An enhanced multi-strategy educational competition optimizer for numerical optimization Authors: Zihan Li, Tianmin Zhang, Hongzhen Hu Published: 2026 ABSTRACT This study proposes an Enhanced Multi-Strategy Educational Competition Optimizer (EMSECO) to address the inherent limitations of the standard Educational Competition Optimizer, specifically its imbalanced exploration–exploitation dynamics, insufficient diversity preservation, and suboptimal convergence behavior. Three synergistic mechanisms are integrated. Firstly, an undergraduate education stage is embedded within the original cyclical framework, leveraging differential vector guidance to augment spatial coverage and sustain diversity. Secondly, a graduate education stage is introduced as an external intensification module, utilizing elite individual and elite subpopulation guidance to strengthen local exploitation. Thirdly, a random walk restart strategy is incorporated to perturb stagnated individuals, thereby suppressing premature convergence. Comparative experiments against AE, MRFO, SAO, EOSMA, ALSHADE, and EPSCA across multiple dimensionalities demonstrate that EMSECO achieves superior performance, attaining average Friedman rankings of 1.2, 1.4, 1.7, and 2.0 for 10-, 30-, 50-, and 100-dimensional CEC 2017 test functions, respectively. Practical applicability is further validated through seven real-world engineering problems. In summary, the proposed EMSECO algorithm exhibits enhanced global exploration capacity, improved local exploitation efficiency, and superior convergence stability. Introduction Optimization refers to the systematic process of identifying the most effective solution from a set of feasible alternatives for a given problem. Three fundamental components govern any optimization endeavor: the objective function (or fitness function), constraints, and decision variables1. The primary aim is to determine optimal or near-optimal values for these variables subject to specified constraints, thereby extremizing the objective function2. The rapid advancement of industrialization and escalating technical demands have substantially amplified the complexity of numerous engineering challenges. Contemporary optimization problems frequently exhibit nonlinear, non-differentiable, multimodal, and discontinuous characteristics, rendering them intractable through conventional analytical approaches3. While both classical optimization techniques and metaheuristic algorithms have been deployed to address such problems, traditional deterministic methods demonstrate considerable limitations: poor adaptability to problem structures, prohibitive computational overhead, pronounced susceptibility to premature convergence, and elevated risk of entrapment in local optima4. These inherent deficiencies have catalyzed growing research interest in metaheuristic methodologies. Metaheuristic algorithms typically integrate stochastic global sampling mechanisms with intensified local search procedures, harnessing the synergistic interplay between exploration and exploitation to navigate vast, structurally complex search landscapes in pursuit of optimal or satisficing solutions5. Exploration ensures comprehensive coverage of the search domain, safeguarding against premature stagnation in suboptimal regions, whereas exploitation concentrates computational resources on promising neighborhoods to accelerate convergence and refine solution quality. By circumventing dependence on rigorous mathematical formulations, these algorithms accommodate black-box and high-complexity systems, thereby enabling the treatment of high-dimensional, heavily constrained, non-convex problems that defy classical approaches. Although theoretical guarantees of global optimality are generally absent, metaheuristics routinely deliver high-quality solutions within acceptable computational budgets. In recent years, metaheuristic algorithms have achieved widespread adoption across diverse engineering and intelligent system applications—owing to their robust optimization capabilities, remarkable versatility, and straight forward implement ability. Representative domains include mission planning6,7, photovoltaic model identification8,9, coverage optimization in wireless sensor networks10,11, hyperparameter optimization12,13, medical image segmentation14,15, and feature selection16,17, among others. Metaheuristic algorithms are conventionally taxonomized into three principal categories based on their inspirational origins and operational mechanisms: evolutionary-based, physics-based, and swarm-based algorithms. Evolutionary-based algorithms, epitomized by Genetic Algorithm (GA)18, Differential Evolution (DE)19, draw inspiration from natural evolutionary processes. Subsequent research endeavors have predominantly focused on variant formulations and hybrid extensions of these foundational paradigms. Physics-based algorithms originated with Simulated Annealing (SA)20, which emulates the thermodynamic annealing process. Light Spectrum Optimizer (LSO)21, Stellar Oscillation Optimizer (SOO)22 are also belong to this category. This category has subsequently diversified into chemistry-inspired algorithms (e.g., Chemical Reaction Optimization (CRO)23 and Henry Gas Solubility Optimization (HGSO)24) and mathematics-inspired algorithms (e.g., Sine Cosine Algorithm (SCA)25, Bezier Curve-based Optimization (BCO)26 and Perpendicular Bisector Optimization Algorithm (PBOA)27), reflecting the broadening of metaphorical foundations beyond classical physics. Swarm-based algorithms trace their lineage to Particle Swarm Optimization (PSO)28 and Ant Colony Optimization (ACO)29. This branch has undergone substantial expansion, ramifying into animals-based algorithms (e.g., Tuna Swarm Optimization (TSO)30, Black-winged Kite Algorithm (BKA)31, Crayfish Optimization Algorithm (COA)32, Fishing Cat Optimizer (FCO)33 and Artificial Meerkat Algorithm (AMA)34), plants-inspired algorithms (e.g., Dandelion Optimizer (DO)35, Moss Growth Optimization (MGO)36 and Ivy Algorithm (IA)37), and human-based algorithms (e.g., Catch Fish Optimization Algorithm (CFOA)38, Rider Optimization Algorithm (ROA)39 and Human Evolutionary Optimization Algorithm (HEOA)40), mirroring the rich behavioral complexity observable across ecological and social systems. The classification of meta-heuristic algorithms is shown in Fig. 1.Fig. 1The classification of meta-heuristic algorithms. The classification of meta-heuristic algorithms. Notwithstanding the demonstrated versatility of these algorithms in addressing complex nonlinear and multimodal landscapes, their empirical performance exhibits marked problem-dependent variability. The No Free Lunch (NFL) theorem41 establishes that no single algorithm universally dominates across all problem classes. Indeed, individual algorithms harbor intrinsic limitations—including deficient exploratory capabilities, compromised convergence reliability toward global optima, and parameter sensitivity—which collectively constrain their efficacy. These deficiencies have motivated sustained research efforts directed toward both de novo algorithmic development and systematic enhancement of existing methodologies. The Educational Competition Optimizer (ECO) was introduced by Lian et al.42. in 2024 as a novel anthropogenic metaheuristic. Drawing inspiration from the competitive academic progression of students through primary to secondary educational stages, ECO distinguishes itself from conventional counterparts through its multi-phase search architecture, which confers enhanced adaptability across diverse optimization landscapes. Since its inception, ECO has been successfully deployed in several engineering applications, including photovoltaic system regulation43, battery parameter identification44, overcurrent relay coordination optimization45, and reservoir production optimization46, demonstrating its practical efficacy in addressing real-world constrained problems. Nevertheless, the canonical ECO formulation exhibits non-negligible deficiencies that constrain its broader applicability. Specifically, three critical challenges persist: rapid population diversity deterioration as iterations progress, pronounced susceptibility to premature convergence, and inherent difficulties in calibrating the exploration–exploitation equilibrium. These limitations collectively impair the algorithm’s capacity to navigate complex multimodal fitness landscapes and maintain sustained search vigor throughout the optimization trajectory. In response to these shortcomings, extensive research efforts have been directed toward algorithmic refinements aimed at bolstering ECO’s global exploration competence and problem-specific adaptability. Tang et al. incorporated a leapfrog strategy to fortify ECO’s global exploration capacity, coupled with an early termination mechanism designed to facilitate escape from local optima entrapment47. Li et al. introduced a Powell local search mechanism during the later evolutionary stages to intensify exploitation, synergistically integrated with a population regeneration protocol to counteract diversity loss48. Hammad et al. proposed a hybrid ECO variant embedding orthogonal learning strategies, specifically targeting the preservation of population diversity in terminal search phases49. Meanwhile, Emam et al. amalgamated a local escape operator with Gaussian perturbation tactics, thereby effectively augmenting both local evasion capabilities and exploitative refinement50. Notwithstanding these multidimensional enhancements to ECO documented