A Hybrid Metaheuristic for High-Dimensional Constrained Optimization: Applications to Logistics and UAV Path Planning Authors: Yarong Li, Chuandong Qin Published: 2026 ABSTRACT Inspired by the hovering, diving, and cooperative hunting behaviors of the pied kingfisher, the Pied Kingfisher Optimizer (PKO) has demonstrated competitive performance in optimization tasks. However, it exhibits several phase-specific limitations, including uneven population distribution caused by random initialization, insufficient use of historical information during exploration, over-reliance on the global best during exploitation, and weakly guided perturbation in the symbiosis phase. To address these issues, this study proposes an Improved Pied Kingfisher Optimizer (IPKO), which incorporates biologically inspired adaptive strategies. Drawing inspiration from the kingfisher’s diverse perching, gaze adjustment during hovering, evasive diving after failed strikes, and territory shifting based on flock position, four mechanisms are developed. Specifically, sine chaotic opposition-based initialization enhances population diversity; adaptive directional search regulates the exploration–exploitation balance; stochastic perturbation-based information fusion improves the ability to escape local optima; and centroid-based adaptive boundary handling strengthens constraint adaptability. The performance of IPKO is evaluated on the CEC2017 benchmark suite (10, 30, 50, and 100 dimensions) and two real-world engineering problems. Experimental results show that IPKO achieves superior overall performance compared with eleven state-of-the-art algorithms, with statistical significance confirmed by the Friedman test and Holm’s post-hoc procedure. Ablation studies further verify the contribution of each strategy. In engineering applications such as cold chain logistics and dynamic multi-UAV cooperative path planning, the IPKO algorithm demonstrates superior solution quality, robustness, and constraint-handling capability compared with competing algorithms. These results demonstrate that IPKO is a robust and effective bio-inspired optimization approach for solving complex, high-dimensional constrained engineering problems. 1. Introduction In nature, the pied kingfisher performs a remarkable sequence of behaviors—hovering to survey prey, diving to capture, and cooperating with conspecifics—to achieve efficient foraging in complex environments. This biological strategy offers a compelling metaphor for solving complex optimization problems in engineering. In the era of Industry 4.0 and smart manufacturing, engineering systems are generating increasingly complex optimization challenges that often involve high-dimensional, nonlinear, and multi-constrained decision spaces, making traditional deterministic methods inadequate. Metaheuristic optimization algorithms have become indispensable tools for solving complex optimization problems in science and engineering, particularly when traditional deterministic methods are ineffective due to nonlinearity, high dimensionality, and the absence of gradient information [1]. These algorithms have been successfully applied across a wide range of domains. In medicine, improved particle swarm optimization has been used to determine optimal mounting positions for medical robots, while enhanced whale optimization algorithms have been employed for COVID-19 detection [2,3]. In image processing, grey wolf optimizer variants have been applied to multilevel thresholding problems [4]. In power systems, metaheuristics have been widely used to solve optimal reactive power dispatch and economic emission dispatch problems [5,6]. In engineering optimization, various algorithms have demonstrated strong performance in solving complex design and planning problems [7]. These applications highlight the effectiveness and versatility of metaheuristic optimization methods. In addition to metaheuristic optimization methods, mathematical optimization approaches have also demonstrated strong capabilities in solving engineering optimization problems. For example, convex relaxation and convex modeling techniques have been successfully applied in hybrid AC/DC microgrid economic dispatch problems due to their computational efficiency and theoretical guarantees. Representative studies such as “Enabling high-efficiency economic dispatch of hybrid AC/DC networked microgrids: Steady-state convex bi-directional converter models” provide effective deterministic optimization frameworks for structured optimization problems [8]. Metaheuristic algorithms are generally classified into four categories: swarm intelligence (SI) algorithms, evolutionary algorithms (EA), physics-based algorithms, and human-based algorithms [9]. Swarm intelligence algorithms simulate collective behaviors of natural organisms, and include particle swarm optimization (PSO), grey wolf optimizer (GWO), whale optimization algorithm (WOA), Harris hawks optimization (HHO), marine predators algorithm (MPA), and golden jackal optimization (GJO) [10,11,12,13,14,15]. Evolutionary algorithms mimic biological evolution and natural selection, including genetic algorithms (GA) and differential evolution (DE) [16,17]. Physics-based algorithms are inspired by physical laws, such as the gravitational search algorithm (GSA) and simulated annealing (SA) [18,19]. Human-based algorithms model human learning or social behaviors, such as teaching–learning-based optimization (TLBO) [20]. According to the No Free Lunch theorem [21], no single optimization algorithm can achieve optimal performance across all problem domains. Therefore, continuous development and improvement of metaheuristic algorithms remain essential. Despite their success, many metaheuristic algorithms still suffer from issues such as premature convergence, insufficient exploration in high-dimensional spaces, and weak adaptability under complex constraints [22]. To address these limitations, recent studies have increasingly focused on integrating multiple strategies within a unified optimization framework to enhance robustness and generalization capability. For example, adaptive and multi-strategy particle swarm optimization variants have been proposed to balance exploration and exploitation more effectively [23,24,25]. Similarly, algorithms such as GJO, MPA, and snake optimizer (SO) [26] have been widely improved through various enhancement strategies, such as opposition-based learning, chaotic mapping, adaptive perturbation, and hybrid update mechanisms, which has become a common approach to improving convergence accuracy and maintaining population diversity in metaheuristic optimization. These developments indicate that hybridization, adaptive mechanisms, and multi-strategy integration have become dominant trends in recent metaheuristic design, particularly for addressing high-dimensional and constrained optimization problems. The Pied Kingfisher Optimizer (PKO) is a recently proposed swarm-based metaheuristic inspired by the hunting behavior of pied kingfishers [27]. In nature, a kingfisher first hovers to survey prey distribution (exploration), then dives sharply toward a selected target (exploitation), and occasionally performs cooperative fishing with conspecifics (symbiosis). Accordingly, PKO organizes its search into four phases: initialization (positioning of virtual kingfishers), exploration (global search via hovering simulation), exploitation (local search via diving), and symbiosis (information sharing among individuals). This biomimetic mapping provides an intuitive optimization framework. PKO has demonstrated competitive performance on benchmark functions and engineering applications, indicating its potential as an effective optimization method. However, like many swarm-based algorithms, PKO exhibits several limitations when dealing with complex optimization problems. First, its random initialization strategy may lead to insufficient population diversity and uneven spatial distribution, limiting early-stage exploration. Second, the exploration phase relies heavily on stochastic parameter switching without effectively utilizing historical information, resulting in reduced search efficiency. Third, the exploitation phase is strongly guided by the current global best solution, which may cause the algorithm to become trapped in local optima and suffer from premature convergence. Finally, the symbiosis phase employs random perturbations without directional guidance, leading to unstable performance, especially in high-dimensional or constrained scenarios. Although several studies have applied PKO to different problems, its phase-specific limitations remain largely unaddressed. These limitations collectively restrict the convergence accuracy, robustness, and scalability of the original PKO. Motivated by the above observations, this paper proposes an Improved Pied Kingfisher Optimizer (IPKO) that further draws biological inspiration from previously unmodeled kingfisher behaviors. Specifically:Diverse perching (sine chaotic opposition-based initialization): In nature, kingfishers rarely perch at the same spot repeatedly; they survey from multiple perches to gain different perspectives. We simulate this by using sine chaotic mapping to generate diverse initial positions, then apply opposition-based learning to expand the search space, mimicking the bird’s ability to consider opposite directions.Gaze adjustment during hovering (adaptive directional search): While hovering, a kingfisher continuously adjusts its gaze based on both the best-known prey location and random movements of nearby birds. Our adaptive directional search guides exploration using both global best information and randomly selected individuals.Evasive diving after failed strikes (stochastic perturbation-based information fusion): When a kingfisher fails to catch prey, it performs an erratic dive to relocate. Similarly, our stochastic perturbation strategy adds controlled randomness to help the algorithm escape local optima.Territory shifting based on flock position (centroid-based adaptive boundary handling): A kingfisher’s territory is not rigid; it adapts based on the average position of the flock. We use the population centroid to dynamically adjust out-of-bound individuals, maintaining feasibility while preserving diversity. Diverse perching (sine chaotic opposition-based initialization): In nature, kingfishers rarely perch at the same spot repeatedly; they survey from multiple perches to gain different perspectives. We simulate this by using sine chaotic mapping to generate diverse initial positions, then apply opposition-based learning to expand the search space, mimicking the bird’s ability to consider opposite directions. Diverse perching (sine chaotic opposition-based