{"title":"A Q-learning enhanced multi-objective memetic algorithm for vehicle routing with simultaneous pickup-delivery and time windows","authors":"Xueyu Lai, Yalan Zhou","doi":"10.1016/j.swevo.2026.102582","DOIUrl":"10.1016/j.swevo.2026.102582","url":null,"abstract":"<div><div>The vehicle routing problem with simultaneous pickup and delivery and time windows (VRPSPDTW) is an NP-hard combinatorial optimization problem. In practical logistics settings, routing decisions need to balance multiple conflicting criteria. Accordingly, a multi-objective VRPSPDTW (MOVRPSPDTW) is formulated to optimize operating cost, makespan, and customer satisfaction under capacity and hard time-window constraints. A Q-learning enhanced multi-objective memetic algorithm (QMMA) is proposed for MOVRPSPDTW. Q-learning-based local search is employed to adapt the sequence of local-search operators during the search. A two-stage move evaluation procedure screens candidate moves using route attributes before exact evaluation of the full three-objective vector. Computational experiments are conducted on the WC, JD, and package-delivery sets under fixed time budgets, and performance is assessed using hypervolume (HV) and inverted generational distance plus (IGD<span><math><msup><mrow></mrow><mrow><mo>+</mo></mrow></msup></math></span>). QMMA achieves the best or joint-best mean HV on 91 out of 96 instances and the best or joint-best mean IGD<span><math><msup><mrow></mrow><mrow><mo>+</mo></mrow></msup></math></span> on 87 out of 96 instances. The ablation results show that the local search procedure and Q-learning-guided operator selection contribute to the performance of QMMA. These results indicate that QMMA can produce high-quality nondominated approximation sets for MOVRPSPDTW.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102582"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354239","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Distribution-guided evolution and tail-adaptive search for robust portfolio optimization","authors":"Tanuja Mathpal, Arun Kumar","doi":"10.1016/j.swevo.2026.102581","DOIUrl":"10.1016/j.swevo.2026.102581","url":null,"abstract":"<div><div>The financial markets often experience continuous shifts in distributions, exhibit fat-tailed return patterns, and have continuously changing risk structures, which make traditional portfolio optimization methods less effective. Many-objective evolutionary algorithms (MOEAs) have proven to perform well in handling multiple investment goals, but their search mechanisms are mainly based on information in the objective space and are, therefore, not very sensitive to return distribution variations. Thus, solutions that have emerged in the past under historical market conditions could be less robust if market regimes change. To overcome this drawback, this paper introduces a distribution-sensitive and tail-aware evolutionary framework named Distribution-Guided Non-dominated Sorting Genetic Algorithm III (DG-NSGA-III) for robust portfolio optimization. The proposed framework introduces Wasserstein distance-based adaptation into the evolutionary search process, which allows the algorithm to adapt to distribution shifts and direct exploration based on observed distributional shifts in return data. The system also incorporates tail-risk characterization and adaptive behavioral stock filtering to handle extreme risk exposure and changing asset quality, improving the system's robustness. The effectiveness of the suggested approach is tested by rolling walk-forward experiments and compared with classical allocation techniques and state-of-the-art evolutionary optimization baselines. Further validation is carried out by statistical significance tests, ablation analysis, convergence assessment, and cross-market robustness evaluation in the NIFTY 50 equity universe and S&P 500 equity universe. The out-of-sample evaluation results achieved by the proposed method are a 25.02% return, a Sharpe ratio of 1.003, a Sortino ratio of 1.419, and an LOR of 0.775. Experimental results show enhanced robustness, downside risk management, and stability within non-stationary market environments.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102581"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354241","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Thin plate spline-based prediction mechanism for dynamic constrained multi-objective optimization","authors":"Hongbiao Tian, Baosheng Li, Xiaochuan Gao, Qianlong Dang, Zhiyang Zhang","doi":"10.1016/j.swevo.2026.102598","DOIUrl":"10.1016/j.swevo.2026.102598","url":null,"abstract":"<div><div>Dynamic constrained multi-objective optimization problems (DCMOPs) have received significant attention due to their wide applications in the real world. However, many dynamic constrained multi-objective evolutionary algorithms (DCMOEAs) struggle to capture dynamic constrained Pareto solution sets during the prediction stage, and constraints cause a delayed response. To address these issues, this paper proposes a DCMOEA with thin plate spline-based prediction mechanism (TPSEA). Specifically, this algorithm establishes a mapping relationship between the solution sets found in adjacent but different environments, which is used to predict unknown populations based on existing historical populations. During the prediction stage, the prediction mechanism adopts the mapping relationship in the objective space to determine a reasonable initial population for the next environment. To avoid discarding well-predicted solutions that may be filtered out during population maintenance, thin plate spline-based fitting is applied to preserve the structural features of the manifold as much as possible. Additionally, a maintenance strategy based on angular diversity is designed to preserve population diversity. This strategy ensures the rationality of population matching during the prediction stage and preserves the manifold structure during the fitting process. Comparison experiments on two benchmark test suites with 19 DCMOPs demonstrate that TPSEA is competitive with seven state-of-the-art DCMOEAs.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102598"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354243","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Yunhe Wang, Zhengyu Du, Li Wang, Qiqi Liu, Qingda Chen, Yaochu Jin, Haoyu Zhang