in the literature, substantial scope for improvement persists with respect to three critical aspects: the dynamic equilibrium between global exploration and local exploitation, the efficacy of local optimum evasion, and the precision of terminal convergence. To address these residual deficiencies, this study proposes an Enhanced Multi-Strategy Educational Competition Optimizer (EMSECO). The main contributions of this study are summarized as follows:An Undergraduate Education Stage (UES) is embedded within the original search framework, leveraging diverse differential vector guidance to enrich population diversity and enhance global exploration capability.A Graduate Education Stage (GES) is designed as an external intensification module, utilizing elite individual and elite subpopulation guidance to curtail ineffective search and strengthen local exploitation capacity.A Random Walk Restart Strategy (RWS) is devised to prevent entrapment in local optima within complex search landscapes, thereby augmenting local optimum escape capability.Extensive comparative experiments conducted on the CEC 2017 benchmark suite and engineering design optimization problems validate that the proposed EM-SECO possesses competitive advantages in optimization performance and robustness. An Undergraduate Education Stage (UES) is embedded within the original search framework, leveraging diverse differential vector guidance to enrich population diversity and enhance global exploration capability. An Undergraduate Education Stage (UES) is embedded within the original search framework, leveraging diverse differential vector guidance to enrich population diversity and enhance global exploration capability. A Graduate Education Stage (GES) is designed as an external intensification module, utilizing elite individual and elite subpopulation guidance to curtail ineffective search and strengthen local exploitation capacity. A Graduate Education Stage (GES) is designed as an external intensification module, utilizing elite individual and elite subpopulation guidance to curtail ineffective search and strengthen local exploitation capacity. A Random Walk Restart Strategy (RWS) is devised to prevent entrapment in local optima within complex search landscapes, thereby augmenting local optimum escape capability. A Random Walk Restart Strategy (RWS) is devised to prevent entrapment in local optima within complex search landscapes, thereby augmenting local optimum escape capability. Extensive comparative experiments conducted on the CEC 2017 benchmark suite and engineering design optimization problems validate that the proposed EM-SECO possesses competitive advantages in optimization performance and robustness. Extensive comparative experiments conducted on the CEC 2017 benchmark suite and engineering design optimization problems validate that the proposed EM-SECO possesses competitive advantages in optimization performance and robustness. The structure of this paper is as follows: Sect. "Educational competition optimizer" briefly introduces the standard ECO. Section "Enhanced multi-strategy educational competition optimizer" details the Enhanced Multi-Strategy Educational Competition Optimizer (EMSECO). Section "Results and analysis of numerical experiments" present the simulation experiments and a detailed analysis of the results. Section "Engineering applications" applies EMSECO to solve six engineering design problems. Section "Conclusions" concludes the paper and provides an outlook for future work. Educational competition optimizer The Educational Competition Optimizer (ECO) draws upon the competitive mechanisms inherent in educational systems as its foundational metaphor. The optimization trajectory is structured analogously to three consecutive scholastic stages: primary education, secondary education, and high school education. Through adaptive search strategies, ECO achieves a seamless transition from global exploration to intensified exploitation. The mathematical formalism is summarized as follows. Population initialization Distinct from conventional metaheuristic paradigms, ECO employs Logistic chaotic mapping for population initialization. Assuming a population size of N, problem dimensionality of D, and bounded search space [lb, ub], the initial position of the i-th individual is computed via Eq. (1).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{0} = lb + \left( {ub - lb} \right) \times x_{i}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i}$$\end{document} denotes the value computed through Logistic chaotic mapping, obtained via Eq. (2).2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{i} = \alpha \times x_{i - 1} \times \left( {1 - x_{i - 1} } \right),0 \le x_{0} \le 1,i = 1,2,...,N$$\end{document}where α represents a control parameter fixed at 4 in this study, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{0}$$\end{document} denotes a stochastic initial value uniformly sampled from the interval (0, 1). Primary education stage In the primary education stage, the ECO algorithm bisects the population of N individuals into two distinct subsets. Following fitness-based ranking, the top-performing 20% constitute the school set, while the remaining 80% form the student set. It is noteworthy that this dichotomous partition is reconstructed at the commencement of each iteration. In this stage, students select geographically proximal schools, and schools adaptively adjust their positions to maximize enrollment rates. This process is mathematically formulated in Eqs. (3) and (4).3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} + \omega \times \left( {X_{imean}^{t} - X_{i}^{t} } \right) \times Levy\left( D \right)$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} + \omega \times \left( {X_{iclose}^{t} - X_{i}^{t} } \right) \times randn$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1}$$\end{document} denote the position of the i-th individual before and after update at iteration t, respectively. ω represents an iteration-dependent adaptive coefficient computed via Eq. (5). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{imean}^{t}$$\end{document} signifies the mean value across all dimensions of the i-th individual. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Levy\left( D \right)$$\end{document} denotes a D-dimensional random vector following Lévy distribution, generated through Eq. (6). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{iclose}^{t}$$\end{document} indicates the school individual closest to the i-th individual at iteration t. randn represents a random number drawn from standard normal distribution.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\omega = 0.1 \times \ln \left( {2 - t/T} \right)$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Levy\left( D \right) = \frac{\mu \times \sigma }{{\left| v \right|^{2/3} }},\mu \sim N\left( {0,D} \right),v \sim N\left( {0,D} \right),\sigma = \left( {\frac{{\Gamma \left( {2.5} \right) \times \sin \left( {0.75\pi } \right)}}{{1.5 \times \Gamma \left( {1.25} \right) \times 2^{0.25} }}} \right)^{2/3}$$\end{document}where T represents the predefined maximum iteration count. Secondary education stage In the secondary education stage, ECO maintains the bipartite population structure, repartitioning the entire swarm based on fitness rankings. The elite 10% constitute the school population, while the remainder form the student population. Schools update their positions through fusion of population centroid and global best individual; students persistently select their nearest school and are further stratified into two cohorts based on academic potential. These behavioral patterns are mathematically articulated in Eq. (7) through Eq. (8).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} + \left( {X_{best} - X_{mean}^{t} } \right) \times e^{{\left( {\frac{t}{T} - 1} \right)}} \times Levy\left( D \right)$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} - \omega \times X_{iclose}^{t} - P \times \left( {E \times \omega \times X_{iclose}^{t} - X_{i}^{t} } \right)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{best}$$\end{document} denotes the position of the best individual discovered thus far (i.e., the global best). \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{mean}^{t}$$\end{document} represents the centroid position of all individuals at generation t. P signifies the patience coefficient of the student population, formulated in Eq. (9). E indicates the learning motivation factor, defined in Eq. (10).9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P = 4 \times randn \times \left( {1 - t/T} \right)$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E = \left\{ {\begin{array}{*{20}l} {\frac{\pi }{P} \times \frac{t}{T},r_{1} > 0.5} \hfill \\ {1,r_{1} \le 0.5} \hfill \\ \end{array} } \right.