In agricultural production, pest detection plays a critical role. Traditional pest monitoring primarily relies on manual inspection, which is inefficient and highly subjective, making it inadequate for the requirements of modern precision agriculture. With the rapid advancement of deep learning techniques in image recognition and related fields, their application in pest detection has demonstrated significant potential. To address this demand, this paper proposes an efficient pest detection method based on the YOLOv10n deep learning model. To overcome the limitations of YOLOv10n, including insufficient multi-scale feature fusion, information bottlenecks, and weak local feature representation capability, this study introduces a re-parameterized general feature pyramid network (RepGFPN) and a spatial pyramid pooling enhanced local attention network (SPPELAN), further optimized by integrating the efficient channel attention (ECA) mechanism. The proposed model is trained and evaluated on the IP102 dataset, while the Pest24 dataset is employed to validate its generalization capability. Experimental results demonstrate that, with relative increases of 19.3
Intelligent agents have emerged as a fundamental paradigm for enabling autonomous, adaptive, and data-driven decision-making across increasingly heterogeneous computational environments. Despite substantial advances in learning algorithms, agent architectures, and deployment infrastructures, the literature remains fragmented across methodological, architectural, and application-specific perspectives. To address this fragmentation, this survey develops a unified taxonomy that systematically integrates learning paradigms, agent architectures, coordination mechanisms, deployment models, application domains, and evaluation frameworks from a common analytical perspective. The present work synthesizes recent developments in reinforcement learning (RL), federated learning (FL), neuro-symbolic reasoning, transfer learning (TL), evolutionary computation, and meta-learning, while examining their interactions with reactive, deliberative, hybrid, hierarchical, and cognitive agent architectures, as well as communication and coordination mechanisms in multi-agent systems (MAS). Particular attention is devoted to the relationship between learning mechanisms, deployment environments, and operational requirements in resource-constrained, dynamic, and large-scale systems. Beyond characterizing intelligent agent technologies, this survey provides a structured analysis of representative benchmarks, evaluation metrics, domain-specific performance indicators, and emerging challenges related to scalability, robustness, trustworthiness, interoperability, energy efficiency, and the development of embodied and large language model (LLM)-based agent systems. By linking methodological foundations with practical deployment considerations across domains such as robotics, healthcare, finance, cybersecurity, smart cities, education, and scientific computing, this study offers a comprehensive cross-domain synthesis of the intelligent agent landscape. The resulting framework facilitates systematic comparison, identifies recurring design trade-offs and research gaps, and provides a structured foundation for future advances in intelligent agent systems.
Large-scale multi-objective optimization problems (LSMOPs) typically involve hundreds or thousands of decision variables, posing significant challenges for evolutionary algorithms in balancing convergence efficiency and population diversity. To address these challenges, this paper proposes a dynamic Bayesian Gaussian mixture model based decision variable grouping evolutionary algorithm (TDGEA). First, a dynamic Bayesian Gaussian mixture model framework is employed to formulate the problem and perform preliminary grouping of decision variables, which categorizes variables into convergence-related variables and diversity-related variables. Then, a dynamic differential grouping strategy is introduced to further partition variables according to their interactions, ensuring strong intra-group coupling and weak inter-group coupling. Finally, an adaptive dual-population cooperative optimization strategy is adopted to independently optimize convergence-related and diversity-related variables. Experimental evaluations on standard benchmark test suites show that the proposed TDGEA consistently outperforms several state-of-the-art large-scale optimization algorithms with respect to convergence speed, optimization accuracy, and scalability.