{"title":"Surrogate-assisted multi-population evolutionary feature selection for high-dimensional classification","authors":"Yunhe Wang, Zhengyu Du, Li Wang, Qiqi Liu, Qingda Chen, Yaochu Jin, Haoyu Zhang","doi":"10.1016/j.swevo.2026.102593","DOIUrl":"10.1016/j.swevo.2026.102593","url":null,"abstract":"<div><div>Existing surrogate-assisted evolutionary algorithms for high-dimensional feature selection often face two key challenges: the risk of getting trapped in local optima due to a loss of solution diversity, and the high computational cost arising from the vast feature selection search space. To address these issues, we propose a Surrogate-assisted Multi-population Evolutionary algorithm for high-dimensional Feature Selection, named SME-FS. In SME-FS, a surrogate model is initially constructed in the high-dimensional feature selection space to alleviate the computational burden. Guided by the surrogate’s predictions, a multi-population particle swarm optimization algorithm is then used to maintain solution diversity in the large search space. Specifically, knowledge sharing across sub-populations is facilitated by a voting-based global best fusion strategy, helping reduce the risk of premature convergence. To further reduce the computational cost of high-dimensional surrogate modeling, the surrogate is retrained only when the global best solutions of all sub-populations fail to improve during the same iteration. SME-FS is evaluated across 14 real-world, high-dimensional datasets with up to nearly 50,000 features and is benchmarked against ten feature selection algorithms. The experimental results demonstrate that SME-FS achieves a favorable trade-off among classification accuracy, computational efficiency, and subset size, with the shortest runtime in all available comparisons with the evolutionary baselines. The selected subsets retain 2.87%–10.81% of the original features, providing compact representations relative to the original feature spaces. We further apply SME-FS to the high-dimensional ALL4 leukemia dataset. Analysis of a 422-gene subset characterizes the selected features through model-specific importance and exploratory functional annotation.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102593"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354244","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"A bi-level reinforcement learning-driven adaptive genetic algorithm for dynamic flexible job shop scheduling in complex product manufacturing","authors":"Qingyang Jia, Jiaqing Li, Yajie Dou, Ziyi Chen, Lining Xing, Kewei Yang","doi":"10.1016/j.swevo.2026.102576","DOIUrl":"10.1016/j.swevo.2026.102576","url":null,"abstract":"<div><div>This paper addresses a dynamic flexible job shop scheduling problem (DFJSP) with new job arrivals and job cancellations, which commonly arises in the manufacture of complex products such as aero-engines, satellites, and missiles. The objective is to minimize the sum of total weighted tardiness cost and rescheduling cost. A bi-level DDQN-QLAGA algorithm is used to solve this problem. The upper level decides when to reschedule, and the lower level decides how to reschedule. At the upper level, a double deep Q-network (DDQN) is used to learn a rescheduling trigger policy, which balances tardiness cost and rescheduling cost based on the real-time shop-floor state. At the lower level, a Q-learning-driven adaptive genetic algorithm (QLAGA) is designed, in which Q-learning is used to select among three search strategies during the evolutionary process. Each strategy combines coordinated settings for the crossover rate, mutation rate, genetic operators, and local search intensity, so that the search behavior remains consistent in each stage. Finally, a series of numerical experiments demonstrates the effectiveness of the proposed algorithm in solving the DFJSP.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102576"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354303","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"GANCS: GAT-enhanced PPO warm-starting of NSGA-III for energy- and SLA-aware cloud workflow scheduling","authors":"Amine Chraibi","doi":"10.1016/j.swevo.2026.102565","DOIUrl":"10.1016/j.swevo.2026.102565","url":null,"abstract":"<div><div>Scientific workflows executed in heterogeneous clouds require schedules that balance completion time, energy consumption, and deadline compliance. Population-based optimizers can represent these trade-offs, but random initialization may waste early generations on poor task-to-virtual-machine (VM) assignments. This study tests whether a graph-based reinforcement-learning policy can provide a better initial population for three-objective evolutionary