$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{1}$$\end{document} represents a uniformly distributed random number in the interval [0, 1]. High school education stage In the high school education stage, ECO retains the identical subpopulation construction criterion employed in the secondary stage. The school population updates its positions through comprehensive integration of centroid, best, and worst individuals, thereby accommodating the broader demands of the student cohort. Conversely, student individuals enroll in the currently best-performing school. These corresponding behaviors are mathematically expressed in Eqs. (11) and (12).11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} + \left( {X_{best} - X_{mean}^{t} } \right) \times randn - \left( {X_{worst} - X_{mean}^{t} } \right) \times randn$$\end{document}12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} - P \times \left( {E \times X_{best} - X_{i}^{t} } \right)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{worst}$$\end{document} denotes the position of the worst individual discovered thus far (i.e., the global worst). In the basic ECO algorithm, the three educational phases execute sequentially in a cyclical manner as iterations progress. Specifically, the primary education stage is activated at iteration 1, the secondary education stage at iteration 2, and the high school education stage at iteration 3. This rotational execution pattern repeats cyclically until termination criteria are satisfied. Enhanced multi-strategy educational competition optimizer This section may be divided by subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn. Undergraduate education stage The canonical ECO algorithm concludes its evolutionary trajectory at the high school education stage, which imposes inherent constraints on its optimization capability. Specifically, this three-phase architecture exhibits pronounced susceptibility to premature convergence in complex multimodal landscapes: the intensified exploitation mechanisms characterizing the terminal phase rapidly erode population diversity in the absence of sustained exploratory provisions. Consequently, the algorithm struggles to maintain search diversity throughout prolonged optimization processes, compromising its reliability for high-dimensional or intricately structured problems. To remedy these deficiencies, this study introduces an intermediate undergraduate education stage (UES) following the high school phase. By deploying the current-to-Xwmean/1 mutation strategy for positional updating, the UES realizes controlled directional perturbations through weighted mean-centric difference vectors. This mechanism rejuvenates population diversity via subpopulation-based information exchange while preserving convergence momentum, thereby enhancing basin-hopping capability across rugged fitness landscapes. The resultant four-phase hierarchical framework improves robustness across varying problem complexities and dimensionalities. The positional update mechanism of the UES is mathematically articulated in the following equations.13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} + 0.8 \times \left( {X_{{{\mathrm{wmean}}}}^{t} - X_{i}^{t} } \right) + 0.8 \times \left( {X_{R1}^{t} - X_{R2}^{t} } \right)$$\end{document}14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{{\mathrm{wmean}}}}^{t} = \frac{2}{N} \times \sum\limits_{i = 1}^{0.5N} {\left( {\theta_{i} \times X_{i}^{t} } \right)}$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta_{i} = \ln \left( {0.5N + 1} \right)/\left( {\sum\limits_{i = 1}^{0.5N} {\left( {\ln \frac{{\left( {0.5N + 1} \right)}}{i}} \right)} } \right)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{{\mathrm{wmean}}}}^{t}$$\end{document} denotes the weighted centroid of the elite subpopulation, with fitness-proportional weighting amplifying the guidance contribution of high-quality solutions. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta_{i}$$\end{document} denotes the weighting coefficient assigned to each individual. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{R1}^{t}$$\end{document} represents an individual randomly selected from the current population, whose introduction reinforces population diversity through stochastic perturbation. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{R2}^{t}$$\end{document} signifies an individual stochastically drawn from the union of current population and external archive, a dual-source selection mechanism that concurrently preserves diversity (via population component) and exploits historical information (via archive component). The external archive is constructed in the same manner as that of L-SHADE. It is noteworthy that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{R1}^{t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{R2}^{t}$$\end{document} are regenerated independently per iteration to maintain stochastic independence and prevent premature convergence. Graduate education stage The UES introduced in the preceding section, in conjunction with the original three educational phases, constitutes a comprehensive search mechanism spanning the entire optimization trajectory. While UES effectively remedies the deficiency of population diversity deterioration characteristic of basic ECO in later evolutionary phases, it does not substantively enhance local exploitation capability. Given that the undergraduate phase precedes the terminal graduate education stage, this study proposes a Graduate Education Stage (GES) update strategy explicitly designed for intensified exploitation. In GES, intensified exploitation is prioritized, rendering elite guidance paramount. However, exclusive reliance on the global best risks premature convergence to local optima. Inspired by the Grey Wolf Optimizer’s hierarchical leadership structure, GES stochastically selects one guiding individual from the top-three ranked candidates per iteration, thereby diversifying elite information sources while maintaining exploitation pressure. It should be noted that the selection probabilities for the three individuals are identical, each being one-third. Furthermore, the elite subpopulation encapsulates the swarm’s evolutionary trajectory. Estimation of Distribution Algorithms leverage probabilistic models to learn high-quality solution distributions, enabling refined local search within promising regions and remedying ECO’s deficient terminal-phase convergence precision. Additionally, model-guided sampling exhibits enhanced directionality, curtailing ineffective search expenditures and accelerating convergence toward optimal basins. Accordingly, we propose a GES positional update strategy that integrates elite individual guidance with subpopulation distribution learning, as conceptually illustrated in Fig. 2. The mathematical formulation is presented as follows.16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Cov = \frac{2}{N} \times \sum\limits_{i = 1}^{0.5N} {\left( {X_{i}^{t} - X_{{{\mathrm{wmean}}}}^{t} } \right) \times \left( {X_{i}^{t} - X_{{{\mathrm{wmean}}}}^{t} } \right)^{T} } ,X_{i}^{t} \in P_{dominant}$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = \left( {X_{{{\mathrm{wmean}}}}^{t} + X_{e}^{t} + X_{i}^{t} } \right)/3 + g_{i} ,g_{i} \sim N\left( {0,Cov} \right)$$\end{document}where Cov denotes the covariance matrix of the elite subpopulation, capturing the distributional dispersion of high-quality solutions. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{e}^{t}$$\end{document} represents one individual stochastically selected from the top-three ranked candidates, serving as the primary directional guide for the mutation operator. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{dominant}$$\end{document} represents the elite subpopulation, specifically the fitness-ranked superior fraction of the current population.Fig. 2The schematic of GES. The schematic of GES. At each search iteration, the appropriate educational phase is deterministically selected from the four-stage hierarchy (including UES) based on evolutionary progress, for population-wide updating. Subsequently, GES is applied to the top G% of individuals ranked by fitness for further refinement. This sequential orchestration of UES and GES achieves functional complementarity: UES reinvigorates population diversity through differentiated mutation, while GES channels this diversified evolutionary potential toward refined exploitation via elite-guided probabilistic sampling. Random walk restart strategy The later evolutionary phases of the basic ECO algorithm, diversity rapidly deteriorates, causing the population to cluster in local optimal regions. Furthermore, when the best individual becomes trapped in local optima, the remaining population struggles to achieve substantive improvements during subsequent updates. Under such circumstances, the basic algorithm lacks effective mechanisms to restore exploratory capabilities. To remedy these deficiencies, this study proposes a Random Walk Restart Strategy (RWS) to address stagnation scenarios that are intractable for canonical ECO. RWS emulates the stochastic wandering behavior observed in natural organisms, incorporating Cauchy mutation to perturb current positional coordinates. The Cauchy distribution exhibits a higher probability of generating large-span jumps compared to Gaussian distributions, effectively “catapulting” stagnated individuals out of local basins. Even when the population has largely converged, Cauchy mutation retains the potential to produce exploratory points distant from current regions, thereby preventing wasted computational effort on futile fine-tuning near suboptimal solutions. On the other hand, Lévy flight predominantly performs small-scale random walks, while occasionally making direct jumps to another local optimum. In contrast, Cauchy mutation exhibits greater stability than Lévy flight and is capable of generating more effective jumps. Moreover, to further enhance overall convergence performance and optimization capability, this study designs a cosine decay factor to dynamically modulate search step sizes, ensuring that the algorithm emphasizes extensive exploration in early iterations and fine-grained exploitation in later phases. The mathematical formulation of RWS is presented