Percussive massage motions impose stringent requirements on efficiency and joint-level impact suppression. To address this problem, a control-oriented multi-objective trajectory optimization framework is developed. Motion rhythm is regulated at the joint-trajectory level by jointly optimizing the spatial position and time allocation of intermediate peak points in quintic B-spline trajectories, enabling an explicit trade-off between efficiency and impact mitigation. For safety-sensitive percussive massage applications, a diversity-aware evolutionary search framework is developed. An EMA-guided diversity stabilization mechanism is employed to smooth adaptive parameter evolution, while adaptive crossover regulation and SOA-inspired local refinement are coordinated within NSGA-II to balance exploration and exploitation and improve Pareto-front distribution. This enables the generation of diverse time–impact trade-off trajectories that can accommodate different massage intensity and safety requirements. ZDT benchmark evaluations demonstrate significant improvements in solution diversity and convergence quality. Simulation studies conducted in CoppeliaSim further show that the optimized trajectories reduce joint impact indicators by 53.6
Chaotic Evolution Optimization (CEO) is a meta-heuristic optimization method inspired by the hyper-chaotic dynamics of two-dimensional discrete amnesia mapping. It enhances global exploration by simulating co-evolution among chaotic individuals and employing chaotic mapping to generate random search directions. Despite its competitiveness, the CEO algorithm has limitations in convergence accuracy, easy to fall into local optima, and premature convergence, particularly in complex problem domains. In this paper, a Multi-strategy Chaotic Evolution Optimization (MCEO) algorithm is proposed. It leverages four improvement strategies to mitigate these issues and synergistically boost the algorithm’s performance. First, a good point set initialization strategy generates a more uniform initial population for faster convergence into promising regions. Second, a novel nonlinear dynamic adjustment factor is designed to balance between global exploration and local exploitation, thereby improving convergence accuracy. Third, a spiral-guided random walk strategy replaces the original position update in the local exploitation phase, preventing search homogenization. Finally, a lens opposition-based learning strategy updates the current best solution after each iteration to reduce the risk of falling into local optima. This experimental work comprised two components. A parameter sensitivity analysis and ablation experiments for the MCEO algorithm were conducted on the IEEE CEC2020 benchmark. Subsequently, the algorithm’s performance in multidimensional optimization was assessed using the CEC2022 and CEC2017 test suites, benchmarking it against nine other peer algorithms. Results show that MCEO algorithm exhibits significantly better convergence accuracy and stability in most scenarios. Additionally, validation on three constrained engineering design problems and one wireless sensor network coverage problem confirms that MCEO algorithm outperforms the basic CEO and eight other algorithms in both optimization performance and practical applicability.
Swarm intelligence optimization algorithms exhibit excellent performance in handling single-objective optimization problems due to their bionic characteristics, but practical engineering applications often involve conflicting multi-objective optimization problems. To mitigate the limitations of the basic dung beetle optimization algorithm in balancing convergence rate, solution diversity, and distribution characteristics for multi-objective scenarios, this work presents an enhanced multi-objective dung beetle optimization algorithm. A dominance degree matrix mechanism is embedded to reduce the count of objective function comparisons in non-dominated sorting, while kernel density estimation crowding degree calculation method is integrated to sustain population diversity. To boost the algorithm’s global search performance, this study further develops an elite opposition-based learning strategy with adaptive learning probability adjustment, enabling the population to explore unexploited search spaces. Finally, comparative experiments are conducted between this algorithm and other algorithms under different multi-objective test functions. The simulation results show that the proposed algorithm has good performance under various indicators, which verifies the applicability and superiority of the algorithm proposed in this paper. Meanwhile, the algorithm developed in this work is implemented to tackle the multi-objective path planning optimization problem for mobile robot, and numerical simulation results verify that the method achieves sound performance in resolving practical engineering issues.
In many real-world scenarios, automated mobile robots are used to navigate through environments with rough terrain containing obstacles. These obstacles can be static as well as dynamic in nature. The robot should maintain a considerable safe distance from these obstacles. Moreover, the path taken by the robot must be smooth as well as energy efficient. This article proposes a Multi-objective Ant Colony Optimization (MOACO) based path planning algorithm called Efficient Multi-objective Ant Colony Optimization (EMOACO) for navigating mobile robots in environments with static and dynamic obstacles. The environments are modeled as grid-based. We consider four objectives to be optimized. These are– path length, safety, path smoothness, and overall elevation change. An ageing and ranking-based novel pheromone update technique is introduced in this paper that increases the diversity of solutions present in the Pareto front. For static environments (SE), a comparative study is performed among the proposed EMOACO-SE, NSGA-II, and A* algorithm. Another comparative study is carried out for the dynamic environment (DE) among the proposed EMOACO-DE and ABACO (Ageing Based Ant Colony Optimization) algorithm. These comparative studies establish the efficacy of the proposed method.