initialization): In nature, kingfishers rarely perch at the same spot repeatedly; they survey from multiple perches to gain different perspectives. We simulate this by using sine chaotic mapping to generate diverse initial positions, then apply opposition-based learning to expand the search space, mimicking the bird’s ability to consider opposite directions. Gaze adjustment during hovering (adaptive directional search): While hovering, a kingfisher continuously adjusts its gaze based on both the best-known prey location and random movements of nearby birds. Our adaptive directional search guides exploration using both global best information and randomly selected individuals. Gaze adjustment during hovering (adaptive directional search): While hovering, a kingfisher continuously adjusts its gaze based on both the best-known prey location and random movements of nearby birds. Our adaptive directional search guides exploration using both global best information and randomly selected individuals. Evasive diving after failed strikes (stochastic perturbation-based information fusion): When a kingfisher fails to catch prey, it performs an erratic dive to relocate. Similarly, our stochastic perturbation strategy adds controlled randomness to help the algorithm escape local optima. Evasive diving after failed strikes (stochastic perturbation-based information fusion): When a kingfisher fails to catch prey, it performs an erratic dive to relocate. Similarly, our stochastic perturbation strategy adds controlled randomness to help the algorithm escape local optima. Territory shifting based on flock position (centroid-based adaptive boundary handling): A kingfisher’s territory is not rigid; it adapts based on the average position of the flock. We use the population centroid to dynamically adjust out-of-bound individuals, maintaining feasibility while preserving diversity. Territory shifting based on flock position (centroid-based adaptive boundary handling): A kingfisher’s territory is not rigid; it adapts based on the average position of the flock. We use the population centroid to dynamically adjust out-of-bound individuals, maintaining feasibility while preserving diversity. The effectiveness of the proposed IPKO is evaluated on the IEEE CEC2017 benchmark suite with multiple dimensional settings (10, 30, 50, and 100). In addition, two real-world engineering problems, namely cold chain logistics routing and dynamic multi-UAV cooperative path planning, are considered to further validate its practical applicability. The performance of IPKO is compared with eleven state-of-the-art algorithms, including GA, PSO, DE, GWO, WOA, HHO, GBO, HBA, and the original PKO. Statistical significance is assessed using the Wilcoxon signed-rank test and Friedman test. Furthermore, ablation studies and parameter sensitivity analysis are conducted to verify the contribution and robustness of each proposed strategy. The main contributions of this paper are summarized as follows:A biomimetic enhancement framework that models four specific kingfisher behaviors (diverse perching, gaze adjustment, evasive diving, and territory shifting) into mathematical optimization strategies, extending the biological fidelity of the original PKO.A sine chaotic mapping combined with opposition-based learning strategy is introduced to enhance population diversity and improve early-stage exploration.An adaptive directional search mechanism is proposed to improve exploration efficiency and accelerate convergence.A stochastic perturbation-based information fusion strategy is developed to prevent premature convergence and enhance the ability to escape local optima.A centroid-based adaptive boundary handling mechanism is designed to effectively maintain feasibility under complex constraints.Comprehensive experiments on benchmark functions and real-world engineering problems demonstrate that the proposed IPKO achieves superior performance in terms of convergence accuracy, robustness, and solution quality. A biomimetic enhancement framework that models four specific kingfisher behaviors (diverse perching, gaze adjustment, evasive diving, and territory shifting) into mathematical optimization strategies, extending the biological fidelity of the original PKO. A biomimetic enhancement framework that models four specific kingfisher behaviors (diverse perching, gaze adjustment, evasive diving, and territory shifting) into mathematical optimization strategies, extending the biological fidelity of the original PKO. A sine chaotic mapping combined with opposition-based learning strategy is introduced to enhance population diversity and improve early-stage exploration. A sine chaotic mapping combined with opposition-based learning strategy is introduced to enhance population diversity and improve early-stage exploration. An adaptive directional search mechanism is proposed to improve exploration efficiency and accelerate convergence. An adaptive directional search mechanism is proposed to improve exploration efficiency and accelerate convergence. A stochastic perturbation-based information fusion strategy is developed to prevent premature convergence and enhance the ability to escape local optima. A stochastic perturbation-based information fusion strategy is developed to prevent premature convergence and enhance the ability to escape local optima. A centroid-based adaptive boundary handling mechanism is designed to effectively maintain feasibility under complex constraints. A centroid-based adaptive boundary handling mechanism is designed to effectively maintain feasibility under complex constraints. Comprehensive experiments on benchmark functions and real-world engineering problems demonstrate that the proposed IPKO achieves superior performance in terms of convergence accuracy, robustness, and solution quality. Comprehensive experiments on benchmark functions and real-world engineering problems demonstrate that the proposed IPKO achieves superior performance in terms of convergence accuracy, robustness, and solution quality. Based on these enhancements, this study provides a complete biomimetic optimization framework that not only faithfully models the pied kingfisher’s hunting behaviors but also demonstrates its practical transferability to two representative engineering systems: cold chain logistics (analogous to efficient prey routing) and dynamic multi-UAV cooperative path planning (analogous to flock hunting). The results confirm that biologically faithful algorithm design can yield superior performance in real-world optimization tasks. The remainder of this paper is organized as follows. Section 2 introduces the original PKO and analyzes its limitations. Section 3 presents the proposed IPKO and its improvement strategies. Section 4 reports experimental results on benchmark functions. Section 5 discusses applications to real-world engineering problems. Finally, Section 6 concludes the paper and outlines future research directions. 2. Original Pied Kingfisher Optimization Algorithm (PKO) The Pied Kingfisher Optimization (PKO) algorithm is a metaheuristic optimization method designed for solving single-objective optimization problems. The algorithm is constructed based on four search operators, namely perching, hovering, diving, and commensalism. These operators are mapped to four sequential phases: initialization, exploration, exploitation, and symbiosis. 2.1. Initialization Phase PKO generates an initial population consisting of N candidate solutions randomly distributed within the feasible search space. The position of the i-th individual in the j-th dimension is defined as: (1)Xi,j=LB+(UB−LB)×rand,1≤i≤N,1≤j≤Dim where rand∈[0,1] is a uniformly distributed random number, and LB and UB represent the lower and upper bounds of the decision space, respectively. This initialization strategy ensures feasibility but does not explicitly promote population diversity or spatial coverage. 2.2. Exploration Phase The exploration phase is responsible for global search and is implemented through two complementary update strategies. The position update rule is given by:(2)Xnew(i,:)=X(i,:)+α·T·(X(j,:)−X(i,:)),i≠j where α is a random vector sampled from a normal distribution, and T is a time-varying control parameter that determines the search mode. The parameter T switches between two update schemes with different step characteristics, facilitating variation in search range. One scheme produces relatively small perturbations, while the other allows larger position adjustments, thereby enhancing population dispersion. 2.3. Exploitation Phase The exploitation phase focuses on refining candidate solutions in promising regions of the search space. The update rule is defined as:(3)Xnew(i,:)=X(i,:)+HuntingAbility×o×α·(b−bestPosition) where HuntingAbility is a scaling factor related to the fitness of individuals, o is a decreasing coefficient controlling the step size over iterations, α is a stochastic vector, and bestPosition denotes the current global best solution. This mechanism intensifies the search around high-quality solutions but introduces a strong dependency on the accuracy of the global best. 2.4. Symbiosis Phase The symbiosis phase introduces stochastic interactions among individuals to improve population diversity and support escape from local optima. The update rule is expressed as:(4)Xnew(i,:)=X(r,:)+o×α·X(i,:)−X(r,:) where r is a randomly selected individual index within the population. This operation perturbs the current solution using the relative distance between individuals. The activation of this phase is controlled by the Predation Efficiency (PE), which decreases linearly during the optimization process, thereby reducing randomness in later stages. 