in Eq. (18):18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{i}^{t + 1} = X_{i}^{t} + \phi \times \left( {X_{i}^{t} - X_{best} } \right) \times cathy$$\end{document}19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\phi = \lambda_{\min } + \left( {\lambda_{\max } - \lambda_{\min } } \right) \times \cos \left( {\frac{\pi t}{{2T}}} \right)$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$cathy = \gamma \times \tan \left( {\pi \left( {rand - 0.5} \right)} \right)$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\max }$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\min }$$\end{document} denote the maximum and minimum values of the cosine attenuation factor ϕ, respectively. In this paper, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\max }$$\end{document} is set to 1 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda_{\min }$$\end{document} to 0.01. γ denotes the scale parameter of the Cauchy function, fixed at 1 in this study. RWS assigns each individual a memory register S to record stagnation frequency, as depicted in Algorithm 1. When S exceeds the prescribed threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\max }$$\end{document}, RWS identifies the individual as stagnated and executes Eq. (18) for positional reinitialization. Recognizing that excessive perturbation in later phases would compromise population convergence, the threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\max }$$\end{document} is progressively increased as the search advances, as formulated in Eq. (21).21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\max } = 1 + \frac{3t}{T}$$\end{document} Algorithm 1: Pseudo-code of RWS. Algorithm 1: Pseudo-code of RWS. Pseudocode of EMSECO The pseudocode of proposed algorithm is shown in Algorithm 2 and its flowchart is depicted in Fig. 3.Algorithm 2: Pseudo-code of EMSECO algorithm.Fig. 3The flowchart of EMSECO. Algorithm 2: Pseudo-code of EMSECO algorithm. The flowchart of EMSECO. Time complexity analysis Time complexity analysis serves as a pivotal metric for evaluating the computational efficiency of meta-heuristic algorithms. For the basic ECO, let N, D and T denote the population size, the dimensionality of the solution space, and the maximum number of iterations, respectively. During the initialization phase, the generation of N individuals across D dimensions yields a time complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {N \times D} \right)$$\end{document}. The computational complexity of all three search stages is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {\frac{N}{3} \times D \times T} \right)$$\end{document}. Consequently, the overall time complexity of the basic ECO is calculated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {N \times D} \right) + O\left( {\frac{N}{3} \times D \times T} \right) + O\left( {\frac{N}{3} \times D \times T} \right) + O\left( {\frac{N}{3} \times D \times T} \right) = + O\left( {N \times D \times T} \right)$$\end{document}. Regarding the proposed EMSECO, the initialization phase similarly involves traversing N individuals across D dimensions, resulting in a complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {N \times D} \right)$$\end{document}. The undergraduate education stage (UES) is integrated into the original cyclical framework alongside the three canonical educational phases, with all four stages executed in rotational succession. Consequently, each stage maintains identical computational complexity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {\frac{N}{4} \times D \times T} \right)$$\end{document}. For the graduate education stage (GES), assuming \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{G}$$\end{document} individuals participate in this phase, the computational complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {N_{G} \times D \times T} \right)$$\end{document}. For the random walk restart strategy (RWS), assuming \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{R}$$\end{document} individuals fall into stagnation per iteration, the computational complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {N_{R} \times D \times T} \right)$$\end{document}. Thus, the total time complexity of EMSECO aggregates to:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {N \times D} \right) + O\left( {\frac{N}{4} \times D \times T} \right) + O\left( {\frac{N}{4} \times D \times T} \right) + O\left( {\frac{N}{4} \times D \times T} \right) + O\left( {\frac{N}{4} \times D \times T} \right) + O\left( {N_{G} \times D \times T} \right) + O\left( {N_{R} \times D \times T} \right) = O\left( {\left( {N + N_{G} + N_{R} } \right) \times D \times T} \right)$$\end{document} Assuming all individuals execute GES and RWS per iteration, the maximum time complexity of EMSECO is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {3N \times D \times T} \right)$$\end{document}. The minimum complexity is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {N \times D \times T} \right)$$\end{document} when neither strategy is triggered. Thus, the primary determinants of EMSECO’s computational overhead are the execution frequencies of GES and RWS. To ensure experimental fairness, all competing algorithms employ maximum fitness evaluation count (FEsMax) as the termination criterion. This standardization prevents unfair advantages arising from multiple internal iterations per algorithmic cycle. Results and analysis of numerical experiments In this section, we conduct systematic validation of the proposed EMSECO algorithm employing comprehensive benchmark suites. To rigorously appraise its optimization efficacy, comparative experiments were orchestrated against a spectrum of canonical and state-of-the-art metaheuristic variants. Quantitative performance indicators—specifically mean objective values and standard deviations—were utilized to dissect convergence behavior and solution robustness. Furthermore, Friedman’s non-parametric statistical test was deployed to ascertain significant performance differentials among competing algorithms, complemented by Nemenyi post-hoc analysis for visualized pairwise comparisons. Additionally, Wilcoxon rank-sum tests were executed to statistically validate the superiority of EMSECO relative to individual counterparts. This multi-faceted analytical framework, integrating descriptive statistics with inferential hypothesis testing, substantiates EMSECO’s competitive advantages across complex test landscapes, thereby furnishing empirical grounding for its deployment in real-world optimization endeavors. Experimental setup and performance metrics To empirically substantiate the efficacy of EMSECO, this section presents statistical assessments and convergence analyses conducted on the CEC 2017 benchmark suite. Comparative investigations were orchestrated against the canonical ECO alongside three foundational metaheuristics—AE51, MRFO52, SAO53—and three enhanced variants: EOSMA54, ALSHADE55, and EPSCA56. Together, these basic, novel, and variants provide a comprehensive comparison for evaluating the effectiveness and robustness of the proposed method. To ensure equitable performance realization, all competitor algorithms were configured with parameter settings prescribed in their original publications, as detailed in Table 1.Table 1Algorithm parameters setting.AlgorithmParameters settingEMSECO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 30,H = 0.5,G_{1} = 0.2,G_{2} = 0.1$$\end{document}ECO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 30,H = 0.5,G_{1} = 0.2,G_{2} = 0.1$$\end{document}AEN = 30,a = 4MRFON = 30,S = 2SAON = 30,a = 0.35,β = 5/7EOSMA\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 30,a_{1} = 2,a_{2} = 1,GP = 0.5,z = 0.7$$\end{document}ALSHADEN = 30D,H = 6,F = 0.5,CR = 0.5,p = 0.11EPSCAEN = 30,a = 2,c = 90 Table 1Algorithm parameters setting.AlgorithmParameters settingEMSECO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 30,H = 0.5,G_{1} = 0.2,G_{2} = 0.1$$\end{document}ECO\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 30,H = 0.5,G_{1} = 0.2,G_{2} = 0.1$$\end{document}AEN = 30,a = 4MRFON = 30,S = 2SAON = 30,a = 0.35,β = 5/7EOSMA\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N = 30,a_{1} = 2,a_{2} = 1,GP = 0.5,z = 0.7$$\end{document}ALSHADEN = 30D,H = 6,F = 0.5,CR = 0.5,p = 0.11EPSCAEN = 30,a = 2,c = 90 Algorithm parameters setting. The IEEE Congress on Evolutionary Computation (CEC) introduced the CEC 2017 benchmark suite, comprising practical engineering optimization challenges, mathematical models, and synthetically designed complex test functions. This suite enjoys extensive international adoption, enabling rigorous evaluation of algorithmic performance across varying problem scales, landscape characteristics, and complexity hierarchies—encompassing unimodal, multimodal, hybrid, and composition functions. Its capacity to comprehensively mirror optimization capabilities and robustness attributes under diverse problem conditions has established it as a globally recognized standard for algorithmic benchmarking. Following the exclusion of F2, the suite comprises 29 test functions: unimodal functions (F1, F3), multimodal functions (F4–F10), hybrid functions (F11–F20), and composition functions (F21–F30). All functions support 10-, 30-, 50-, and 100-dimensional instantiations within a uniform search space of [− 100, 100]. Table 2 provides detailed specifications for the CEC 2017 suite.Table 2The information of the CEC2017 benchmark test functions.TypeNoFunctions nameDimensionMinUnimodal functionsF1Shifted