scheduling. GANCS couples graph-attention-network (GAT)-encoded proximal policy optimization (PPO) with the non-dominated sorting genetic algorithm III (NSGA-III). Stochastic policy rollouts generate VM-assignment chromosomes that seed Pareto search over makespan, energy, and service-level agreement (SLA) violation. Experiments comprised ten paired replications on three synthetic Pegasus-inspired workflow families, supplemented by tighter-deadline, ablation, training-budget, transfer, and workflow-scale evaluations. Warm starting reduced generations-to-90%-hypervolume (HV) from 4.2 ± 1.2 to 1.8 ± 1.7 on CyberShake-30 (<span><math><mrow><mi>p</mi><mo>=</mo><mn>0.027</mn></mrow></math></span>) and usually reached the threshold before evolution on Montage-25 (<span><math><mrow><mi>p</mi><mo>=</mo><mn>0.002</mn></mrow></math></span>), without a detectable final-HV loss. Under tighter deadlines, the SLA objective became active, but GANCS did not significantly differ from the Pareto baselines after multiplicity correction. The backward channel did not accelerate convergence, supporting a simpler primary configuration without behavior cloning. GANCS also remained Pareto-competitive on configurations with 100 and 121 realized tasks. Overall, the results show that policy-seeded evolutionary search can reduce early search effort while preserving final Pareto quality across the tested workflows and scales.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102565"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354242","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Non-linear search-space transformations break scale invariance in bounded particle swarm optimization","authors":"Jörg Kunkelmoor, Paolo Mercorelli","doi":"10.1016/j.swevo.2026.102579","DOIUrl":"10.1016/j.swevo.2026.102579","url":null,"abstract":"<div><div>Normalizing the search space is standard practice in PSO implementations, and it is usually justified by the claim that the algorithm is invariant under coordinate transformation. That claim holds for affine maps only. We show that non-linear maps change convergence behavior under an unchanged objective, and we trace the effect to the resolution geometry of a transformation: the local derivative <span><math><mrow><mi>d</mi><mi>T</mi><mo>/</mo><mi>d</mi><mi>ξ</mi></mrow></math></span>, which sets where in the bounded domain a swarm can resolve fine detail. A transformation should therefore favor optima that lie where it concentrates resolution, and fail where it coarsens it. We test this on a constructed mirror pair, two Sphere landscapes that differ only in the position of the global optimum. Under absorbing boundary handling, one transformation reaches machine precision on the first landscape and misses the second by five orders of magnitude. The asymmetry survives a control for initialization density. A full-factorial study of four transformations, two boundary handling strategies (clamping, rebound) and three reseeding strategies (none, soft, hard) over 12 benchmark functions, 5 dimensionalities and 50 independent runs each, 72,000 runs in total, shows that the three factors are not separable. Rebound handling recovers every transformation that clamping causes to fail, so bound-handling results obtained under linear scaling do not carry over to transformed spaces. Soft reseeding degrades performance on all but one function. Since the optimum location is unknown in practice, we recommend <span>norm01+rebound+hard</span> as a default.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102579"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354235","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Lei Yue, Yuhao Zhu, Dan Luo, Leilei Meng, Linshan Ding
{"title":"Preference-conditioned graph reinforcement learning for dynamic integrated scheduling of AGVs and unrelated parallel machines with finite buffers","authors":"Lei Yue, Yuhao Zhu, Dan Luo, Leilei Meng, Linshan Ding","doi":"10.1016/j.swevo.2026.102589","DOIUrl":"10.1016/j.swevo.2026.102589","url":null,"abstract":"<div><div>Dynamic integrated scheduling of multiple automated guided vehicles (AGVs) and unrelated parallel machines is challenging because transportation, processing, finite-buffer blocking, and conflicting operational objectives must be coordinated under a variable-size action space. This study proposes PC-TAG-MAPPO, a preference-conditioned task-augmented graph multi-agent proximal policy optimization method for jointly minimizing total order tardiness and controllable energy consumption. Feasible dispatch and pickup operations are represented as task nodes in a heterogeneous graph, which turns the combinatorial scheduling decision into graph-based task selection. Machine assignment and processing-speed choices are embedded in dispatch-task nodes. A preference-conditioned actor–critic architecture with objective-specific reward channels and value heads enables a single policy to serve different trade-offs between tardiness and energy consumption, and attention-based coordination combined with sequential task claiming yields conflict-free joint actions for simultaneously idle AGVs. Across 36 benchmark cases covering three problem scales, three due-date tightness levels, and two speed-scaling scenarios, PC-TAG-MAPPO improves the average normalized hypervolume by 29.5% and reduces the average inverted generational distance by 31.9% relative to the strongest deep reinforcement learning baseline, and by 70.1% relative to the best priority dispatching rule. Ablation and statistical analyses support these improvements. The method maintains reliable order completion with average decision times below 10 ms, supporting real-time multi-objective scheduling.