2.5. Summary and Motivation for Improvement Although PKO provides a structured framework that integrates exploration and exploitation, several inherent limitations can be identified in its core components. These limitations directly motivate the design of the proposed Improved Pied Kingfisher Optimizer (IPKO). Initialization limitation: Initialization limitation: The population is generated using a purely random initialization strategy, which does not guarantee uniform distribution or sufficient diversity. This may result in clustered individuals and inadequate coverage of the search space, particularly in high-dimensional problems. This limitation motivates the introduction of a diversity-enhanced initialization strategy. Exploration limitation: Exploration limitation: The exploration phase primarily depends on stochastic parameter sampling without incorporating any historical or population-level guidance. As a result, the search process may suffer from low efficiency, redundant exploration, and slow convergence. This limitation motivates the development of a guided exploration mechanism that leverages population information. Exploitation limitation: Exploitation limitation: The exploitation phase relies heavily on the current global best solution to guide the search. When the global best is trapped in a local optimum, the population tends to converge prematurely, reducing solution quality. This limitation motivates the design of an adaptive exploitation strategy to alleviate premature convergence. Symbiosis limitation: Symbiosis limitation: The symbiosis phase introduces random perturbations without directional or adaptive control. This leads to unstable performance, as the mechanism may either fail to escape local optima or disrupt promising solutions. This limitation motivates the incorporation of a controlled and adaptive local escape mechanism. Overall motivation: Overall motivation: Overall motivation: To address the above limitations, this study proposes an Improved Pied Kingfisher Optimizer (IPKO), which enhances population diversity, improves search guidance, strengthens exploitation robustness, and introduces adaptive escape strategies. The details of these improvements are presented in Section 3. 3. Improved Pied Kingfisher Optimization Algorithm (IPKO) 3.1. Design Rationale of IPKO Although the original Pied Kingfisher Optimizer (PKO) provides a biologically inspired framework that combines exploration and exploitation behaviors, its search dynamics still exhibit several limitations when solving high-dimensional constrained optimization problems. As analyzed in Section 2, these limitations mainly include: (i) insufficient population diversity during initialization, (ii) weak utilization of directional and historical information during exploration, (iii) premature convergence caused by excessive dependence on the global best solution during exploitation, and (iv) unstable boundary adaptation and local escape capability in constrained search spaces. From a biomimetic perspective, pied kingfishers exhibit highly adaptive hunting behaviors in complex natural environments. During hovering, kingfishers continuously adjust gaze orientation to track moving prey and optimize attack direction. When prey capture fails, they perform evasive diving and dynamic repositioning to avoid repeated ineffective attacks. In addition, flock-level territory adjustment enables individuals to maintain spatial adaptability under environmental constraints. These adaptive biological behaviors provide important inspiration for improving the search dynamics of PKO. Motivated by these observations, this study proposes an Improved Pied Kingfisher Optimizer (IPKO) based on a coordinated adaptive search framework. The proposed framework establishes behavior-oriented cooperation among different search mechanisms across various optimization stages. Specifically, four complementary mechanisms are developed: (1) a sine chaotic opposition-based initialization strategy inspired by diversified perch selection behavior to improve population distribution and spatial coverage; (2) an adaptive directional search mechanism inspired by hovering gaze adjustment behavior to enhance directional exploration and adaptive exploitation; (3) a stochastic binary perturbation-based information fusion strategy inspired by evasive diving behavior after failed predation to improve local-optimum escape capability; (4) a centroid-based adaptive boundary handling mechanism inspired by territory relocation and cooperative flock positioning behavior to strengthen feasible-region adaptability under boundary constraints. The proposed IPKO constructs a coordinated adaptive enhancement framework in which each mechanism targets a specific deficiency of PKO while cooperatively improving population diversity, search guidance, exploitation robustness, and constrained search adaptability. The overall framework of IPKO is illustrated in Figure 1, where 3.2–3.5 in the figure correspond to Section 3.2, Section 3.3, Section 3.4 and Section 3.5 respectively. 3.2. Sine Chaotic Opposition-Based Initialization Strategy To overcome the limitations of purely random initialization, a sine chaotic opposition-based initialization strategy is introduced [28,29,30]. This strategy replaces random sampling with a structured population generation mechanism to improve spatial distribution and diversity. First, a sine chaotic sequence is generated as:(5)sine(i,j)=sin(π·sine(i,j−1)),j=2,…,dim where sine(i,1) is initialized randomly in [0,1]. The sequence is then mapped to the search space:(6)X(i,:)=lb+sine(i,:)·(ub−lb) To further enhance diversity, an opposition-based learning strategy is incorporated:(7)Xop(i,:)=r·(ub−lb)−X(i,:) where r∈[0,1] is a random scalar. Boundary constraints are enforced as:(8)Xop(i,:)=max(Xop(i,:),lb),Xop(i,:)=min(Xop(i,:),ub) The original and opposite populations are merged and ranked:(9)Xall=[X;Xop] and the best N individuals are retained:(10)pop=Xall(index(1:N),:) Compared with the purely random initialization in the original PKO, the proposed sine chaotic opposition-based strategy introduces a structured population generation mechanism. By leveraging the ergodicity of sine chaotic mapping and the symmetry of opposition-based learning, it produces an initial population with improved spatial coverage and reduced clustering. This leads to enhanced exploration capability in early iterations and faster convergence toward promising regions, providing a more reliable foundation for subsequent adaptive search processes, particularly in high-dimensional and multimodal optimization problems. 3.3. Adaptive Directional Search Mechanism To overcome the insufficient guidance and low search efficiency of the original PKO during the exploration stage, an adaptive directional search mechanism is proposed. This mechanism is inspired by the hovering and visual tracking behaviors of pied kingfishers, as well as the environmental perception mechanism in crayfish optimization [31,32]. In natural hunting processes, pied kingfishers continuously adjust their hovering posture and gaze orientation according to prey movement and environmental feedback. By integrating global guidance, peer-based directional perception, adaptive intensity regulation, and stochastic angular perturbation, the proposed mechanism establishes a direction-aware neighborhood search framework. The adaptive directional search is activated when(11)r≤(1−PE) where PE denotes the predation efficiency and r∈[0,1] is a uniformly distributed random number. A low predation efficiency indicates that the current search fails to approach promising regions, thereby triggering enhanced directional exploration. The update procedure consists of the following five steps, each fully defined for reproducibility. Step 1: Peer Individual Selection Step 1: Peer Individual Selection A random individual is selected from the population to provide auxiliary directional information:(12)r1=randi([1,Popsize]) This peer-based interaction reduces over-dependence on the global best solution and improves population-level directional adaptability. Step 2: Global Direction Construction Step 2: Global Direction Construction A directional vector toward the current global best solution is constructed as:(13)X2=(Best_position−X(i,:))×r where r∈[0,1] is the same random number used in the activation condition. This vector represents the primary search tendency toward promising regions. Step 3: Adaptive Search Intensity Regulation (B) Step 3: Adaptive Search Intensity Regulation (B) An adaptive modulation coefficient is introduced to balance exploration and exploitation over iterations:(14)B=c×cosπ2×1−tMaxiteration,c=2 Here, t is the current iteration and Maxiteration is the maximum number of iterations. B increases from 0 to 2 as the optimization progresses. In early iterations (t small, B≈0), the search step is nearly zero, preserving population diversity. In later iterations (t close to Maxiteration, B≈2), the directional movement is fully activated, enabling strong local exploitation around promising regions. The constant c=2 defines the upper bound of the modulation intensity. Step 4: Stochastic Angular Perturbation (θ) Step 4: Stochastic Angular Perturbation (θ) To simulate omnidirectional gaze adjustment, an angular perturbation mechanism is introduced:(15)θ=RouletteWheelSelection(p),p=1,2,…,360 Each angle θ (in degrees) is assigned an equal probability 1/360. The roulette wheel selection is implemented as:(16)θ=⌈360·u⌉,u∼U(0,1) The angle determines the direction of the sine/cosine components in the position update. A full circle (0∘ to 360∘) allows unbiased omnidirectional sampling. For example, θ=0∘ makes the cosine term dominate (moving directly toward the peer individual), while θ=90∘ makes the sine term dominate (lateral movement). This mechanism prevents repeated movement along fixed directions and significantly enhances spatial exploration diversity. Step 5: Direction-Aware Position Update with Perturbation Factor (V) Step 5: Direction-Aware Position Update with Perturbation Factor (V) The final position update is formulated as:(17)Xnew(i,:)=X2+X(r1,:)−X(i,:)×cos(θ)×B×V×rand+X(r1,:)×sin(θ)×B×rand where: rand∈[0,1] is a uniformly distributed random number regenerated each time the equation is evaluated.V is a discrete random variable representing the environmental perception intensity. Based on the crayfish-inspired environmental quality evaluation in the reference, V is independently sampled for each activation:(18)V=randi([2,5])(integeruniformlydrawnfrom{2,3,4,5}) A smaller value (V=2 or 3) indicates that the current local environment is perceived as favorable, encouraging conservative directional contraction. A larger value (V=4 or 5) emulates poor environmental quality (e.g., oxygen deficiency or food scarcity), triggering a stronger perturbation to escape local optima. The stochastic nature of V introduces non-deterministic step-size bursts, further improving robustness. For illustrative purposes, Figure 2c presents two representative cases (V=2 and V=5), corresponding to conservative and aggressive perturbation behaviors, respectively. rand∈[0,1] is a uniformly distributed random number regenerated each time the equation is evaluated. rand∈[0,1] is a uniformly distributed random number regenerated each time the equation is evaluated. V is a discrete random variable representing the environmental perception intensity. Based on the crayfish-inspired environmental quality evaluation in the reference, V is independently sampled for each activation:(18)V=randi([2,5])(integeruniformlydrawnfrom{2,3,4,5}) V is a discrete random variable representing the environmental perception intensity. Based on the crayfish-inspired environmental quality evaluation in the reference, V is independently sampled for each activation:(18)V=randi([2,5])(integeruniformlydrawnfrom{2,3,4,5}) The three parameters jointly determine the magnitude and direction of the search move:B provides an iteration-dependent global scaling that grows from 0 to 2, making early moves tentative and later moves become more directionally focused.