and Rotated Bent Cigar Function10/30/50/100100F3Shifted and Rotated Zakharov Function10/30/50/100300Multimodal functionsF4Shifted and Rotated Rosenbrock’s Function10/30/50/100400F5Shifted and Rotated Rastrigin’s Function10/30/50/100500F6Shifted and Rotated Expanded Scaffer’s F6 Function10/30/50/100600F7Shifted and Rotated Lunacek Bi_Rastrigin Function10/30/50/100700F8Shifted and Rotated Non-Continuous Rastrigin’s Function10/30/50/100800F9Shifted and Rotated Levy Function10/30/50/100900F10Shifted and Rotated Schwefel’s Function10/30/50/1001000Hybrid functionsF11Hybrid Function 1 (N = 3)10/30/50/1001100F12Hybrid Function 2 (N = 3)10/30/50/1001200F13Hybrid Function 3 (N = 3)10/30/50/1001300F14Hybrid Function 4 (N = 4)10/30/50/1001400F15Hybrid Function 5 (N = 4)10/30/50/1001500F16Hybrid Function 6 (N = 4)10/30/50/1001600F17Hybrid Function 6 (N = 5)10/30/50/1001700F18Hybrid Function 6 (N = 5)10/30/50/1001800F19Hybrid Function 6 (N = 5)10/30/50/1001900F20Hybrid Function 6 (N = 6)10/30/50/1002000Composition functionsF21Composition Function 1 (N = 3)10/30/50/1002100F22Composition Function 2 (N = 3)10/30/50/1002200F23Composition Function 3 (N = 4)10/30/50/1002300F24Composition Function 4 (N = 4)10/30/50/1002400F25Composition Function 5 (N = 5)10/30/50/1002500F26Composition Function 6 (N = 5)10/30/50/1002600F27Composition Function 7 (N = 6)10/30/50/1002700F28Composition Function 8 (N = 6)10/30/50/1002800F29Composition Function 9 (N = 3)10/30/50/1002900F30Composition Function 10 (N = 3)10/30/50/1003000 Table 2The information of the CEC2017 benchmark test functions.TypeNoFunctions nameDimensionMinUnimodal functionsF1Shifted and Rotated Bent Cigar Function10/30/50/100100F3Shifted and Rotated Zakharov Function10/30/50/100300Multimodal functionsF4Shifted and Rotated Rosenbrock’s Function10/30/50/100400F5Shifted and Rotated Rastrigin’s Function10/30/50/100500F6Shifted and Rotated Expanded Scaffer’s F6 Function10/30/50/100600F7Shifted and Rotated Lunacek Bi_Rastrigin Function10/30/50/100700F8Shifted and Rotated Non-Continuous Rastrigin’s Function10/30/50/100800F9Shifted and Rotated Levy Function10/30/50/100900F10Shifted and Rotated Schwefel’s Function10/30/50/1001000Hybrid functionsF11Hybrid Function 1 (N = 3)10/30/50/1001100F12Hybrid Function 2 (N = 3)10/30/50/1001200F13Hybrid Function 3 (N = 3)10/30/50/1001300F14Hybrid Function 4 (N = 4)10/30/50/1001400F15Hybrid Function 5 (N = 4)10/30/50/1001500F16Hybrid Function 6 (N = 4)10/30/50/1001600F17Hybrid Function 6 (N = 5)10/30/50/1001700F18Hybrid Function 6 (N = 5)10/30/50/1001800F19Hybrid Function 6 (N = 5)10/30/50/1001900F20Hybrid Function 6 (N = 6)10/30/50/1002000Composition functionsF21Composition Function 1 (N = 3)10/30/50/1002100F22Composition Function 2 (N = 3)10/30/50/1002200F23Composition Function 3 (N = 4)10/30/50/1002300F24Composition Function 4 (N = 4)10/30/50/1002400F25Composition Function 5 (N = 5)10/30/50/1002500F26Composition Function 6 (N = 5)10/30/50/1002600F27Composition Function 7 (N = 6)10/30/50/1002700F28Composition Function 8 (N = 6)10/30/50/1002800F29Composition Function 9 (N = 3)10/30/50/1002900F30Composition Function 10 (N = 3)10/30/50/1003000 The information of the CEC2017 benchmark test functions. To guarantee objectivity, fairness, and reliability in experimental outcomes, all algorithms were executed under identical hardware-software configurations. The experimental platform utilized MATLAB R2024a (which is available at https://www.mathworks.com/products/matlab.html) on Windows 11, powered by an AMD Ryzen 9 7945HX CPU @2.5 GHz processor with 32 GB RAM. Algorithmic parameters were standardized as follows: maximum function evaluations (FEsMax) = 1000 × D, with 30 independent replications per test case. Performance was quantified through three critical metrics: best, mean, and std. Statistical significance was rigorously established via Friedman’s test, Nemenyi post-hoc analysis, and Wilcoxon rank-sum tests, collectively corroborating the superiority of EMSECO. Exploration and exploitation analysis In the analytical framework of population-based metaheuristics, meticulous examination of two distinct yet interdependent facets is imperative: individual agent behavior and collective population dynamics, conventionally conceptualized as exploration and exploitation. Exploration entails extensive traversal of the solution space to identify promising regions, whereas exploitation involves intensive refinement within previously identified high-quality neighborhoods to pinpoint local optima. In this study, the proportional contributions of exploration and exploitation are quantified via Eqs. (22) and (23), respectively. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Div_{\max }$$\end{document} denotes the maximum population diversity observed throughout the entire evolutionary trajectory, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Div_{t}$$\end{document} represents the population diversity at the tth iteration, computed through Eq. (24).22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Exploration\% = \frac{{Div_{t} }}{{Div_{\max } }} \times 100\%$$\end{document}23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Exploitation\% = \frac{{\left| {Div_{t} - Div_{\max } } \right|}}{{Div_{\max } }} \times 100\%$$\end{document}24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Div_{t} = \frac{1}{D}\sum\limits_{j = 1}^{D} {\left( {\frac{1}{N}\sum\limits_{i = 1}^{N} {\left| {median\left( {X_{i}^{j} \left( t \right)} \right) - X_{i}^{j} \left( t \right)} \right|} } \right)}$$\end{document} To empirically validate the capability of EMSECO in balancing exploitation and exploration, the algorithm was deployed on the CEC 2017 benchmark suite. The corresponding experimental outcomes are visualized in Fig. 4.Fig. 4The exploration and exploitation of EMSECO. The exploration and exploitation of EMSECO. As illustrated in Fig. 4, EMSECO exhibits a relatively elevated exploration rate during the initial evolutionary phases, which facilitates comprehensive coverage of the entire solution space. As the optimization trajectory progresses, the exploration rate gradually diminishes while the exploitation rate concomitantly intensifies. This dynamic transition enables the algorithm to more effectively identify and refine high-quality solutions. Such behavioral patterns reflect a smooth shift from exploration-oriented to exploitation-oriented search modes. Particularly noteworthy is that even in later stages dominated by exploitation, the algorithm maintains a residual exploration capacity, which serves as an effective safeguard against entrapment in local optima. Parameter sensitivity analysis Parameter sensitivity analysis constitutes a critical component in metaheuristic algorithm research, serving to evaluate the impact of parameter perturbations on algorithmic performance and thereby elucidate the degree of dependency on specific parameter configurations. Through systematic sensitivity analysis, the adaptability of proposed algorithms under diverse parameter settings can be validated, mitigating performance reliance on particular parameter values while furnishing rational guidance for parameter selection. Consequently, rigorous analysis of critical parameters not only deepens theoretical comprehension of underlying algorithmic mechanisms but also establishes foundations for reliable deployment in complex real-world optimization scenarios. This subsection presents comprehensive parameter sensitivity analysis for the proposed EMSECO algorithm, providing theoretical and empirical support for parameter calibration. Specifically, parameter G as the control parameter governing the graduate education stage, determining the proportion of individuals participating in this phase. Its value directly modulates the exploration–exploitation equilibrium. When G assumes larger values, a greater fraction of individuals is guided by elite subpopulation and global best information, enabling EMSECO to conduct intensified refinement within promising regions and remedy the deficient convergence precision characteristic of canonical ECO in terminal phases. However, excessive G values risk premature focalization on local neighborhoods, potentially compromising global optimality discovery. Conversely, diminished G values reduce graduate-stage participation, impairing EMSECO’s capacity to maintain adequate population diversity. To empirically assess the performance implications of varying G values, EMSECO was benchmarked on the CEC 2017 test suite across G ∈ {10, 20, 30, 40, 50, 60}. All remaining experimental parameters were standardized according to Sect. "Experimental setup and performance metrics" specifications. Figure 5 presents the Friedman test rankings of EMSECO with varying G values across different dimensional configurations. As evidenced by Fig. 5, EMSECO consistently outperforms the canonical ECO algorithm irrespective of G magnitude, thereby substantiating the efficacy of the GES in enhancing ECO’s optimization capability. Notably, G = 30 yields the superior average ranking for EMSECO. The ranking trajectories across dimensional settings reveal a non-monotonic performance pattern: algorithmic efficacy initially deteriorates with increasing G, subsequently recovers and strengthens. This inverse-U relationship indicates that both insufficient and excessive G values impair the full realization of GES potential, underscoring the criticality of calibrated parameter selection.Fig. 5The Friedman rankings of EMSECO with different G on CEC 2017 test suite. The Friedman rankings of EMSECO with