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102589"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354238","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"Evolutionary memory-guided co-evolution of search and reasoning for manufacturing value chain scheduling","authors":"Wenjun Xu, Shuang Zheng, Kaipu Wang, Xun Ye, Zhanzheng Xiao, Yue Li, Jianyuan Xie, Yixin Zhang","doi":"10.1016/j.swevo.2026.102590","DOIUrl":"10.1016/j.swevo.2026.102590","url":null,"abstract":"<div><div>In customized manufacturing environments, procurement decisions, assembly operations, and order completion are closely coupled, making it difficult to reconcile local efficiency with global coordination. This study investigates integrated procurement and assembly scheduling with order-based dispatch, jointly optimizing supplier selection and assembly scheduling. The objectives are to minimize procurement costs, finished-product inventory holding costs, and order tardiness penalties, with each order dispatched only after all its products are completed. To address this scheduling task, we propose an Evolutionary Memory-Centered Agent-Enhanced Evolutionary Approach (AEEA). The framework couples evolutionary search with contextual reasoning by a fixed large language model (LLM). Across generations, selected elite chromosomes and their evaluated three-objective vectors are retained as evolutionary memory. A subset of these records forms the prompt context used by the LLM to generate new candidates. A dual-path strategy combines LLM-generated candidates with evolutionary offspring, while constraint checking and repair address infeasible candidates. Experiments in a representative CubeSat-inspired manufacturing setting compare AEEA with the Non-dominated Sorting Genetic Algorithm III, Multi-objective Evolutionary Algorithm Based on Decomposition, Optimization by PROmpting, and the Multi-Agent Large Language Model-Driven Evolutionary Framework. The results indicate improved solution quality and convergence in the evaluated cases. These results highlight the potential of evolutionary memory to bridge evolutionary search and LLM-based contextual reasoning in complex integrated manufacturing scheduling.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102590"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149364151","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
{"title":"ChiSao: A fully GPU-resident optimizer for cataloguing the modes of multimodal black-box functions","authors":"Ira Wolfson","doi":"10.1016/j.swevo.2026.102574","DOIUrl":"10.1016/j.swevo.2026.102574","url":null,"abstract":"<div><div>Bayesian inference and scientific computing often call for cataloguing all modes of a multimodal black-box objective, not just its global optimum. Classical multimodal and niching methods keep their search strategy on the host, parallelizing at most one generation, and pay a per-generation cost, an <span><math><mrow><mi>O</mi><mrow><mo>(</mo><msup><mrow><mi>d</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow></mrow></math></span> covariance update or <span><math><mrow><mi>O</mi><mrow><mo>(</mo><msup><mrow><mi>N</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow></mrow></math></span> pairwise distances. We present <span>ChiSao</span>, a memetic multimodal optimizer whose entire search strategy resides on the device: seeding, batched L-BFGS convergence, deduplication, reseeding, smoothing, and anti-convergence all run on the GPU with no expensive host round-trips. To our knowledge, it is the first fully GPU-resident multimodal optimizer. GPU throughput rather than dimension becomes the limiting factor: the per-evaluation cost stays <span><math><mrow><mi>O</mi><mrow><mo>(</mo><mi>d</mi><mo>)</mo></mrow></mrow></math></span> against the baselines’ <span><math><mrow><mi>O</mi><mrow><mo>(</mo><msup><mrow><mi>d</mi></mrow><mrow><mn>2</mn></mrow></msup><mo>)</mo></mrow></mrow></math></span> to <span><math><mrow><mi>O</mi><mrow><mo>(</mo><msup><mrow><mi>d</mi></mrow><mrow><mn>3</mn></mrow></msup><mo>)</mo></mrow></mrow></math></span>, a full polynomial order lower on identical hardware. On a single NVIDIA RTX 3080, <span>ChiSao</span> solves <span><math><mrow><mi>d</mi><mo>=</mo><mn>512</mn></mrow></math></span> in about twenty seconds where restart CMA-ES needs over four minutes, and recovers shifted optima to <span><math><mrow><mi>d</mi><mo>=</mo><mn>2048</mn></mrow></math></span>, a regime the host-resident baselines cannot enter. On single-global recovery, a memetically fair, L-BFGS-polished restart CMA-ES matches <span>ChiSao</span> at equal budget, and dedicated niching methods lead on rotated-composition landscapes. However, recovery quality is not the claim; the contribution is the device-resident regime itself, in which cataloguing coverage scales with a GPU-sized population at high dimension, plus an on-device configuration portfolio at constant-factor cost. We map reach and limits across the 42 Simon Fraser University (SFU) functions, the CEC’2013 niching suite, a dimensional multi-Ackley benchmark, and a noisy landscape. <span>ChiSao</span> is released as open-source.</div></div>","PeriodicalId":48682,"journal":{"name":"Swarm and Evolutionary Computation","volume":"110 ","pages":"Article 102574"},"PeriodicalIF":9.6,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149354236","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}