The dynamics of smoking cessation demonstrate long-term memory effects, which integer-order models cannot properly capture. This research addresses this gap by developing a Stochastic Levenberg-Marquardt Backpropagation Artificial Neural Network (SLMB-ANN) framework for solving a fractional-order quit smoking model (FOQSM) with a harmonic mean uptake function. The developed model divides the complete population into four disease transmission groups which include potential smokers P(t), occasional smokers L(t), chain smokers S(t) and quit smokers Q(t). The SLMB-ANN is trained on reference solutions generated by the Adam’s method for fractional orders α = 0.6, 0.7, 0.8, 0.9 . A comprehensive validation using error histograms, regression analysis ( R=1 ) and mean squared error ( MSE≈ 10^-9 ) demonstrates that the proposed method achieves absolute errors between 10^-5 and 10^-7 . The results confirm that the SLMB-ANN framework provides a stable, accurate and computationally efficient tool for analyzing non-linear fractional-order epidemiological systems and public health modeling.
Cancer gene data encompasses extensive information on gene expression, which can facilitate accurate diagnosis, personalized treatment and prognosis assessment of cancer. However, this type of data is characterized by high dimension, multiple text and multiple classification. Effectively reducing dimensionality and implementing precise prediction present significant challenges in feature selection (FS) for cancer gene expression data. To address these challenges, a binary sea-horse optimizer based on staged adjustment and improved breeding strategy is proposed. Initially, it puts forward a staged adjustment strategy that divides the iterative process into four distinct stages. The weight coefficient in the updating step is dynamically adjusted, thereby enhancing the algorithm's search capability, increasing the likelihood of escaping local optima and improving local convergence precision. Subsequently, a multi-species breeding strategy is proposed to increase population diversity and improve the algorithm's convergence speed. In the FS section, various binary algorithms were derived by selecting and exploring S-shaped, V-shaped, U-shaped and Z-shaped transfer functions. A total of 8 such algorithms were generated. The simulation experiment is divided into five parts. In the first two parts, the performance of the proposed strategy and algorithm is validated by using the CEC-2022 test functions. Experimental results demonstrate that SSHOB3 exhibits superior performance, offering advantages in average fitness, optimal fitness and convergence speed. For the last three parts of the experiment, the proposed method is first tested on 9 public UCI datasets, selecting the best binary variant V1SSHOB3. Then, by comparing with other binary algorithms, it is shown that V1SSHOB3 achieves lower average fitness values, higher average classification accuracy and fewer selected features. Finally, to verify the performance and superiority of the proposed algorithm in solving the FS of cancer gene expression data, V1SSHOB3 is extended to cancer gene expression datasets. The results indicate that V1SSHOB3 outperforms other binary swarm intelligence optimization algorithms in addressing the FS problem of cancer gene expression data.
The rapid expansion of cloud services has made selecting the right cloud service provider (CSP) increasingly complex. CSP services are evaluated based on various quality-of-service (QoS) attributes, including availability, best practices, compliance, durability, latency, reliability, response time, accessibility, and throughput. However, different CSPs may excel in different QoS attributes. Traditional multi-attribute decision-making (MADM) algorithms are used to rank CSPs but do not integrate QoS attributes across multiple CSPs. To address this limitation, a composite CSP can be created by combining the best QoS attributes from multiple CSPs, resulting in a unified, flexible cloud service. In this context, the optimization algorithm is crucial for selecting the best combination of QoS attributes across CSPs, thereby ensuring an efficient, high-performing composite CSP. This paper enhances the recent Rao algorithm and its three variants (Rao-1 to Rao-3) by proposing an enhanced Rao-based optimization (ERBO) algorithm with three variants (ERBO-1 to ERBO-3) for determining the composite CSP. The performance of the ERBO algorithm is compared with the existing genetic algorithm (GA) and the Rao algorithm to evaluate its effectiveness in constructing a composite CSP. Simulation experiments are conducted using the QoS for web services (QWS) dataset to assess the performance of both the existing and proposed algorithms. Additionally, a sensitivity analysis is performed to examine their robustness under varying weight conditions and analysis of variance (ANOVA) is used for statistical validation. The results demonstrate that the ERBO algorithm outperforms the GA and Rao algorithm in selecting the best-performing QoS attributes, leading to a more effective composite CSP.
