ABSTRACT Rapid urban development and climate change reveal limits of traditional construction, highlighting a need to integrate adaptive and innovative technologies in architecture. A systematic review of sustainable projects and technologies connects architectural strategies with remote sensing data. Case studies, including Malakoff, Tianbao, Roskilde, and Macau, assess spatial, structural, and functional flexibility to evaluate adaptive, energy-efficient solutions and long-term environmental performance. Key principles of adaptive design are highlighted, encompassing genetic algorithms, parametric methods, machine learning, intelligent and bio-inspired systems, AI-based technologies such as micro-GAs (exemplified by a building façade in Boston, U.S.A.), evolutionary computation, artificial neural networks, generative design, ML algorithms, digital twins for urban forecasting, a bio-inspired ventilation system at the Eastgate Centre in Harare, Zimbabwe, and energy-efficient façade and spatial systems at the Vancouver Convention Centre West in Canada, while factors limiting widespread application of these intelligent technologies are identified. Achieving sustainable architecture goals requires a comprehensive approach integrating intelligent and adaptive systems tailored to specific conditions, thereby creating harmonious buildings that meet societal demands and support environmental sustainability.

This paper presents Monkey Jumping Optimization (MJO), a novel nature-inspired metaheuristic algorithm that emulates the arboreal locomotion behavior of monkeys to address complex global optimization problems. The MJO algorithm integrates three biologically inspired mechanisms: energy-aware leap dynamics, probabilistic branch selection, and canopy memory preservation to balance exploration and exploitation within multimodal search spaces. By representing candidate solutions as monkeys navigating a virtual tree structure, MJO employs a population-based framework that combines stochastic and deterministic strategies for efficient traversal of the solution landscape. Extensive benchmarking against eight state-of-the-art algorithms, including Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Grey Wolf Optimizer (GWO), demonstrates competitive performance under standard benchmarking assumptions. Specifically, MJO achieves up to 28.7% faster convergence under standard experimental settings compared to PSO on challenging deceptive functions and attains 15-22% higher success rates in identifying global optima across the Congress on Evolutionary Computation (CEC 2024) benchmark suite. Beyond empirical evaluation, theoretical convergence properties are analyzed using a Markov chain framework. The algorithm's biologically inspired design, combined with computational efficiency, makes it suitable for engineering applications such as unmanned aerial vehicle (UAV) path planning and neural architecture search. Despite its strengths, MJO may exhibit relatively slower convergence on well-conditioned unimodal problems and shows sensitivity to parameter settings, highlighting opportunities for further refinement. The proposed MJO algorithm, accompanied by an open-source implementation, provides a flexible and extensible framework for solving complex optimization problems across diverse domains.

Differential evolution has become one of the mainstream solvers for complex optimization problems due to its concise structure and strong global search ability. However, the performance of the DE algorithm is highly sensitive to its mutation and crossover strategies and related control parameters. Traditional adaptive strategies are mostly based on heuristic rules and lack online learning capabilities, which limits their further application in expensive black box optimization scenarios. Therefore, this article proposes Reinforcement Learning-controlled Differential Evolution with Limited-memory Broyden-Fletcher-Goldfarb-Shanno Refinement (RL-DE). The algorithm has constructed a three-layer closed-loop framework of “parameters policy refinement”, achieving a paradigm shift from rule driven to data-driven. The system evaluation on the Congress on Evolutionary Computation 2017 (CEC2017) benchmark test set shows that the algorithm is robust in high-dimensional scenarios and achieves better results compared to classical adaptive DE variants; Its successful application in flexible job shop scheduling problems also validates the generalization ability of the framework towards the field of discrete combinatorial optimization. This study provides a learnable unified architecture for parameter adaptation in DE, offering efficient and intelligent solutions for expensive black box optimization and engineering scheduling problems.