θ decides how the movement is split into a radial component (toward the peer individual, via cosθ) and a lateral component (perpendicular to the peer direction, via sinθ).V randomly modulates the cosine-dominated radial step, amplifying it by a factor of 2 to 5. The sine term is not multiplied by V, preserving a baseline lateral exploration regardless of environmental perception. B provides an iteration-dependent global scaling that grows from 0 to 2, making early moves tentative and later moves become more directionally focused. B provides an iteration-dependent global scaling that grows from 0 to 2, making early moves tentative and later moves become more directionally focused. θ decides how the movement is split into a radial component (toward the peer individual, via cosθ) and a lateral component (perpendicular to the peer direction, via sinθ). θ decides how the movement is split into a radial component (toward the peer individual, via cosθ) and a lateral component (perpendicular to the peer direction, via sinθ). V randomly modulates the cosine-dominated radial step, amplifying it by a factor of 2 to 5. The sine term is not multiplied by V, preserving a baseline lateral exploration regardless of environmental perception. V randomly modulates the cosine-dominated radial step, amplifying it by a factor of 2 to 5. The sine term is not multiplied by V, preserving a baseline lateral exploration regardless of environmental perception. Table 1 provides a concrete example for different parameter combinations under a fixed peer direction. Figure 2 provides a schematic illustration to improve intuitive understanding:(a) Angular perturbation: A circle centered at the current individual X(i). The angle θ (sampled uniformly from 0∘ to 360∘) determines the search direction. The cosine component points toward the selected peer individual X(r1), while the sine component is perpendicular.(b) Adaptive intensity B over iterations: A plot showing B increasing from 0 to 2 following a cosine curve, with early iterations (exploration) in gray and late iterations (exploitation) in blue.(c) Perturbation factor V: Two arrows representing step sizes for V=2 (short arrow) and V=5 (long arrow), illustrating how a larger V triggers a longer jump. (a) Angular perturbation: A circle centered at the current individual X(i). The angle θ (sampled uniformly from 0∘ to 360∘) determines the search direction. The cosine component points toward the selected peer individual X(r1), while the sine component is perpendicular. (a) Angular perturbation: A circle centered at the current individual X(i). The angle θ (sampled uniformly from 0∘ to 360∘) determines the search direction. The cosine component points toward the selected peer individual X(r1), while the sine component is perpendicular. (b) Adaptive intensity B over iterations: A plot showing B increasing from 0 to 2 following a cosine curve, with early iterations (exploration) in gray and late iterations (exploitation) in blue. (b) Adaptive intensity B over iterations: A plot showing B increasing from 0 to 2 following a cosine curve, with early iterations (exploration) in gray and late iterations (exploitation) in blue. (c) Perturbation factor V: Two arrows representing step sizes for V=2 (short arrow) and V=5 (long arrow), illustrating how a larger V triggers a longer jump. (c) Perturbation factor V: Two arrows representing step sizes for V=2 (short arrow) and V=5 (long arrow), illustrating how a larger V triggers a longer jump. Comparison with the Original PKO The original PKO relies primarily on stochastic movement without explicit directional coordination or adaptive modulation. In contrast, the proposed mechanism integrates:global guidance (X2),peer-based directional perception (X(r1,:)−X(i,:)),iteration-dependent scaling (B),unbiased omnidirectional angle sampling (θ),stochastic step-size modulation (V). This design yields a balanced exploration–exploitation transition, stronger convergence, and improved robustness in high-dimensional constrained optimization problems. global guidance (X2), global guidance (X2), peer-based directional perception (X(r1,:)−X(i,:)), peer-based directional perception (X(r1,:)−X(i,:)), iteration-dependent scaling (B), iteration-dependent scaling (B), unbiased omnidirectional angle sampling (θ), unbiased omnidirectional angle sampling (θ), stochastic step-size modulation (V). stochastic step-size modulation (V). 3.4. Stochastic Binary Perturbation-Based Information Fusion Strategy To mitigate premature convergence caused by excessive dependence on the global best solution, a stochastic binary perturbation-based information fusion strategy is proposed [33]. This mechanism combines dimension-wise selective perturbation with information fusion from non-elite individuals. The update rule is defined as:(19)Xnew(i,:)=U1·X(i,:)+(1−U1)·y+rand·X(r1,:)−X(r2,:) where:U1=rand(1,Dim)>r is a binary vector controlling dimension-wise updates. It performs dimension-wise selective inheritance from the current individual, allowing part of the original solution structure to be preserved while perturbing only selected dimensions. This selective information fusion mechanism reduces destructive updates and improves search stability during local escape.y=X(i,:)+X(r3,:)/2.r1,r2,r3 are randomly selected indices. Compared with the original PKO, which tends to reinforce the region of the current global best and suffers from premature convergence, the proposed strategy introduces controlled stochasticity while preserving partial solution structure. By incorporating information from multiple randomly selected individuals and applying dimension-wise binary perturbations, it prevents search trajectory homogenization. This leads to enhanced ability to escape local optima and sustained population diversity throughout the optimization process. U1=rand(1,Dim)>r is a binary vector controlling dimension-wise updates. It performs dimension-wise selective inheritance from the current individual, allowing part of the original solution structure to be preserved while perturbing only selected dimensions. This selective information fusion mechanism reduces destructive updates and improves search stability during local escape. U1=rand(1,Dim)>r is a binary vector controlling dimension-wise updates. It performs dimension-wise selective inheritance from the current individual, allowing part of the original solution structure to be preserved while perturbing only selected dimensions. This selective information fusion mechanism reduces destructive updates and improves search stability during local escape. y=X(i,:)+X(r3,:)/2. y=X(i,:)+X(r3,:)/2. r1,r2,r3 are randomly selected indices. r1,r2,r3 are randomly selected indices. 3.5. Centroid-Based Adaptive Boundary Handling Mechanism To ensure feasibility under boundary constraints, a centroid-based adaptive boundary repair mechanism is introduced. Unlike direct truncation, this method utilizes population distribution information to perform adaptive correction [34]. First, boundary violations are detected:(20)flagub=X(i,:)>ub,flaglb=X(i,:)0.5)32: Xnew=U1·Xi(t)+(1−U1)·(y+rand()·(Xr1(t)−Xr2(t))) (Equation (19))33: Xi(t)=CentroidBoundary(Xnew)34: stuck=035: end if36: end for37:end for38:return Gbest, fbest 1:Step 1: Sine chaotic opposition-based initialization (Equations (5)–(10)) Step 1: Sine chaotic opposition-based initialization (Equations (5)–(10)) 2:Initialize population X, evaluate fitness, find Gbest, fbest Initialize population X, evaluate fitness, find Gbest, fbest 3:Set stuck=0 Set stuck=0 4:Step 2: Main loop Step 2: Main loop 5:for t=1 to MaxIter do for t=1 to MaxIter do 6: for i=1 to N do for i=1 to N do 7: Step 3: Adaptive directional search Step 3: Adaptive directional search 8: Compute PE Compute PE 9: if rand() ≤(1−PE)(Equation(11)) then if rand() ≤(1−PE)(Equation(11)) then 10: r1=randi([1,N]), r=rand() r1=randi([1,N]), r=rand() 11: X2=(Gbest−Xi(t−1))×r X2=(Gbest−Xi(t−1))×r 12: B=2×cos(π2×(1−tMaxIter)) B=2×cos(π2×(1−tMaxIter)) 13: θ=⌈360×rand()⌉ θ=⌈360×rand()⌉ 14: V=randi([2,5]) V=randi([2,5]) 15: Xnew=X2+(Xr1(t−1)−Xi(t−1))×cosθ×B×V×rand() Xnew=X2+(Xr1(t−1)−Xi(t−1))×cosθ×B×V×rand() 16: +Xr1(t−1)×sinθ×B×rand() +Xr1(t−1)×sinθ×B×rand() 17: Xi(t)=CentroidBoundary(Xnew) (Equations (20)–(23)) Xi(t)=CentroidBoundary(Xnew) (Equations (20)–(23)) 18: else else 19: Xi(t)=OriginalPKODiving(Xi(t−1)) Xi(t)=OriginalPKODiving(Xi(t−1)) 20: Xi(t)=CentroidBoundary(Xi(t)) Xi(t)=CentroidBoundary(Xi(t)) 21: end if end if 22: Step 4: Stochastic binary perturbation fusion Step 4: Stochastic binary perturbation fusion 23: if f(Xi(t))0.5) U1=(rand(1,Dim)>0.5) 32: Xnew=U1·Xi(t)+(1−U1)·(y+rand()·(Xr1(t)−Xr2(t))) (Equation (19)) Xnew=U1·Xi(t)+(1−U1)·(y+rand()·(Xr1(t)−Xr2(t))) (Equation (19)) 33: Xi(t)=CentroidBoundary(Xnew) Xi(t)=CentroidBoundary(Xnew) 34: stuck=0 stuck=0 35: end if end if 36: end for end for 37:end for end for 38:return Gbest, fbest return Gbest, fbest 3.9. Complexity Analysis of the IPKO Algorithm Algorithmic complexity is a critical indicator of computational feasibility and scalability, commonly evaluated in terms of time complexity (runtime cost) and space complexity (memory usage) [37]. An effective algorithmic enhancement should improve convergence behavior and solution quality without introducing additional asymptotic computational burden beyond that of the baseline method [38]. 