The irregular topology and high dimensionality of the search landscape frequently hinder data clustering in high-dimensional spaces. Although metaheuristic algorithms have significant potential, most use inflexible update strategies that cannot adapt to the specifics of the dataset, leading to ineffective refinement. The paper proposes a framework called Multi-Strategy Surrogate Optimisation (MSSO), in which the evolving population is treated as a dynamic surrogate of the objective space. MSSO is a derivative-free mechanism that dynamically adjusts the search radius based on local landscape regularity to stabilise convergence. MSSO combines data-driven exploration, hybrid attraction, and surrogate-guided refinement in a multi-phase optimisation scheme. MSSO achieves directed search by inferring local descent trajectories directly from population spatial interactions via local linear surrogates, combined with a derivative-free adaptive search radius mechanism. Evaluation against nine state-of-the-art metaheuristics, i.e., Particle Swarm Optimization (PSO), Whale Optimisation Algorithm (WOA), Differential Evolution (DE), Opposition Based Harris-Hawks Optimisation (OHHO), Harris-Hawks Optimisation (HHO), Sine Cosine Algorithm (SCA), Salp Swarm Algorithm (SSA), Grey Wolf Optimizer (GWO), and Butterfly Optimisation Algorithm (BOA) across 11 datasets demonstrates that MSSO achieves a superior average rank of 1.91. Statistical analysis confirms that MSSO achieves statistically significant improvement over six of nine compared baselines while maintaining competitive performance with the top three methods.
The symmetric and asymmetric traveling salesman problems require different local-search moves. Segment reversal is valid for symmetric costs, but in ATSP it changes the direction and cost of every internal arc. This paper presents a topology-aware dual-mode adaptive hybrid genetic algorithm (DM-AHGA) that estimates matrix asymmetry and selects an STSP or ATSP search track. Both tracks use collaborative pheromone induction (CPI) to reuse edges or arcs from elite tours and a gain-driven triggering mechanism (GDTM) to decide when offspring receive additional local search. The STSP track combines candidate-set 2-opt with pheromone-guided and bad-edge refinement. The ATSP track uses order-preserving relocation during evolution and VND/ILS refinement afterward. Experiments on 23 STSP and 14 ATSP instances from TSPLIB, together with synthetic Euclidean instances, show that DM-AHGA obtains consistently low optimality gaps on STSP and remains competitive with strong ATSP baselines under matched wall-clock budgets. Ablation results show complementary effects of CPI and GDTM and further improvements from terminal ATSP refinement. Overall, the results support selecting evolutionary operators according to distance-matrix symmetry.
Multi-robot path planning under partial observability requires each robot to infer useful spatial information from limited local observations while coordinating with teammates. We present Spatial Policy-Conditioned Actor-Critic (SPARC), a centralized-training/decentralized-execution framework that combines an egocentric spatial actor with per-agent conditional twin critics. The actor preserves local-grid topology through convolution, spatial self-attention, recurrent state modeling, and an explicit goal-direction tensor. During training, each critic estimates values for all candidate actions of the focal agent while conditioning on the global state and sampled teammate-action contexts; the update exactly sums over the nine focal actions and samples only the combinatorial teammate action space. Experiments on benchmark scenarios demonstrate that SPARC achieves higher task returns and lower collision rates compared with QMIX, MAPPO, and IPPO. Ablation studies provide descriptive support for the contributions of spatial representation, attention mechanisms, goal information integration, dynamic-obstacle history, and teammate-conditioned value estimation. Overall, SPARC improves cooperative multi-robot path planning performance under partial observability.
Atom search optimization (ASO) is a newly developed metaheuristic technique inspired by basic molecular dynamics. However, ASO employs a fixed parameter control scheme and a single learning strategy, which results in insufficient local exploitation capability and makes it susceptible to premature convergence when tackling complex multimodal problems. To alleviate these drawbacks, a competitive learning-based ASO with parameter adaptation (CLASO-PA) is proposed in this work. Specifically, in CLASO-PA, a parameter adaptive mechanism is first introduced to dynamically adjust the control parameters according to the evolutionary state of the population, thereby achieving a better balance between exploration and exploitation. Subsequently, a competitive learning strategy driven by historical population information is developed to improve convergence performance while maintaining population diversity. On this basis, a competitive perturbation mechanism is designed to escape from local optima. The effectiveness of CLASO-PA is evaluated on 42 benchmark functions against ASO and 23 well-established metaheuristic algorithms, and further validated on multi-robot path planning problems. Experimental results demonstrate that CLASO-PA significantly outperforms its competitors in terms of solution accuracy, convergence speed, and statistical rankings. The source code of CLASO-PA is publicly available at https://github.com/lqhgxgl/CLASO-PA .
With rising environmental concerns and fossil fuel depletion, electric vehicles (EVs) have emerged as a promising alternative for sustainable transportation; however, selecting the most suitable EV remains challenging due to multiple conflicting criteria and uncertainty in expert evaluations. Since criterion weights can significantly influence the final decision, relying solely on subjective weighting may introduce bias due to differences in expert preferences and judgments. Therefore, along with subjective weighting, an objective weighting mechanism is needed to derive the importance of criteria. To address this issue, an entropy measure is used to determine objective criterion weights, while SWARA-based subjective weights and the integrated weights are considered to examine the effect of different weighting perspectives on EV ranking through q-ROFS-COPRAS. The entropy-derived weights are further compared with existing q-ROFS entropy measures and employed in q-ROFS-TOPSIS, q-ROFS-VIKOR, q-ROFS-WASPAS, q-ROFS-ARAS, and q-ROFS-MABAC to examine the consistency of the resulting rankings. The q-ROFS-COPRAS method yields the ranking E_2 ≻ E_4 ≻ E_6 ≻ E_1 ≻ E_3 ≻ E_5 , which is identical to the rankings obtained using q-ROFS-TOPSIS, q-ROFS-MABAC, and q-ROFS-WASPAS, indicating high degree of consistency among these methods. Furthermore, the applicability of the entropy-based framework is demonstrated through an existing EV selection case study, where the obtained results show substantial agreement with those reported in the literature. Sensitivity analysis for q ranging from 2 to 10 yields an unchanged final ranking, indicating the stability of the decision outcome with respect to variations in the q-rung parameter.