different G on CEC 2017 test suite. Strategy effectiveness analysis To rigorously validate the efficacy of the proposed enhancement mechanisms, ablation studies are conducted in this subsection. Three single-strategy ECO variants (UECO, GECO, RECO) and three dual-strategy ECO variants (UGECO, URECO, GRECO) are included in the comparative assessment. The specific configurations of these six ECO variants are detailed in Table 3. In Table 3, “Yes” indicates incorporation of the corresponding strategy, whereas “No” signifies its exclusion. For instance, UECO denotes the ECO algorithm augmented with UES, while UGECO represents the configuration with RWS removed (i.e., combining UES and GES only). To ensure experimental fairness, identical experimental conditions are strictly enforced across all competitors. Furthermore, to guarantee statistical stability and reliability, each test function is independently executed 30 times. All experimental data, including Best, Mean, and Std, are compiled in Tables 9, 10, 11, 12 of Appendix A.Table 3The information of the ECO variants.StrategyUECOGECORECOUGECOURECOGRECOEMSECOUESYesNoNoYesYesNoYesGESNoYesNoYesNoYesYesRWSNoNoYesNoYesYesYes Table 3The information of the ECO variants.StrategyUECOGECORECOUGECOURECOGRECOEMSECOUESYesNoNoYesYesNoYesGESNoYesNoYesNoYesYesRWSNoNoYesNoYesYesYes The information of the ECO variants. Table 4 summarizes the Friedman test results for the ablation studies at a significance level of 0.05. As indicated in Table 4, statistically significant differences are detected among all competing algorithms across four dimensional settings. Specifically, EMSECO, incorporating all three enhancement mechanisms, achieves the top rank, demonstrating that the synergistic integration of mechanisms outperforms individual or pairwise combinations. Furthermore, the Friedman test results unequivocally confirm that each proposed mechanism contributes positively to ECO’s performance improvement.Table 4The information of the ECO variants.Test suiteDimensionECOUECOGECORECOUGECOURECOGRECOEMSECOp-valueCEC-2017107.4484.2765.8286.7242.5173.1384.5171.5525.34E-28307.7594.9664.9317.1032.3454.1723.6211.1035.21E-33507.8285.2764.8976.9662.5524.0343.3451.1036.10E-331007.5523.4835.6556.7592.9313.2763.8972.4482.46E-23Average rank7.6474.5005.3286.8882.5863.6553.8451.552N/AOverall rank85672341N/A Table 4The information of the ECO variants.Test suiteDimensionECOUECOGECORECOUGECOURECOGRECOEMSECOp-valueCEC-2017107.4484.2765.8286.7242.5173.1384.5171.5525.34E-28307.7594.9664.9317.1032.3454.1723.6211.1035.21E-33507.8285.2764.8976.9662.5524.0343.3451.1036.10E-331007.5523.4835.6556.7592.9313.2763.8972.4482.46E-23Average rank7.6474.5005.3286.8882.5863.6553.8451.552N/AOverall rank85672341N/A The information of the ECO variants. Figure 6 provides visual representation of the ranking distributions for different ECO variants across four dimensionalities. As evidenced by Fig. 6, UECO attains the best performance among the three single-strategy variants, indicating that UES provides the most substantial contribution to EMSECO. Conversely, GRECO exhibits the poorest performance among the three dual-strategy variants, suggesting that removing UES inflicts the most severe performance degradation. The contrasting behaviors of these two variants collectively substantiate the efficacy of UES. Similarly, it can be inferred that GES contributes more substantially than RWS to EMSECO’s overall enhancement. UES ensures sufficient population diversity through differential guidance, while the directional perturbation of superior individuals guarantees the convergence trend, thereby achieving performance enhancement. GES focuses on strengthening local search; consequently, the improvement is limited for complex functions that require a favorable balance between exploration and exploitation. RWS is primarily designed to escape premature convergence, rather than concentrating on enhancing exploitation and exploration capabilities.Fig. 6The Friedman rankings of ECO variants on CEC 2017 test suite. The Friedman rankings of ECO variants on CEC 2017 test suite. Figure 7 visualizes the Wilcoxon rank-sum test results for the ablation studies. In the figure, “superior” indicates that the ECO variant significantly outperforms the basic ECO algorithm; “similar” denotes comparable performance between the variant and basic ECO; and “inferior” signifies that the ECO variant is significantly weaker than basic ECO. As evidenced by Fig. 7, none of the single-mechanism ECO variants exhibit significant inferiority to basic ECO, confirming that individual improvement mechanisms exert no detrimental effects on ECO’s performance. Although EMSECO demonstrates significant inferiority to basic ECO in both 10D and 100D settings, its overall performance remains superior, as indicated by the tallest red bar representing the largest count of functions where EMSECO significantly outperforms basic ECO.Fig. 7The Wilcoxon rank sum test results of ECO variants on CEC 2017 test suite. The Wilcoxon rank sum test results of ECO variants on CEC 2017 test suite. The convergence curves depicted in Fig. 8 demonstrate that the proposed enhancement mechanisms substantially improve ECO’s convergence efficiency. EMSECO achieves the optimal convergence speed and accuracy in most scenarios, with the sole exception of 100D F1 where it ranks second. Collectively, these experimental results ultimately corroborate the efficacy of each strategy in enhancing algorithmic performance, while simultaneously validating the soundness and reliability of the ablation study methodology employed herein.Fig. 8The convergence curves of ECO variants on CEC 2017 test suite. The convergence curves of ECO variants on CEC 2017 test suite. Performance comparison analysis In this subsection, EMSECO is benchmarked against three foundational metaheuristics and three enhanced variants. The comparative results are presented in Tables 13, 14, 15, 16 of Appendix A. Figure 9 illustrates the ranking distributions of EMSECO relative to competing algorithms across diverse test functions, preliminarily demonstrating the superior performance of the proposed EMSECO algorithm. As depicted in Fig. 9, the ranking distributions of each algorithm across different functions form a surface; smaller surface area indicates superior overall performance. It is evident that EMSECO encloses the minimal surface area, signifying its best overall performance. In stark contrast, basic ECO predominantly occupies lower ranks, resulting in the largest enclosed area.Fig. 9The ranking radars of EMSECO and its competitors on CEC 2017 test suite. The ranking radars of EMSECO and its competitors on CEC 2017 test suite. Non-parametric tests constitute a class of statistical methods that do not rely on assumptions regarding population distributions, making them particularly suitable for scenarios involving non-normally distributed data or limited sample sizes. Subsequent analyses will statistically evaluate the experimental results through Friedman test, Nemenyi post-hoc test, and Wilcoxon rank-sum test, complemented by convergence curves and box plots to assess convergence behavior and robustness. The Wilcoxon rank-sum test was employed to conduct pairwise comparisons between EMSECO and seven competing algorithms. The significance level was set at α (= 0.05). According to the testing principle, if the calculated P-value is less than α, the null hypothesis is rejected, indicating a statistically significant difference between the two samples. In the resulting tables, the symbol ( +) denotes that EMSECO is significantly superior to the comparative algorithm, (−) indicates that EMSECO is significantly inferior, and ( =) signifies no statistically significant difference between the two methods. The Wilcoxon rank-sum test results are summarized in Table 5. As indicated in Table 5, EMSECO consistently achieves more ( +) symbols than (−) symbols across comparisons with different algorithms, implying that the proposed EMSECO is significantly superior to competing algorithms in aggregate. When benchmarked against the three foundational metaheuristics, EMSECO attains a ( +) count exceeding the combined total of (−) and ( =), demonstrating significant superiority on the majority of test functions. Similarly, in comparisons with EOSMA and ALSHADE, EMSECO again realizes a ( +) majority over the sum of (−) and ( =), confirming marked advantages even relative to enhanced algorithms. Regarding EPSCA, EMSECO only achieves ( +) dominance in 30D and 50D, whereas 10D and 100D exhibit balanced performance profiles. This suggests functional similarity between EMSECO and EPSCA on most 10D and 100D functions, though EMSECO maintains superior overall performance.Table 5The Wilcoxon rank sum test results of EMSECO and its competitors on CEC 2017 test suite.EMSECO vs. (+ / = /−)CEC-2017 test suite10D30D50D100DECO27/2/029/0/029/0/025/4/0AE28/0/122/4/322/3/419/5/5SAO23/6/027/1/125/3/121/8/0MRFO26/3/022/4/317/6/617/3/9EOSMA24/3/219/7/317/9/319/6/4ALSHADE24/3/222/3/423/3/319/3/7EPSCA12/13/416/10/311/15/316/4/9 Table 5The Wilcoxon rank sum test results of EMSECO and its competitors on CEC 2017 test suite.EMSECO vs. (+ / = /−)CEC-2017 test suite10D30D50D100DECO27/2/029/0/029/0/025/4/0AE28/0/122/4/322/3/419/5/5SAO23/6/027/1/125/3/121/8/0MRFO26/3/022/4/317/6/617/3/9EOSMA24/3/219/7/317/9/319/6/4ALSHADE24/3/222/3/423/3/319/3/7EPSCA12/13/416/10/311/15/316/4/9 The Wilcoxon rank sum test results of EMSECO and its competitors on CEC 2017 test suite. To comprehensively evaluate overall