3.9.1. Time Complexity Analysis The time complexity of IPKO is analyzed by decomposing the algorithm into the initialization phase and the iterative optimization phase. Initialization Phase The initialization phase employs sinusoidal chaotic mapping combined with opposition-based learning to generate the initial population. For a population size of N in a D-dimensional search space, chaotic sequence generation, opposition solution construction, and fitness evaluation all require linear operations with respect to both N and D. Therefore, the total computational cost of the initialization phase is: Oinit=O(N·D). Main Iterative Phase The main optimization loop is executed for Tmax iterations. In each iteration, the following operations are performed:Position updating through exploration or exploitation strategies;Adaptive Directional Search Mechanism (conditionally triggered);Stochastic binary perturbation-based information fusion;Centroid-based adaptive boundary handling;Fitness evaluation of updated individuals. Position updating through exploration or exploitation strategies; Position updating through exploration or exploitation strategies; Adaptive Directional Search Mechanism (conditionally triggered); Adaptive Directional Search Mechanism (conditionally triggered); Stochastic binary perturbation-based information fusion; Stochastic binary perturbation-based information fusion; Centroid-based adaptive boundary handling; Centroid-based adaptive boundary handling; Fitness evaluation of updated individuals. Fitness evaluation of updated individuals. All position update, perturbation, and repair operations are applied on a dimension-wise basis and involve simple arithmetic computations. Even though some mechanisms (e.g., environmental update and stochastic perturbation) are conditionally activated, their worst-case computational cost per iteration remains linear with respect to population size and dimensionality. Consequently, the per-iteration complexity is bounded by: Oiter=O(N·D). Thus, the total time complexity of the iterative phase is: Oloop=O(Tmax·N·D). Overall Time Complexity By combining the initialization and iterative phases, the overall time complexity of IPKO is given by:(24)Ototal=O(N·D)+O(Tmax·N·D)=O(Tmax·N·D). 3.9.2. Space Complexity Analysis The space complexity of IPKO is primarily determined by the storage requirements of population-related variables. Initialization Phase During initialization, the algorithm stores the population matrix and corresponding fitness values, requiring: Oinit=O(N·D). Main Iterative Phase During the optimization process, IPKO maintains the current population, temporary candidate solutions, and the population centroid vector. The centroid requires O(D) memory, which is negligible compared to the dominant population storage. Hence, the total memory requirement remains: Oloop=O(N·D). Overall Space Complexity Combining both phases, the overall space complexity of IPKO is:(25)Ototal=O(N·D). 3.9.3. Comparative Complexity Discussion Compared with the original PKO algorithm, IPKO exhibits identical asymptotic time and space complexities, namely O(Tmax·N·D) for time and O(N·D) for space. Although IPKO introduces multiple enhancement strategies, including chaotic initialization, adaptive directional search, diversity-preserving perturbation, and centroid-based handling, these components only incur constant-factor computational overhead and do not alter the algorithm’s asymptotic complexity. Therefore, IPKO achieves improved convergence quality and robustness while maintaining the lightweight computational characteristics and scalability of the baseline PKO. 4. Results on Benchmark Functions 4.1. Experimental Setup 4.1.1. Benchmark Functions and Comparative Algorithms To comprehensively evaluate the effectiveness and robustness of the IPKO algorithm in comparison to the original PKO algorithm, this study conducted comparative experiments using all 29 benchmark functions from the CEC2017 test suite across 10, 30, 50, and 100 dimensions. These functions are systematically classified into four distinct groups: (1) Unimodal Functions (F1–F10): These fundamental functions are mainly intended for evaluating the algorithm’s local search capacity and convergence rate. (2) High-Condition-Number Functions (F11–F20): Characterized by notably disparate parameter scales, this category examines the algorithm’s proficiency in dealing with imbalanced and ill-conditioned search landscapes. (3) Composite Multimodal Functions (F21–F25): These functions amalgamate multiple basic functions to mimic intricate, multi-faceted search scenarios and pose a challenge to the algorithm’s ability to evade local optima. (4) Hybrid Composition Functions (F26–F29): By integrating diverse function characteristics, this category offers a rigorous assessment for evaluating the overall global optimization performance and robustness. The performance of the proposed IPKO algorithm was comprehensively compared with that of 11 well-established optimization algorithms, namely the original (PKO), Newton-Raphson Based Optimizer (NRBO) [39], K-means Optimizer (KO) [40], Pelican Optimization Algorithm (POA) [41], Honey Badger Algorithm (HBA) [42], Gradient-Based Optimizer (GBO) [43], Whale Optimization Algorithm (WOA), Grey Wolf Optimizer (GWO), Linear Population Size Reduction SHADE (L-SHADE), Covariance Matrix Adaptation Evolution Strategy (CMA-ES), and Differential Evolution (DE). Although some of these algorithms are newly developed, they have been thoroughly validated in their corresponding sources and exhibit strong competitive capabilities, thus serving as appropriate and challenging benchmarks for comparison. 4.1.2. Parameter Settings and Experimental Environment To ensure a fair, reproducible, and statistically reliable evaluation, all comparative experiments were conducted under identical experimental settings. For all algorithms, the population size was fixed at N=30, and the maximum number of iterations was set to Tmax=500, resulting in an identical upper bound on the total number of function evaluations across methods. Each algorithm was independently executed 30 times on every benchmark function to alleviate the influence of stochastic variability. From these runs, mean value, standard deviation, and average runtime (art) were recorded as performance indicators, providing a comprehensive assessment in terms of accuracy, robustness, and computational efficiency. To guarantee fairness in comparison, all competing algorithms employed their original parameter settings as recommended in the corresponding literature. No problem-specific parameter tuning was performed for any algorithm, ensuring that performance differences arise from algorithmic mechanisms rather than manual optimization. All simulations were implemented in MATLAB R2024a and executed on a 64-bit Windows 11 operating system. The hardware platform consisted of a 13th Gen Intel® Core i7–13700H CPU operating at 2.40 GHz with 16 GB of RAM. To enhance reproducibility, all algorithms were executed under the same software environment and computational resources. All experiments were repeated 30 independent runs per problem instance. Since the performance data of stochastic metaheuristics typically violate normality assumptions, we employ non-parametric statistical tests following standard practice in evolutionary computation. Specifically, the Friedman test is used to detect overall significant differences among multiple algorithms across benchmark functions. When the Friedman test rejects the null hypothesis (p<0.05), we apply Holm’s post-hoc procedure to control the family-wise error rate in pairwise comparisons between IPKO and each competitor. All reported p-values are two-tailed with a significance level α=0.05. 4.2. Comprehensive Performance Evaluation on CEC2017 The comprehensive comparative results of all twelve algorithms across the 29 test functions are presented in the subsequent sections. 4.2.1. Numerical Results and Statistical Analysis The comparative performance of IPKO against 11 state-of-the-art algorithms on the CEC2017 benchmark suite is evaluated across four dimensional settings: 10D, 30D, 50D, and 100D. The numerical results for each dimension are presented in Table 4, Table 5, Table 6 and Table 7, where Table 4 reports 30D results, Table 5 reports 10D results, Table 6 reports 50D results, and Table 7 report 100D results. To statistically validate the overall performance differences, the Friedman test is conducted for each dimension, with results summarized in Table 8, Table 9, Table 10 and Table 11. The corresponding Holm post-hoc test results for each dimension are presented in Table 12, Table 13, Table 14 and Table 15, where the proposed IPKO is compared with other competitor algorithms. Overall Performance Analysis Overall Performance Analysis The numerical results demonstrate that IPKO consistently achieves superior optimization performance across the majority of benchmark functions. In terms of solution accuracy, IPKO obtains the best mean values on 14, 17, 17, and 15 functions in the 10D, 30D, 50D, and 100D cases, respectively. These results indicate that IPKO is capable of producing high-quality solutions across diverse optimization landscapes. Regarding solution stability, IPKO also demonstrates competitive robustness. It achieves the lowest standard deviation on 8 functions (10D), 10 functions (30D), 4 functions (50D), and 3 functions (100D). These consistently low variance values indicate that the algorithm maintains stable convergence behavior and reliable performance across independent runs. Notably, the superiority of IPKO becomes particularly evident on multimodal and hybrid benchmark functions, which typically exhibit complex and challenging search landscapes. This improvement can be attributed to the proposed enhancement strategies, including sine chaotic initialization and stochastic binary perturbation, which increase population diversity and enable the algorithm to escape local optima more effectively. B.Scalability with Increasing Dimensionality Scalability with Increasing Dimensionality Another important observation is that the performance advantage of IPKO remains stable as the problem dimensionality increases. Although optimization difficulty increases significantly when the dimension grows from 10D to 100D, IPKO still maintains a high number of best-performing functions in terms of both mean and standard deviation. For example, even in the 100D scenario, which represents the most challenging case, IPKO still achieves the best mean results on 15 benchmark functions. This demonstrates that the proposed algorithm maintains strong scalability and robustness when dealing with high-dimensional optimization problems. This sustained performance advantage indicates that the proposed mechanisms—particularly adaptive directional search and centroid-based handling—help maintain population diversity and guide individuals toward promising regions while avoiding premature convergence. C.Friedman Statistical Test and Discussion Friedman Statistical Test and Discussion To further verify the statistical significance of the observed performance differences, the Friedman test is conducted across all benchmark functions. The results show highly significant differences among the compared algorithms (p≪0.001). IPKO consistently achieves the lowest average ranks of 2.25, 1.93, 1.64, and 2 for the 10D, 30D, 50D, and 100D cases, respectively, indicating the best overall performance. The ranking results reveal that LSHADE generally maintains competitive positions, while GBO shows improved performance as dimensionality increases. In contrast, the baseline PKO exhibits relatively weaker rankings in higher-dimensional scenarios, highlighting the effectiveness of the proposed enhancement strategies. Further Holm post-hoc tests (α=0.05) show that IPKO significantly outperforms all or the vast majority of the compared algorithms (all adjusted p-values are less than 0.05, with only no significant difference from PKO and LSHADE in a few dimensions). This fully confirms that the proposed improved strategies enable IPKO to achieve a significant and stable advantage in overall optimization performance. Overall, the consistently lowest average ranks and highly significant p-values confirm that the superior performance of IPKO is statistically reliable rather than incidental, demonstrating its strong optimization capability and robustness on the challenging CEC2017 benchmark suite. 