This study investigates the application of quantum-inspired optimization to solve the NP-hard problem of feature selection in high-dimensional biological datasets. While classical machine learning approaches often struggle with the curse of dimensionality and redundant feature spaces in genomic modeling, we propose a Quantum Annealing-Based Gene Selection (QAGS) framework. The core of this framework is the formulation of the gene selection task as a Quadratic Unconstrained Binary Optimization (QUBO) problem, specifically designed to leverage the energy minimization properties of quantum annealing. The proposed architecture integrates a noise-removal preprocessing phase with a quantum-driven optimization core, followed by a hybrid deep neural network for phenotype prediction. By mapping genomic interactions into a Hamiltonian representation, the QAGS framework identifies compact feature subsets that yield reduced feature dimensionality and improved predictive efficiency within the evaluated setting. Experimental validation was conducted using large-scale crop genomic data exceeding 18,000 features. Results demonstrate that the quantum annealing-driven approach achieves a 94.6
This paper proposes a new single-solution optimization algorithm called Metaheuristic Exponential Search Optimization with Covariance Matrix Adaptation (MES-CM). The method was designed as a non-metaphor-based metaheuristic that uses an exponential pseudo-random sampling mechanism to generate search steps over multiple distance scales. This mechanism enables both short-range refinements and longer exploratory moves without relying on a population of candidate solutions. To improve directional search, MES-CM incorporates a covariance-guided transformation inspired by covariance matrix adaptation. In the proposed framework, however, the covariance information is estimated intermittently from previously accepted solutions and used within a single-solution search model. Therefore, the main contribution of MES-CM lies in the integration of exponential multi-scale sampling, accepted-solution-based covariance estimation, and a compact single-solution optimization structure. The performance of MES-CM was evaluated on the CEC 2017 and CEC 2022 benchmark suites, as well as on four constrained engineering design problems. The experimental study includes accuracy analysis, empirical convergence analysis, running time comparison, and non-parametric statistical tests based on the Wilcoxon signed-rank test and the Friedman mean rank test. The results show that MES-CM achieves highly competitive performance compared with recent metaheuristic algorithms from 2022 to 2026. On the CEC 2017 benchmark, MES-CM obtained the best mean objective value for 24 out of 30 functions and achieved the best overall Friedman mean rank. On the CEC 2022 benchmark, MES-CM achieved the best mean objective value for 7 out of 12 functions and again obtained the best overall Friedman ranking. The engineering design results further confirm that the proposed method can obtain high-quality solutions for constrained practical problems. The source code of MES-CM is publicly available at: https://github.com/grzegorzbies/MES-CM .
Balancing convergence and diversity while traversing infeasible regions remains a formidable challenge in constrained multi-objective optimization problems (CMOPs) with discrete or narrow feasible domains. However, existing multi-population methods typically employ static collaboration strategies, which sometimes struggle to adapt to the differentiated demands of search behaviors across different evolutionary stages. To address this issue, this paper proposes a two-stage multi-population constrained multi-objective evolutionary algorithm with a dynamic collaborative mechanism (TMDCEA). The optimization process of TMDCEA is partitioned into two distinct stages: global exploration and local exploitation. In the first stage (the global exploration stage), the algorithm establishes a weak coevolutionary mode among a main population and two auxiliary populations (a diversity-enhancing auxiliary population and an unconstrained auxiliary population). In the second stage (the local exploitation stage), the algorithm establishes a strong coevolutionary mode between the main population and the diversity-enhancing auxiliary population. Through this dynamic coevolutionary framework, the algorithm achieves an excellent balance among convergence, feasibility, and diversity. To drive the algorithm towards the constrained Pareto front while maintaining solution diversity, this paper employs a reference vector-based diversity-enhancing strategy in the environmental selection of the diversity-enhancing auxiliary population. Furthermore, a dynamic constraint-guided environmental selection strategy is proposed for the main population, which assists the population in progressively traversing vast infeasible regions by dynamically activating constraints step by step. Experimental results on three CMOP benchmark test suites and twelve real-world CMOPs demonstrate that, compared with ten state-of-the-art algorithms, TMDCEA exhibits competitive performance in handling complex constrained problems.