performance distributions, this study employs the Friedman test. This non-parametric statistical method, based on rank assignment, serves as a robust tool for analyzing differences among multiple related samples, particularly suited for experimental designs involving repeated measurements. The Friedman test enables global assessment of aggregate optimization performance across different algorithms without imposing assumptions on underlying data distributions, thereby effectively mitigating evaluation biases potentially arising from reliance on single test functions or isolated performance metrics. Table 6 presents the Friedman test results comparing EMSECO with competing algorithms, visually illustrated in Fig. 10. As evidenced by the p-values in Table 6, statistically significant performance differences exist between EMSECO and all competing algorithms across four dimensional settings. As depicted in Fig. 10, EMSECO achieves the top rank in all dimensionalities, indicating its dominant position on the majority of test functions. Although the margin of superiority narrows with increasing dimensionality, EMSECO maintains its leading position. In stark contrast, basic ECO consistently lags behind, further corroborating the efficacy of the proposed methodology. Detailed Friedman analysis is provided below.Table 6The Friedman test results of EMSECO and its competitors on CEC 2017 test suite.Test suiteDimensionEMSECOECOAESAOMRFOEOSMAALSHADEEPSCAp-valueCEC-2017101.7246.8625.3455.0695.5174.2073.5173.7595.47E-15301.9317.5174.3795.5865.0004.0693.7933.7241.89E-16502.0697.1384.8975.5174.4483.7594.1034.0695.09E-131002.6906.2414.6215.0694.1385.2413.8284.1722.87E-06Average rank2.1036.9404.8105.3104.7764.3193.8103.931N/AOverall rank18765423N/AFig. 10The Friedman rankings of EMSECO and its competitors on CEC 2017 test suite. Table 6The Friedman test results of EMSECO and its competitors on CEC 2017 test suite.Test suiteDimensionEMSECOECOAESAOMRFOEOSMAALSHADEEPSCAp-valueCEC-2017101.7246.8625.3455.0695.5174.2073.5173.7595.47E-15301.9317.5174.3795.5865.0004.0693.7933.7241.89E-16502.0697.1384.8975.5174.4483.7594.1034.0695.09E-131002.6906.2414.6215.0694.1385.2413.8284.1722.87E-06Average rank2.1036.9404.8105.3104.7764.3193.8103.931N/AOverall rank18765423N/A The Friedman test results of EMSECO and its competitors on CEC 2017 test suite. The Friedman rankings of EMSECO and its competitors on CEC 2017 test suite. On the 10-dimensional test suite, EMSECO, ALSHADE, and EPSCA occupy the top three positions, followed sequentially by EOSMA, SAO, AE, MRFO, and ECO. Collectively, EMSECO secures the foremost position with a Friedman score of 1.724, while basic ECO ranks last at 6.862. On the 30-dimensional test suite, EMSECO, EPSCA, and ALSHADE occupy the top three positions, followed sequentially by EOSMA, AE, MRFO, and ECO. Collectively, EMSECO secures the foremost position with a Friedman score of 1.931, while basic ECO ranks last at 7.517. On the 50-dimensional test suite, EMSECO, EOSMA, and EPSCA occupy the top three positions, followed sequentially by ALSHADE, MRFO, SAO, AE, and ECO. Collectively, EMSECO secures the foremost position with a Friedman score of 2.069, while basic ECO ranks last at 7.138. On the 100-dimensional test suite, EMSECO, ALSHADE, and MRFO occupy the top three positions, followed sequentially by EPSCA, SAO, AE, EOSMA, and ECO. Collectively, EMSECO secures the foremost position with a Friedman score of 2.690, while basic ECO ranks last at 6.241. It should be noted that the Friedman test merely indicates whether significant differences exist among all algorithms, without quantifying the magnitude of disparity between EMSECO and any specific competitor. Consequently, the Nemenyi post-hoc test is employed for more granular analysis. Based on Friedman ranking results, the Nemenyi method computes a critical difference value (CDV), which serves as a threshold for assessing algorithmic equivalence: if the difference in mean ranks between EMSECO and another algorithm falls below the CDV, the two methods are deemed statistically indistinguishable. The CDV is calculated via Eq. (25).25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CDV = q_{a} \times \sqrt {\frac{{K\left( {K + 1} \right)}}{6M}}$$\end{document}where M denotes the number of test functions, K represents the number of competing algorithms, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$q_{a}$$\end{document} is the critical value obtained from statistical tables (set to 3.0310 in this study). Figure 11 presents the Nemenyi post-hoc test results derived from Friedman rankings comparing EMSECO with competing algorithms. As depicted in Fig. 11, EMSECO is connected to ALSHADE at 10D, to EPSCA at 30D, and to EOSMA at 50D through critical difference bars. This indicates statistical equivalence between EMSECO and these three algorithms at respective dimensionalities. Regarding 100D, EMSECO is linked to ALSHADE, EPSCA, and EOSMA via CDV-length segments, signifying no statistically significant superiority or inferiority among these four enhanced algorithms on 100-dimensional functions. In stark contrast, no connections exist between EMSECO and basic ECO or other foundational algorithms. These results demonstrate that EMSECO significantly outperforms basic ECO, AE, SAO, and MRFO on the CEC-2017 test suite, and significantly surpasses ALSHADE, EPSCA, and EOSMA in most dimensional settings. Thus, the superior performance of EMSECO is corroborated through multiple statistical validation methods.Fig. 11The Nemenyi test results of EMSECO and its competitors on CEC 2017 test suite. The Nemenyi test results of EMSECO and its competitors on CEC 2017 test suite. To evaluate the convergence performance of the proposed EMSECO, representative test functions from diverse categories were selected for in-depth analysis. The comparative results are presented in Fig. 12. As evidenced by Fig. 12, EMSECO exhibits rapid descent in early search phases and swift convergence to superior solutions in later phases across most functions, outperforming competing algorithms. Even when initial convergence speed is suboptimal, EMSECO demonstrates sustained convergence capability throughout subsequent iterations.Fig. 12The convergence curves of EMSECO and its competitors on CEC 2017 test suite. The convergence curves of EMSECO and its competitors on CEC 2017 test suite. Specifically, for the unimodal function F1, EMSECO achieves the fastest convergence speed and precision across 30D, 50D, and 100D, with the sole exception of 10D where EPSCA exhibits superior initial convergence. Notably, EMSECO maintains continuous convergence throughout the entire search cycle without stagnation, attributable to the elite guidance mechanism of GES which enhances exploitation capability. Regarding the multimodal function F8, EMSECO exhibits inferior convergence speed to EPSCA at 10D and 30D, yet attains superior convergence precision. At 50D, EMSECO demonstrates both faster convergence speed and higher precision. At 100D, EMSECO maintains rapid convergence though precision falls slightly below EPSCA. Collectively, EMSECO mitigates premature convergence tendencies through UES and sustains search vigor via robust global exploration. For the hybrid function F14 and the composition function F25, EMSECO achieves optimal convergence speed and precision, demonstrating superior efficacy when addressing complex landscapes. This enhanced performance stems from diversity preservation via UES, evolutionary trajectory capture through GES, and local optimum escape capability enabled by RWS. In summary, EMSECO exhibits exceptional convergence capability across functionally diverse problems, attributable to its capacity for rapid local optimum escape and reliable convergence toward global optima. To visually assess the stability and robustness of EMSECO, this study employs box plot analysis to compare its distributional characteristics with competing algorithms. In standard box plots, the median line represents the central tendency, while the upper and lower edges of the box denote the third and first quartiles, respectively. The vertical height of the box reflects data dispersion, and the extending whiskers indicate the maximum and minimum values within the effective range, with outliers marked as individual points beyond the whiskers. Consequently, shorter boxes signify more concentrated distributions and superior stability, whereas lower box positions indicate higher solution quality. The comparative results are presented in Fig. 13.Fig. 13The box plots of EMSECO and its competitors on CEC 2017 test suite. The box plots of EMSECO and its competitors on CEC 2017 test suite. Observation of Fig. 13 reveals that EMSECO exhibits outliers solely on 100D F8; in all remaining cases, including the relatively poor-performing 10D F25, no outliers are detected, substantiating its favorable robustness. Furthermore, EMSECO demonstrates lower-positioned boxes on F1 and F14, indicating superior solution quality. For the remaining three scenarios of F25, EMSECO exhibits shorter boxes with more concentrated distributions. Regarding F8, although EMSECO’s box is not the shortest, its solution quality remains superior to competing algorithms. The results from 30 independent runs ultimately validate that EMSECO not only excels in global optimization but also demonstrates remarkable stability in solution quality, thereby consolidating its standing as a highly competitive algorithm with comprehensive advantages. Engineering applications To validate the