4.2.2. Box Plot Behavior Analysis (D = 30) As shown in the box plot analysis (Figure 3 and Figure 4), IPKO generally exhibits competitive performance in terms of both accuracy and robustness compared with the other algorithms. For unimodal functions (e.g., F1 and F3), IPKO achieves relatively low fitness values with a compact distribution, indicating stable convergence behavior. While some algorithms such as NRBO or DE occasionally reach comparable results, several methods (e.g., WOA and GWO) show larger dispersion or higher median values, suggesting less reliable convergence. In multimodal scenarios (e.g., F7, F12, F14, and F19), IPKO maintains a good balance between exploration and exploitation. It consistently ranks among the top-performing algorithms and demonstrates relatively stable distributions. In contrast, certain algorithms such as POA and HBA exhibit wider interquartile ranges or outliers, indicating susceptibility to local optima. For more complex or high-dimensional functions (e.g., F25 and F30), IPKO continues to show strong robustness, with comparatively low variance and competitive median performance. Although some algorithms (e.g., CMA-ES or LSHADE variants) achieve similar or occasionally better results on specific functions, IPKO maintains more consistent behavior across different problem types. Overall, the box plots indicate that IPKO provides a favorable trade-off between solution quality and stability, particularly in complex and multimodal optimization landscapes. 4.2.3. Convergence Analysis (D = 30) The convergence behaviors of all algorithms on the CEC2017 benchmark suite are illustrated in Figure 5. Overall, IPKO demonstrates stable and competitive convergence behavior across diverse function types. On unimodal functions (e.g., F1 and F5), IPKO exhibits faster early-stage convergence than most competing algorithms, indicating that its initialization and adaptive mechanisms effectively guide the population toward promising search regions. For multimodal functions (e.g., F9 and F10), IPKO maintains optimization momentum in later stages, continuing to refine solutions while several baseline methods experience convergence stagnation. The advantage becomes more evident on hybrid functions (e.g., F16–F24), where IPKO consistently reduces fitness values throughout the search process, while comparative algorithms tend to reach convergence plateaus at earlier stages. On highly challenging functions (e.g., F22 and F30), IPKO achieves sustained fitness reduction despite premature convergence observed in several baseline algorithms. Overall, IPKO exhibits strong convergence characteristics in both early-stage search efficiency and late-stage solution refinement. These observations further support the effectiveness of the proposed mechanisms in enhancing both the exploration capability and exploitation efficiency of the algorithm. 4.2.4. Computational Efficiency and Robustness Analysis Beyond solution accuracy, computational efficiency is also an important criterion for evaluating metaheuristic algorithms. This section investigates the efficiency of IPKO from two perspectives: actual runtime and performance under limited function evaluation (FEs) budgets. Runtime Analysis Runtime Analysis The empirical runtime analysis was conducted under the same experimental settings described previously. Table 4 report the average runtime (art) of each algorithm on the 30 CEC2017 benchmark functions, while the overall mean runtime and standard deviation are summarized in Table 16. As shown in Table 16, IPKO requires more runtime than the original PKO due to the additional mechanisms introduced in the proposed framework, including chaotic initialization, adaptive directional search, stochastic perturbation, and centroid-based boundary handling. The mean runtime of IPKO is approximately 2.15 times that of PKO (0.768 s versus 0.358 s). However, these operations mainly consist of population-level vector computations and therefore only introduce constant-factor overhead. Consequently, the asymptotic computational complexity remains unchanged:(26)O(Tmax·N·D) where Tmax denotes the maximum number of iterations, N is the population size, and D is the problem dimension. To further analyze the trade-off between runtime and optimization performance, Figure 6 presents a runtime–accuracy scatter plot, where the horizontal axis represents the average CPU runtime and the vertical axis denotes the number of benchmark functions on which each algorithm achieves the best mean objective value. WOA and LSHADE exhibit relatively low runtime costs; however, their optimization performance remains inferior to that of IPKO on most benchmark functions. Although IPKO incurs a moderate runtime increase compared with lightweight algorithms, it achieves the largest number of best-performing results, indicating a favorable balance between computational cost and solution quality. Compared with CMAES and KO, which also exhibit relatively high computational costs, IPKO provides superior optimization performance while maintaining a comparable runtime level. In particular, KO shows the highest average runtime (1.031 s) without achieving competitive solution quality, whereas IPKO obtains substantially better optimization results with lower computational cost. These observations indicate that the additional computational overhead introduced by IPKO is effectively translated into improved search capability and robustness. Overall, the runtime analysis confirms that IPKO maintains the same asymptotic complexity as other population-based metaheuristics, while the measured runtime increase remains moderate relative to the achieved improvement in optimization performance and stability. B.Performance Under Limited Function Evaluation Budgets Performance Under Limited Function Evaluation Budgets To further evaluate computational efficiency under constrained resources, experiments were conducted on the CEC2017 benchmark suite using FE budgets of {5000,10,000,15,000} under 30D, 50D, and 100D settings. Table 17 reports the corresponding average rankings. The results show that IPKO achieves the best average rank in all settings (1.6–2.5), demonstrating stable performance across different dimensions and evaluation budgets. GBO and LSHADE generally occupy the second tier, whereas NRBO and WOA consistently rank among the weakest algorithms. To quantify how effectively additional evaluations improve optimization performance, a Budget Efficiency Ratio (BER) is defined as(27)BER=f5000−f15,000f5000×100%, where f5000 and f15,000 denote the average fitness values obtained under FE budgets of 5000 and 15,000, respectively. Table 18 reports the geometric mean of final errors and the corresponding BER values for four representative algorithms in 30D and 100D. Among all compared algorithms, IPKO consistently achieves the lowest average errors under both small and large FE budgets. Notably, even at 5k evaluations, the error obtained by IPKO (30D:6.42×102;100D:2.34×104) remains lower than the final errors achieved by most competing algorithms at 15k evaluations, highlighting its strong early-stage search capability. In terms of budget utilization efficiency, IPKO achieves BER values of approximately 66% in 30D and 85% in 100D, indicating that additional evaluations can be effectively exploited to further refine solution quality. The larger BER observed in 100D suggests that IPKO maintains strong scalability in high-dimensional search spaces. GBO and LSHADE also achieve relatively high BER values in 30D (greater than 90%), indicating substantial improvements when additional evaluations are available. However, their final optimization errors at 15k evaluations remain approximately one order of magnitude higher than those of IPKO. In 100D, the BER values of GBO and LSHADE decrease to around 80%, implying reduced search efficiency under higher-dimensional optimization landscapes. By contrast, WOA exhibits the weakest budget utilization capability, with BER values below 30% in both dimensions. Moreover, its optimization errors remain consistently two to three orders of magnitude larger than those of IPKO across all FE budgets, indicating limited suitability for optimization scenarios with strict evaluation constraints. Overall, both the ranking analysis and BER evaluation consistently verify the effectiveness of IPKO under limited FE budgets. IPKO not only achieves superior solution quality during the early search stage, but also maintains stable convergence as additional evaluations become available. Although GBO and LSHADE become more competitive under larger budgets, they still fail to surpass IPKO in overall optimization accuracy. C.Summary Summary The strong efficiency of IPKO under limited computational budgets can be attributed to the complementary roles of its enhancement mechanisms. Chaotic initialization improves early population diversity, adaptive guidance strengthens exploitation toward promising regions, and centroid-based boundary handling reduces ineffective evaluations caused by infeasible solutions. As a result, IPKO achieves a balanced combination of computational efficiency, budget utilization capability, and robust optimization performance. 