In Electronic Health Records (EHR), variables such as medications and diagnoses are recorded repeatedly over time, providing temporal information for clinical decision-making. Deep Learning (DL) has been applied to a variety of tasks using EHR time-series data, yet their characteristics pose major analytical challenges. This review examines studies that have applied DL to the analysis of multivariate time series (MTS) in EHRs, identifying prediction tasks, data types, and strategies used to address these challenges. Methods:Following PRISMA-ScR, we searched Scopus, Web of Science, PubMed, the ACM Digital Library, and IEEE Xplore, for studies published through December 2025. We included peer-reviewed English-language studies applying DL to MTS analysis of structured EHR data. Results:We included 182 articles analyzing key aspects of DL model development in MTS analysis of EHRs. The studies covered diverse tasks and data types across multiple databases, revealing common challenges addressed with shared strategies. MIMIC-III was the most commonly used database, particularly in studies focused on disease management and monitoring tasks involving laboratory results. Most studies addressed single-step classification tasks (e.g., mortality prediction), whereas relatively few addressed MTS forecasting. Attention mechanisms were the predominant interpretability approach, and post-hoc explainability methods were increasingly used. Conclusions:The task-oriented taxonomy indicates that predictive objectives influence the choice of data, model architecture, and explanation methods. Validation on public databases is essential for replicability, while external datasets are required to assess generalizability. Interpretability and explainability remain critical for clinical adoption, and clinician collaboration is essential to ensure clinical relevance.
This paper introduces the Cat Reflex Optimization (CRO) algorithm, a metaheuristic inspired by the multi-stage adaptive movement behavior of felids. The proposed algorithm addresses fundamental challenges in optimization including high dimensionality, non-convexity, and complex constraints through a biologically grounded framework that systematically balances global exploration and local exploitation. CRO is modeled on how cats navigate discontinuous environments by exploring multiple candidate positions, performing pre-transition evaluation, selecting the most promising targets, and adapting their future decisions based on previous outcomes. CRO incorporates three key innovations: a reflexive behavioral repertoire that concurrently generates diverse candidate solutions; a meta-controller architecture for intelligent candidate selection based on fitness improvement, population diversity, and proximity to the global best; and adaptive homeostasis mechanisms that preserve population vigor while mitigating premature convergence. Comprehensive experiments on the CEC2017 (30, 50, and 100 dimensions), CEC2019, and CEC2020 benchmark suites show that CRO performs competitively against ten state-of-the-art metaheuristics, with statistical significance assessed using the Wilcoxon rank-sum test and Friedman ranking. Additional evaluations on the CEC2014 and CEC2022 benchmark suites further support the effectiveness of the proposed algorithm. The applicability of CRO is also demonstrated on seven constrained engineering design problems, where it achieves competitive solution quality and stable performance across multiple statistical measures. These results demonstrate that CRO is an effective optimization framework for challenging continuous and constrained optimization problems.
This selection of environmentally sustainable suppliers is a critical and challenging task in green supply chain management due to the presence of uncertainty, subjectivity, and interaction among evaluation criteria. Multicriteria group decision-making methods are widely used for supplier selection; however, the reliability of experts’ judgments and handling of linguistic hesitation remain significant concerns. To address these issues, this study introduces fractional linguistic fuzzy sets, which provide a more flexible framework for representing experts’ assessments through linguistic membership and non-membership information. A novel score function and distance measure for fractional linguistic fuzzy numbers are developed to enhance discrimination among alternatives. Furthermore, several aggregation operators are proposed, including the fractional linguistic partitioned geometric Heronian mean, fractional linguistic fuzzy interactional weighted partitioned geometric Heronian mean, and fractional linguistic fuzzy interactional partitioned geometric Heronian mean operators, along with their fundamental properties. To improve ranking accuracy, a new grey relational analysis method is extended under the fractional fuzzy environment. The proposed approach is applied to a green supplier selection problem to demonstrate its practicality and effectiveness. Comparative analysis with existing methods confirms the robustness and reliability of the proposed decision-making framework.