applicability of the proposed EMSECO in addressing real-world engineering challenges, this study employs seven representative problems from the CEC 2020 "Real-World Constrained Optimization" benchmark suite (detailed in Table 7). Constraint handling is accomplished via the penalty function method: any violation of hard constraints incurs a penalized fitness value, thereby rapidly eliminating infeasible solutions. EMSECO and seven selected competing algorithms were independently executed 30 times on each problem; the best, mean, and standard deviation of obtained results, together with Friedman test and Wilcoxon rank-sum test statistics, are summarized in Table 8.Table 7The information of engineering problems.ProblemNameDghEP01Tension/compression spring design problem330EP02Pressure vessel design problem440EP03Three-bar truss design problem230EP04Welded beam design problem450EP05Gear train design problem411EP06Cantilever beam design problem510EP07Step-cone pulley problem583Table 8Comparison results of EMSECO and its competitors on engineering problems.NoIndexEMSECOECOAESAOMRFOEOSMAALSHADEEPSCAEP01best1.2835E-021.4016E-021.2990E-021.5717E-021.3344E-021.3870E-021.5937E-021.5990E-02mean1.3865E-026.5958E + 012.4629E-022.3104E-022.6476E-023.6772E-021.4617E + 032.9484E + 02Std1.1231E-032.2062E + 029.6223E-035.5197E-031.0777E-021.9232E-027.9617E + 031.5416E + 03rank16324587EP02best7.7545E + 038.9223E + 031.3624E + 041.8554E + 041.8022E + 044.0077E + 043.9586E + 041.7124E + 04mean1.6589E + 041.4738E + 053.7166E + 045.9214E + 047.3121E + 041.1932E + 051.5788E + 051.1810E + 05Std5.9216E + 031.5004E + 053.1859E + 044.0533E + 045.8329E + 047.7228E + 049.4704E + 049.1683E + 04rank17234685EP03best2.6389E + 022.6397E + 022.6389E + 022.6395E + 022.6397E + 022.6412E + 022.6412E + 022.6411E + 02mean2.6395E + 022.6669E + 022.6430E + 022.6504E + 022.6750E + 022.6578E + 022.6633E + 022.6565E + 02Std7.8455E-022.4781E + 004.2027E-018.2399E-013.2451E + 001.3442E + 001.5416E + 001.8907E + 00rank17238564EP04best1.7847E + 001.9763E + 001.8099E + 002.0530E + 002.1520E + 002.0467E + 001.9943E + 001.9724E + 00mean2.1592E + 004.4229E + 042.6261E + 002.9363E + 009.2936E + 038.0184E + 011.8124E + 023.0654E + 00Std1.7384E-011.6186E + 055.8880E-015.3879E-012.8089E + 044.1958E + 026.6512E + 026.0955E-01rank18237564EP05best2.3078E-112.3576E-092.3078E-112.3576E-092.3576E-091.1834E-099.9216E-108.8876E-10mean1.9233E-092.1308E-061.0065E-072.8539E-061.6977E-063.6511E-075.5065E-065.0772E-07Std2.9374E-098.4937E-063.9310E-079.4291E-063.4561E-065.9794E-077.1191E-068.2642E-07rank16275384EP06best1.5002E + 001.4264E + 001.9829E + 001.4221E + 001.8238E + 002.6589E + 002.4739E + 001.4828E + 00mean2.5374E + 005.4442E + 004.2987E + 001.9916E + 003.4047E + 004.7831E + 005.2400E + 003.6935E + 00Std7.4254E-013.1220E + 001.3990E + 003.2397E-018.6999E-011.3840E + 001.1060E + 001.6658E + 00rank28513674EP07best1.7152E + 014.8694E + 012.0211E + 015.3227E + 021.5916E + 028.6605E + 026.4410E + 026.3558E + 01mean2.1635E + 015.3008E + 023.3131E + 021.1104E + 081.6337E + 033.1506E + 089.3921E + 081.8976E + 03Std6.4543E + 003.4193E + 027.6300E + 024.0792E + 081.1469E + 037.3565E + 081.4246E + 094.1479E + 03rank13264785Friedman Rank1.1436.4292.5713.5715.0005.2867.2864.714 + / = /–N/A7/0/07/0/06/0/17/0/07/0/07/0/07/0/0 Table 7The information of engineering problems.ProblemNameDghEP01Tension/compression spring design problem330EP02Pressure vessel design problem440EP03Three-bar truss design problem230EP04Welded beam design problem450EP05Gear train design problem411EP06Cantilever beam design problem510EP07Step-cone pulley problem583 The information of engineering problems. Table 8Comparison results of EMSECO and its competitors on engineering problems.NoIndexEMSECOECOAESAOMRFOEOSMAALSHADEEPSCAEP01best1.2835E-021.4016E-021.2990E-021.5717E-021.3344E-021.3870E-021.5937E-021.5990E-02mean1.3865E-026.5958E + 012.4629E-022.3104E-022.6476E-023.6772E-021.4617E + 032.9484E + 02Std1.1231E-032.2062E + 029.6223E-035.5197E-031.0777E-021.9232E-027.9617E + 031.5416E + 03rank16324587EP02best7.7545E + 038.9223E + 031.3624E + 041.8554E + 041.8022E + 044.0077E + 043.9586E + 041.7124E + 04mean1.6589E + 041.4738E + 053.7166E + 045.9214E + 047.3121E + 041.1932E + 051.5788E + 051.1810E + 05Std5.9216E + 031.5004E + 053.1859E + 044.0533E + 045.8329E + 047.7228E + 049.4704E + 049.1683E + 04rank17234685EP03best2.6389E + 022.6397E + 022.6389E + 022.6395E + 022.6397E + 022.6412E + 022.6412E + 022.6411E + 02mean2.6395E + 022.6669E + 022.6430E + 022.6504E + 022.6750E + 022.6578E + 022.6633E + 022.6565E + 02Std7.8455E-022.4781E + 004.2027E-018.2399E-013.2451E + 001.3442E + 001.5416E + 001.8907E + 00rank17238564EP04best1.7847E + 001.9763E + 001.8099E + 002.0530E + 002.1520E + 002.0467E + 001.9943E + 001.9724E + 00mean2.1592E + 004.4229E + 042.6261E + 002.9363E + 009.2936E + 038.0184E + 011.8124E + 023.0654E + 00Std1.7384E-011.6186E + 055.8880E-015.3879E-012.8089E + 044.1958E + 026.6512E + 026.0955E-01rank18237564EP05best2.3078E-112.3576E-092.3078E-112.3576E-092.3576E-091.1834E-099.9216E-108.8876E-10mean1.9233E-092.1308E-061.0065E-072.8539E-061.6977E-063.6511E-075.5065E-065.0772E-07Std2.9374E-098.4937E-063.9310E-079.4291E-063.4561E-065.9794E-077.1191E-068.2642E-07rank16275384EP06best1.5002E + 001.4264E + 001.9829E + 001.4221E + 001.8238E + 002.6589E + 002.4739E + 001.4828E + 00mean2.5374E + 005.4442E + 004.2987E + 001.9916E + 003.4047E + 004.7831E + 005.2400E + 003.6935E + 00Std7.4254E-013.1220E + 001.3990E + 003.2397E-018.6999E-011.3840E + 001.1060E + 001.6658E + 00rank28513674EP07best1.7152E + 014.8694E + 012.0211E + 015.3227E + 021.5916E + 028.6605E + 026.4410E + 026.3558E + 01mean2.1635E + 015.3008E + 023.3131E + 021.1104E + 081.6337E + 033.1506E + 089.3921E + 081.8976E + 03Std6.4543E + 003.4193E + 027.6300E + 024.0792E + 081.1469E + 037.3565E + 081.4246E + 094.1479E + 03rank13264785Friedman Rank1.1436.4292.5713.5715.0005.2867.2864.714 + / = /–N/A7/0/07/0/06/0/17/0/07/0/07/0/07/0/0 Comparison results of EMSECO and its competitors on engineering problems. As indicated in Table 8, EMSECO delivers superior solutions to competing algorithms on 6 out of 10 problems, with the sole exception of EP06 where it underperforms relative to SAO. EMSECO achieves the top Friedman ranking of 1.143, followed by AE and SAO at 2.571 and 3.571, respectively. Although the ranking gap between EMSECO and AE is modest, significant performance disparities exist across individual engineering problems. Notably, despite the larger numerical difference between EMSECO and SAO rankings, EMSECO exhibits significant inferiority to SAO specifically on EP06. This outcome reaffirms the No Free Lunch (NFL) theorem: algorithms demonstrating inferior performance on certain problems may excel on others. The Wilcoxon rank-sum test corroborates that EMSECO possesses significant advantages on the majority of engineering problems. In summary, the proposed EMSECO algorithm exhibits outstanding performance and substantial application potential when addressing real-world constrained engineering design optimization tasks. Conclusions To enhance the search capability of the basic Educational Competition Optimizer (ECO), this study proposes an Enhanced Multi-Strategy Educational Competition Optimizer (EMSECO). The proposed EMSECO integrates three synergistic improvement mechanisms: the Undergraduate Education Stage (UES), the Graduate Education Stage (GES), and the Random Walk Restart Strategy (RWS). Firstly, UES is embedded into ECO’s inherent search framework, leveraging diverse differential vector guidance to enrich population diversity. Secondly, GES is introduced as an external intensification module, utilizing elite individual and elite subpopulation guidance to curtail ineffective search and strengthen exploitation capability. Thirdly, RWS is imposed on stagnated individuals to facilitate new solution generation and assist the algorithm in escaping local optima. To validate the efficacy of the proposed methodology, this study employs the CEC 2017 benchmark suite, widely recognized in the academic community. Through Friedman test, Nemenyi post-hoc test, and Wilcoxon rank-sum test, the superiority of EMSECO relative to advanced algorithms is substantiated. Furthermore, six engineering design optimization problems provide additional validation of the practical applicability of the proposed approach. Notwithstanding the favorable overall performance of EMSECO, certain limitations remain that warrant further investigation. On one hand, the algorithmic validation has predominantly relied on benchmark functions and specific engineering cases, which may not fully mirror the complex and dynamic nature of real-world application scenarios, potentially constraining comprehensive assessment of applicability across diverse practical problems. On the other hand, the incorporation of multiple enhancement strategies, while improving optimization efficacy, introduces additional computational overhead, which may impose constraints on real-time optimization applications requiring rapid response capabilities. Regarding future research directions, we envision extending the algorithm’s application domain to resource scheduling, investment decision-making, and industrial planning. Simultaneously, a multi-objective variant will be developed to address increasingly intricate real-world optimization challenges. Furthermore, we intend to explore the utilization of large language models for designing more potent search operators, thereby further fortifying the search capability of the proposed methodology.