4.2.5. Parameter Sensitivity Analysis To systematically evaluate the influence of key parameters on the performance of the proposed IPKO algorithm and to justify the parameter configuration adopted in the comparative experiments, a comprehensive sensitivity analysis is conducted on the population size (Pop) and the balance factor (c). The experiments are performed on eight representative functions from the CEC2017 benchmark suite (F1, F3, F7, F10, F13, F20, F23, F27) with a dimensionality of 30, covering unimodal, multimodal, hybrid, and composition optimization problems. All results are averaged over 30 independent runs. One-Dimensional Sensitivity Analysis One-Dimensional Sensitivity Analysis (1)Population Size Population Size To examine the impact of population size, c is fixed at 2.0, while Pop varies from 20 to 80 with an increment of 5. The Friedman rankings across all test functions are summarized in Table 19. The Friedman test yields a p-value of 1.68×10−4, which is significantly lower than 0.05, indicating that the differences among population sizes are statistically significant. As shown in Table 19, Pop = 25 achieves the best average rank (3.00), followed by Pop = 30 (3.75), demonstrating competitive performance. Although Pop = 25 exhibits the best overall ranking, Pop = 30 provides a favorable trade-off between solution quality and computational cost. Therefore, Pop = 30 is selected as the default setting in subsequent experiments to ensure both efficiency and consistency with commonly adopted configurations in the literature. (2)Balance Factor c Balance Factor c To evaluate the sensitivity to the balance factor, Pop is fixed at 30, while c varies from 0.5 to 3.0 with a step size of 0.5. The corresponding Friedman rankings are presented in Table 20. The Friedman test produces a p-value of 0.623 (>0.05), indicating that no statistically significant differences are observed among the tested values of c. This suggests that the proposed IPKO algorithm is relatively insensitive to the choice of this parameter. Among all candidates, c = 2.5 achieves the best average rank (2.50), followed by c = 2.0 (3.25). Considering both performance consistency and stability across different problem types, c = 2.0 is selected as the default setting. This observation further confirms the robustness of IPKO with respect to variations in the balance factor. B.Two-Dimensional Sensitivity Analysis (Pop × c) Two-Dimensional Sensitivity Analysis (Pop × c) To further investigate the combined effect of the two parameters, a full grid search is conducted over Pop∈[20,80] and c∈[0.5,3.0]. The mean fitness values averaged over all test functions are reported in Table 21. In addition, the sensitivity landscape is visualized in Figure 7 using a heatmap representation, which provides an intuitive illustration of the performance distribution over the parameter space. The star marker highlights the globally optimal parameter combination. From both Table 21 and Figure 7, several important observations can be made:The best performance is achieved at (Pop,c) = (25,2.0), yielding the lowest mean fitness value of 3.2587.A clear performance gradient can be observed along the population size axis, where larger populations generally lead to improved optimization results.In contrast, variations along the c−axis are relatively mild, indicating that the algorithm is less sensitive to this parameter. The best performance is achieved at (Pop,c) = (25,2.0), yielding the lowest mean fitness value of 3.2587. The best performance is achieved at (Pop,c) = (25,2.0), yielding the lowest mean fitness value of 3.2587. A clear performance gradient can be observed along the population size axis, where larger populations generally lead to improved optimization results. A clear performance gradient can be observed along the population size axis, where larger populations generally lead to improved optimization results. In contrast, variations along the c−axis are relatively mild, indicating that the algorithm is less sensitive to this parameter. In contrast, variations along the c−axis are relatively mild, indicating that the algorithm is less sensitive to this parameter. These results suggest that the two parameters jointly influence the optimization performance, while the population size plays a more dominant role. Nevertheless, the default configuration adopted in this study, (Pop,c) = (30,2.0), achieves a mean fitness value of 3.2862, resulting in a marginal performance gap of only 0.84% compared to the global optimum. This negligible difference demonstrates that the selected parameter setting is near-optimal. C.Summary of Parameter Settings Summary of Parameter Settings Based on the above analyses, the parameter configuration used in the subsequent experiments is determined as follows:Pop = 30: This setting achieves competitive performance in the one-dimensional analysis while ensuring a fair and unbiased comparison with existing algorithms.c = 2.0: This value provides stable and consistent performance across all test functions and exhibits strong robustness. Pop = 30: This setting achieves competitive performance in the one-dimensional analysis while ensuring a fair and unbiased comparison with existing algorithms. Pop = 30: This setting achieves competitive performance in the one-dimensional analysis while ensuring a fair and unbiased comparison with existing algorithms. c = 2.0: This value provides stable and consistent performance across all test functions and exhibits strong robustness. c = 2.0: This value provides stable and consistent performance across all test functions and exhibits strong robustness. The two-dimensional analysis further confirms that this configuration achieves performance within 0.84% of the global optimum, validating its effectiveness as a well-balanced and reliable default setting. 4.3. Behavioral and Mechanism Analysis 4.3.1. Ablation Study (D = 30) To further examine the contribution of each enhancement strategy, a progressive ablation study was conducted. The experiment settings are the same as the previous experiments. Specifically, four algorithmic variants were constructed: IPKOa (only sine chaotic opposition-based learning), IPKOab (with the addition of the Adaptive Directional Search Mechanism), IPKOabc (further incorporating the stochastic binary perturbation-based information fusion), and the complete IPKO, which additionally includes the centroid-based adaptive boundary handling mechanism. The baseline PKO was also included for comparison. All variants were run 30 independent times on ten representative CEC2017 functions (F3, F4, F5, F8, F9, F10, F14, F18, F24, F28) under 30 dimensions. Quantitative Results Quantitative Results Table 22 presents the mean, standard deviation, and within-function ranking results (lower rank indicates better performance) for all algorithmic variants. In addition, the Wilcoxon signed-rank test at a significance level of 0.05 was employed to evaluate the statistical significance between each variant and the complete IPKO. Several important observations can be drawn from the results:IPKO consistently achieves the best mean fitness values on all benchmark functions, obtaining an average rank of 1.00. This result demonstrates the superior optimization capability and robustness of the proposed complete framework.IPKOabc ranks second overall with an average rank of 2.00. This variant excludes only the centroid-based adaptive boundary handling mechanism. Moreover, no statistically significant difference is observed between IPKOabc and IPKO on five functions (F3, F14, F18, F24, and F28), indicating that the first three enhancement strategies already provide substantial optimization improvements. However, the remaining performance gap on the other functions further confirms the importance of the proposed boundary handling mechanism in maintaining search stability and solution quality.PKO, IPKOa, and IPKOab are all significantly inferior to IPKO on every benchmark function (p<0.05). Their average ranks range from 3.70 to 4.20, which are considerably worse than those of IPKO and IPKOabc. This finding demonstrates that the four proposed strategies complement each other effectively and should be integrated synergistically to achieve optimal performance. IPKO consistently achieves the best mean fitness values on all benchmark functions, obtaining an average rank of 1.00. This result demonstrates the superior optimization capability and robustness of the proposed complete framework. IPKO consistently achieves the best mean fitness values on all benchmark functions, obtaining an average rank of 1.00. This result demonstrates the superior optimization capability and robustness of the proposed complete framework. IPKOabc ranks second overall with an average rank of 2.00. This variant excludes only the centroid-based adaptive boundary handling mechanism. Moreover, no statistically significant difference is observed between IPKOabc and IPKO on five functions (F3, F14, F18, F24, and F28), indicating that the first three enhancement strategies already provide substantial optimization improvements. However, the remaining performance gap on the other functions further confirms the importance of the proposed boundary handling mechanism in maintaining search stability and solution quality. IPKOabc ranks second overall with an average rank of 2.00. This variant excludes only the centroid-based adaptive boundary handling mechanism. Moreover, no statistically significant difference is observed between IPKOabc and IPKO on five functions (F3, F14, F18, F24, and F28), indicating that the first three enhancement strategies already provide substantial optimization improvements. However, the remaining performance gap on the other functions further confirms the importance of the proposed boundary handling mechanism in maintaining search stability and solution quality. PKO, IPKOa, and IPKOab are all significantly inferior to IPKO on every benchmark function (p<0.05). Their average ranks range from 3.70 to 4.20, which are considerably worse than those of IPKO and IPKOabc. This finding demonstrates that the four proposed strategies complement each other effectively and should be integrated synergistically to achieve optimal performance. PKO, IPKOa, and IPKOab are all significantly inferior to IPKO on every benchmark function (p<0.05). Their average ranks range from 3.70 to 4.20, which are considerably worse than those of IPKO and IPKOabc. This finding demonstrates that the four proposed strategies complement each other effectively and should be integrated synergistically to achieve optimal performance. B.Convergence Behavior Convergence Behavior The convergence behaviors of all algorithmic variants are illustrated in Figure 8. It can be observed that IPKOa achieves noticeably faster convergence during the early search stage compared with the original PKO, indicating that the sine chaotic opposition-based learning strategy effectively enhances population diversity and exploration ability. After introducing the Adaptive Directional Search Mechanism, IPKOab exhibits improved convergence speed during the middle stage of the optimization process, reflecting stronger local exploitation capability and improved search guidance. Furthermore, the incorporation of the stochastic binary perturbation-based information fusion strategy in IPKOabc effectively alleviates premature convergence and enables continuous fitness improvement during the later iterations. Among all variants, the complete IPKO consistently demonstrates the fastest convergence rate and achieves the lowest final fitness values. This superior performance can be attributed to the centroid-based adaptive boundary handling mechanism, which improves population feasibility, stabilizes the search dynamics, and enhances the balance between exploration and exploitation. Overall, the progressive performance improvements observed from IPKOa to the complete IPKO clearly demonstrate that each proposed enhancement strategy contributes positively to the optimization process. The quantitative results and convergence analyses jointly verify the effectiveness and rationality of the integrated IPKO framework. 