financial services and operations. Furthermore, we discuss the challenges associated with integrating GenAI into financial applications. This work aspires to be a reference for researchers and practitioners navigating the new era of generative AI in finance.
As modern manufacturing increasingly emphasizes production efficiency and flexibility, the scheduling of automated guided vehicles (AGVs) within matrix manufacturing workshops has emerged as a significant area of research. Effectively scheduling AGVs to fulfill production requirements while optimizing multi-objective performance is crucial for addressing production scheduling challenges. To tackle the complexities of AGV scheduling, the paper introduces the dual-criteria collaborative evolutionary algorithm (DCCEA), a sophisticated multi-objective evolutionary algorithm designed to optimize scheduling performance while meeting production demands. The DCCEA employs a dual-criteria heuristic to improve convergence rates and solution quality by prioritizing customers with higher fitness levels, thus starting the search from a robust initial population. It incorporates a multi-strategy co-evolutionary framework that features elite-oriented population selection, a bimodal adaptive crossover operator, and a dynamic window mutation strategy, which collectively enhance exploration in high-dimensional objective spaces while preserving diversity and minimizing premature convergence. Additionally, a local optimizer based on route load is introduced to refine the solution distribution and prevent local optima convergence. Experimental evaluations demonstrate that DCCEA outperforms other eight algorithms. Specifically, compared with the most competitive comparison algorithm, DCCEA achieves a 13.47
Benchmark test suites play a critical role in the evaluation of evolutionary algorithms. In bilevel single-objective optimization, however, benchmarks tailored to evolutionary computation remain limited. The widely used SMD test suite exhibits several shortcomings, including performance saturation, insufficient coverage of non-separable problems, and limited representational coverage. To address these issues, this paper proposes a new benchmark construction methodology that extends the original SMD design philosophy while alleviating its inherent limitations. The proposed approach aims to mitigate overfitting caused by performance saturation, enables scalable problem construction without relying on variable separability, and preserves the core characteristics of bilevel benchmarks while supporting the inclusion of new problem types that better reflect practical scenarios. Furthermore, a minimal implementation of the proposed methodology is provided, and its feasibility is demonstrated through empirical studies using several bilevel optimization algorithms. The source code is publicly available at https://gitee.com/jingyp/blopbench
Constrained multiobjective optimization problems (CMOPs) are prevalent in engineering and scientific applications, where the core difficulty is to simultaneously optimize multiple conflicting objectives subject to complex constraints. Existing multi-population co-evolutionary algorithms often suffer from imbalanced resource allocation and insufficient adaptability when coordinating feasible-region exploitation with exploratory search beyond the feasible boundary, which limits their overall effectiveness. To address these issues, this paper proposes an Information-Competitive Dual-Population Evolutionary Algorithm (ICDEA). The proposed framework establishes an information-competitive interaction mechanism between a primary population and an auxiliary population to enable adaptive allocation of search resources. The primary population performs intensive exploitation within the feasible region to promote convergence and constraint satisfaction, whereas the auxiliary population conducts broader exploration over infeasible and boundary-adjacent regions to maintain search coverage and support diversity preservation. In addition, a dynamic offspring allocation strategy, termed InfoGameAlloc, is designed to regulate inter-population resource distribution by jointly considering convergence tendencies, feasibility status, population diversity, distributional information volume, and improvement potential. Extensive experimental evaluations on 47 benchmark problems from the CF, DAS-CMOP, LIR-CMOP, and MW test suites and 12 real-world engineering problems, in comparison with nine competitive constrained multiobjective evolutionary algorithms, demonstrate that ICDEA achieves competitive and well-balanced overall performance in terms of convergence, solution distribution quality, and feasible solution ratio. These results indicate that the proposed framework provides an effective and adaptive way to coordinate feasible-region exploitation and exploratory search in constrained multiobjective optimization.
The influence maximization problem (IM) involves identifying K seed nodes in a given network capable of producing the maximum influence range. A great stream of literature has devoted to the construction of diffusion models and seed determination approaches. The majority of existing studies focus on networks with stable structures; in this manner, the impact of typological changes on the influence diffusion process remains to be investigated. Meanwhile, network systems are inevitably subject to perturbations or even structural damages during operation. Investigating the robustness of the influence diffusion process (RIM problem) holds substantial practical significance. Related studies indicate that the cascading failure may incur severe perturbances over the connectivity, but its impact on the attached diffusion process has not been touched upon. In order to solve the RIM problem under cascading failures. In this paper, we investigate the robust influence maximization (RIM) problem under cascading failures caused by link-based attacks. A numerical metric is designed for comprehensively assessing the robust influence performance of selected seeds. The failure model is also analyzed to give reasonable parameters in the destruction process. In addition, a Memetic algorithm with niches and population recombination strategy is designed to search for seeds with stable influence aiming at solving the RIM problem, named NMA-RIMCF-L. Finally, experiments on synthetic and authentic networks validate the competitiveness of NMA-RIMCF-L with respect to existing approaches.