Complex real-world decision-making problems are usually distinguished by uncertain, imprecise, and multi-source information, especially in environmental assessment cases such as air quality assessment. To deal with such issues it is important to have strong mathematical models that can incorporate interval uncertainty, expert hesitation, and multiple criteria in a single decision-making framework. This paper presents a cubic intuitionistic multi-fuzzy soft set (CIMFSS)-based uncertainty-aware decision framework. The proposed model effectively represents interval-valued membership and non-membership data while incorporating multiple expert judgments and parameterized uncertainty. Weighted arithmetic and geometric aggregation operators CIMFSWAA and CIMFSWGA are developed to facilitate multi-criteria decision-making, and their key mathematical properties (idempotency, boundedness, monotonicity) are rigorously established. A systematic decision-making algorithm based on entropy-based objective weighting and distance measures is presented. The proposed framework is illustrated through a multi-city air quality assessment case study. Comparative structural and numerical analyses with existing fuzzy soft set models demonstrate that the presented approach offers superior information representation and consistent ranking stability under uncertainty.
Transportation problems often involve uncertainty in costs, supply, and demand parameters. This study presents a double-parametric framework for solving fully fuzzy transportation problems (FFTPs) using triangular fuzzy numbers (TFNs). By combining the α -cut concept together with a secondary parameter β , the fuzzy transportation model is transformed into a family of parametric transportation problems. For each parameter pair (α ,β )∈ [0,1]× [0,1] , an initial basic feasible solution is obtained using the TOCM–MT method [1]. The proposed framework enables systematic analysis of transportation cost behaviour under uncertainty through a bounded spectrum of feasible solutions. For the primary benchmark problem, the transportation cost ranges from 234 to 1314, with a center value of 743. The framework was further validated using ten benchmark problems under three uncertainty spread levels (k=1,3,5) . The results show that increasing uncertainty systematically enlarges the feasible transportation cost interval, whereas the center transportation cost remains unchanged for each problem. In addition, sensitivity analysis based on transportation cost bounds demonstrates the influence of uncertainty on transportation cost variability across different spread levels. These findings show that the proposed framework provides a practical approach for analysing optimistic, intermediate, and pessimistic transportation scenarios in fully fuzzy transportation environments.
With the aggravation of energy problems and the promotion of sustainable manufacturing, the low-carbon concept has become increasingly prominent in workshop scheduling. As the saying goes, "When carbon is priced, the lowest-cost emission reduction measures will be implemented first." In this paper, the flexible job-shop scheduling problem with Type-2 Fuzzy Processing Time (T2FPT) is studied. At the same time, the maximum processing time, total carbon emissions, and machine load are considered. To address this problem, a Feedback Learning-based Evolution Algorithm (FLEA) was designed, incorporating five initialization strategies, four crossover and mutation operators. In addition, Q-learning is integrated as a feedback mechanism to dynamically adjust operator involvement during the evolutionary process, supported by a population state observation metric. Meanwhile interpolation method effectively reduces the carbon emission in the scheduling process. At the end of the algorithm, five neighborhood structures are designed for specific problems. Extensive experiments were conducted to evaluate the algorithm’s performance, and the results demonstrated that this method can effectively reduce the processing time and carbon emissions in the production process, showing great potential for workshop scheduling.
Decarbonizing transportation fleets is a multifaceted challenge that involves optimizing long-term fleet composition while carefully balancing three competing objectives: minimizing operational costs, lowering carbon emissions, and consistently fulfilling annual transportation demand. Most existing methods address these goals in isolation, limiting their practical applicability. This study proposes a hybrid evolutionary optimization framework inspired by NSGA-II, which integrates TOPSIS-based preference modeling to generate fleet plans that are cost-efficient, low-emission, and operationally viable across a multi-year planning horizon. This hybrid model leverages TOPSIS scores during both initialization and fitness evaluation phases of the evolutionary optimization (EO), enabling the generation of Pareto-optimal fleet compositions that minimize cost and emissions while satisfying year-wise demand constraints. The framework also incorporates an iterative planning mechanism that reuses viable fleet assets across years, reducing redundant acquisitions and promoting capital efficiency. Compared to the standard EO framework, the proposed method achieved a 28.37 https://github.com/yashyaks/FleetStructureOptimization .