4.3.2. Search Behavior Analysis Population diversity and the balance between exploration and exploitation are two critical factors influencing the effectiveness of metaheuristic optimization algorithms. In the proposed IPKO, both aspects are enhanced through several complementary mechanisms that shape the search dynamics throughout the optimization process. In terms of population diversity, IPKO implicitly maintains diverse search patterns through multiple strategies. During the initialization stage, the sine chaotic opposition-based initialization generates a well-dispersed population, enabling broader coverage of the search space. During the optimization process, the Adaptive Directional Search Mechanism and stochastic binary perturbation introduce dimension-wise and directional variations, continuously injecting new search directions and preventing rapid diversity loss commonly observed in swarm-based algorithms. In addition, the centroid-based handling mechanism preserves population coherence without forcing individuals toward rigid boundaries, thereby avoiding abrupt diversity collapse. Based on the maintained diversity, IPKO further achieves an effective balance between exploration and exploitation across different search stages. In the early stage, exploration is emphasized through chaotic initialization and large-amplitude stochastic updates, enabling rapid traversal of the global search space. As the optimization progresses, the information fusion strategy guides individuals toward high-quality regions while preserving selective randomness at the dimension level, allowing effective solution refinement. Overall, this dynamic search behavior enables IPKO to maintain fast early-stage convergence while preserving refinement capability in later iterations, which explains its strong performance on complex multimodal and high-dimensional optimization problems. 4.3.3. Relationship Between Improvement Strategies and Experimental Performance To visually demonstrate how the proposed enhancement strategies contribute to IPKO’s experimental performance, Figure 9, Figure 10 and Figure 11 illustrate the ranking dynamics of all algorithms under varying computational budgets and dimensional settings. Competitive Advantage Under Limited Budgets Competitive Advantage Under Limited Budgets Figure 9 shows the average ranking dynamics at 30D under increasing FEs budgets. IPKO maintains the lowest rank across all budget levels (1.6→1.7), demonstrating strong early exploration enabled by chaotic initialization and continued refinement from its adaptive mechanisms. In contrast, WOA and NRBO exhibit higher ranks with minimal budget sensitivity, indicating ineffective use of additional evaluations. B.Comprehensive Performance Landscape Comprehensive Performance Landscape The heatmap in Figure 10 presents average rankings across three dimensions and three budgets. IPKO consistently occupies the darkest blue region (lowest ranks) across all twelve configurations, confirming robust performance regardless of dimensionality. LSHADE and GBO emerge as the closest competitors, while PKO consistently lags behind IPKO—directly quantifying the benefit of the proposed enhancements. C.Dimensional Scalability of IPKO Dimensional Scalability of IPKO Figure 11 shows that IPKO’s average rank slightly increases with dimensionality at low FEs (e.g., from 1.7 in 30D to 2.5 in 100D at 5K), reflecting increased optimization difficulty. However, with increasing FEs, the ranking improves consistently and the performance gap across dimensions narrows, demonstrating strong scalability and effective utilization of computational resources. D.Summary Summary Figure 9, Figure 10 and Figure 11 establish a clear link between IPKO’s enhancement strategies and its experimental performance: chaotic initialization enables strong early exploration, adaptive mechanisms maintain search effectiveness across budgets, and continued refinement ensures improvement with increasing evaluations—particularly in high-dimensional spaces. 4.4. Experimental Summary The experimental results on the CEC2017 benchmark suite demonstrate that the proposed IPKO algorithm consistently outperforms eleven state-of-the-art metaheuristic algorithms in terms of solution accuracy, convergence speed, and stability, particularly on high-dimensional and multimodal problems. These performance gains arise from the synergistic interaction of multiple enhancement mechanisms rather than any single modification. Specifically, the sine chaotic opposition-based learning strategy improves the initial population distribution and enhances early-stage exploration, while the Adaptive Directional Search Mechanism strengthens adaptive exploitation. The stochastic binary perturbation strategy effectively maintains population diversity and prevents premature convergence, and the centroid-based handling mechanism stabilizes the search process by preserving feasibility. From a behavioral perspective, these mechanisms collectively establish a dynamic search process in which exploration and exploitation are adaptively coordinated across different optimization stages, leading to both rapid convergence and sustained refinement capability. Furthermore, statistical analyses, including non-parametric significance tests and ranking evaluations, confirm that the observed performance improvements are highly significant and not due to random variation. Overall, IPKO exhibits strong robustness and scalability, demonstrating its effectiveness as a reliable optimization framework for complex benchmark problems. 5. Application of the IPKO Algorithm to Engineering Optimization Problems To further validate the practical applicability and engineering relevance of the proposed IPKO algorithm, this section investigates its performance on two representative real-world optimization problems: cold chain logistics routing and dynamic multi-UAV cooperative path planning. These applications encompass diverse optimization characteristics, including mixed discrete–continuous decision variables, high-dimensional search spaces and strong nonlinearity. Unlike benchmark test functions, these engineering problems require not only superior objective values but also strict feasibility, solution stability, and robustness under complex constraints. Therefore, they constitute a rigorous testbed for evaluating the practical effectiveness of IPKO. Comparative results demonstrate that IPKO consistently outperforms the baseline PKO and several state-of-the-art metaheuristics in terms of solution quality, convergence behavior, and robustness, highlighting its strong potential for real-world engineering decision support. 5.1. Cold Chain Logistics Distribution Optimization 5.1.1. Problem Definition and Dataset The cold chain logistics routing problem considered in this study is defined as follows. A single distribution center located at coordinates (15,39) dispatches refrigerated vehicles to serve 15 customer demand points. Each vehicle has a maximum load capacity of 5 t and a maximum driving range of 180 km. Each customer is associated with a specific demand, a soft time window, and a hard time window. Vehicles must arrive within the hard time window; otherwise, a large penalty is imposed. If a vehicle arrives earlier than the allowable service time, waiting is permitted until service can begin. During transportation, fresh products deteriorate continuously over time and additionally during unloading operations caused by door opening. Meanwhile, the refrigeration system continuously consumes energy and produces carbon emissions. The customer data (coordinates, demand, and time windows) are taken from the illustrative example in [44] and are summarized in Table 23. The departure time from the depot is set to 5:00, the average vehicle speed is 45 km/h, and the unloading efficiency is 4 t/h. Other experimental parameters are listed in Table 24. 5.1.2. Mathematical Model Decision variables: A real vector of length 2Nc (Nc=15). The first Nc components are sorted in descending order to determine the customer visit sequence; the last Nc components are linearly mapped to [1,K] (with K=3 available vehicles) to assign each customer to a vehicle. Objective function: Minimize the total cost f, which consists of fixed cost F1, transportation cost F2, cargo damage cost F3, refrigeration cost F4, transport carbon cost F5, refrigeration carbon cost F6, time-window penalty F7, and penalty terms for load/mileage violations (limit1, limit2):f=F1+F2+F3+F4+F5+F6+F7+limit1+limit2. Each component is defined as follows:Fixed cost: F1=∑j=1Kδj·p1, where δj=1 if vehicle j is used.Transportation cost: F2=∑j=1KDj·p2, where Dj is the total distance (km) traveled by vehicle j.Cargo damage cost:F3=p3∑j=1K∑i∈pathjQnowarf1·(Ti−t0)+arf2·qiω, where Qnow is the current load before serving customer i (t), Ti is the arrival time (h), qi is the demand (t), and ω=4 t/h.Refrigeration cost: The refrigeration system is driven by the vehicle engine. The instantaneous cooling powers (in watts) are first calculated and then converted to kilowatts:Ptrans=(1+ϕ)RSwSn(Tw−Tn)1000[kW],Popen=u(0.54V+3.22)(Tw−Tn)1000[kW].The total refrigeration energy consumption (kWh) is:Eref=∑j=1KPtrans·Djv+∑i∈pathjPopen·qiω.Let SFC be the specific fuel consumption of the refrigeration unit (L/kWh). In this study, SFC=0.28 L/kWh (typical for diesel-driven refrigeration units). The diesel fuel consumed for cooling is Fuelref=SFC·Eref. Then the refrigeration cost is:F4=pfuel·Fuelref, where pfuel is the diesel price (yuan/L), set to 7.5 yuan/L in this study.Transport carbon emission cost:F5=p5e1∑j=1K∑i∈pathjρ1+ρ2−ρ1Q2Qnowdi,i+1, where di,i+1 is the distance between consecutive customers (or depot). The term in parentheses is the fuel consumption rate (L/km), and e1 (kg/L) is the CO2 emission factor of diesel.Refrigeration carbon emission cost: The carbon emissions from refrigeration are proportional to the diesel consumed:F6=p5·e2·Fuelref, where e2 (kg/L) is the CO2 emission factor of diesel used for refrigeration. For consistency with the transport carbon factor, we set e2=e1=2.63 kg/L.Time-window penalty:F7=∑i=1Nc105Tihard_endi100×(soft_starti−Ti)hard_starti≤Tihard_endi100×(soft_starti−Ti)hard_starti≤Tihard_endi100×(soft_starti−Ti)hard_starti≤Ti