This paper proposes a multi-strategy improved pied kingfisher optimizer (MSIPKO), a novel metaheuristic algorithm designed to address constrained optimization problems (COPs). COPs are widely encountered in engineering and industrial applications and are characterized by complex constraints that restrict the feasible solution space and often lead to multiple local optima. To enhance the performance of the original pied kingfisher optimizer (PKO), three strategies are incorporated: (i) a reverse differential crossover mechanism to improve global exploration and maintain population diversity; (ii) an enhanced diving-fishing operator to strengthen local exploitation; and (iii) an improved commensalism phase to enrich search directions and increase robustness. The performance of MSIPKO is evaluated on 12 benchmark functions from the IEEE Congress on Evolutionary Computation 2006 (CEC 2006) test suite and six classical engineering optimization problems. Experimental results demonstrate that MSIPKO outperforms several state-of-the-art algorithms in terms of optimization accuracy, convergence speed, and stability, particularly for high-dimensional, nonlinear, and multi-constrained problems. Moreover, MSIPKO achieves superior or comparable solutions with fewer function evaluations, indicating its high efficiency and adaptability. These results confirm that MSIPKO is a promising tool for solving complex real-world constrained optimization problems. Future work will focus on extending the proposed algorithm to multi-objective and large-scale optimization scenarios.

Two genetically close filamentous Aspergillus species, the non-toxic A. oryzae used for food fermentation for thousands of years (e.g., koji mold, sake, soy sauce) playing huge role in food industry, and the main aflatoxin producer of A. flavus are studied here based on genetic analysis in silico. The evolutionary TimeLine computer program of Ascomycota fungi showed 520 MYA origins. Phylogenetic tree of genome sequences of Ascomycota fungi was found to group in three main clades. The genome size of A. flavus (37.7 x106 bpDNA) but not A. sojae (41.1 x106 bpDNA) showed smaller size than the descendent (90-100 MYA) A. oryzae (37.9 x106 bpDNA). In phylogeny of mitogenomes (mtDNA) A. terreus showed a distinct clade, and the toxic A. parasiticus grouped to the non-toxic A. sojae. Fungal production of aflatoxins, cellulases, and the ribosomal post-translationally modified peptides (RiPPs) of nisin, asperigimycin, gliotoxin; the fungal non-ribosomal peptide (NRPs) fusahexin; the polyketide taxane and mycolactone are reviewed and discussed.

Multitasking optimization (MTO), addressing multiple optimization problems synchronously, has achieved significant success in the field of evolutionary computation. However, in practice, few tasks are accomplished synchronously due to asynchronous initialization. In this article, an asynchronous MTO (AMTO) paradigm is proposed, which aims to deal with multiple optimization problems with asynchronous arrivals. Due to the asynchronous characteristic of tasks, there is multiple tenses knowledge in an AMTO environment. Transferring multitense knowledge may accelerate the optimization process of the target task. Also, an AMTO algorithm is proposed to transfer multitense knowledge. The past-tense knowledge is transferred by an initialization strategy, which selects effective knowledge to deal with mismatched tenses. And the present-tense knowledge is transferred by knowledge reuse, which aligns convergence intervals to handle mismatched evolutionary states. Finally, several AMTO test problem sets and a practical problem are designed to verify the performance of the proposed algorithm. The experimental results show that the performance of the algorithm can be improved by multitense knowledge transfer.

HighlightsWhat are the main findings?We developed and validated new pediatric malnutrition screening models based on non-invasive indicators; the best-performing models (GP, ANN, and ANFIS) showed high apparent diagnostic accuracy in this dataset.Simplified decision tree models showed lower accuracy but offered greater transparency and feasibility for routine ward-level use.What are the implications of the main findings?Machine learning and evolutionary approaches show strong potential to improve pediatric screening of risk of malnutrition.These models show potential as supportive tools in settings where a full subjective malnutrition assessment is not feasible, but further external validation is required before clinical implementation.Background/Objectives: Child malnutrition is a global health challenge linked to poor growth, impaired development, weakened immunity, and adverse outcomes. Early risk detection is essential, but current screening tools differ in accuracy and feasibility. This study aimed to develop and validate new bedside pediatric malnutrition screening models based on machine learning and evolutionary computation methods that can capture complex patterns in non-invasive clinical indicators while remaining practical for routine ward use. Methods: We conducted a cross-sectional study including 180 hospitalized children (1 month–18 years) recruited consecutively from six pediatric wards. The required sample size (minimum 138 participants) was calculated a priori using national prevalence estimates of pediatric undernutrition (4–9.5%) to ensure adequate precision at a 95% confidence level. Data collection included a questionnaire, anthropometry, subjective malnutrition risk assessment, and the Subjective Global Nutritional Assessment (SGNA) tool. Screening models were developed using decision trees, random forests, XGBoost, lasso regression, artificial neural networks, ANFIS, and genetic programming. Their performance was evaluated against the SGNA tool and physician-based subjective malnutrition risk assessment using sensitivity, specificity, AUC, and Cohen’s κ. Results: Machine learning and intelligent evolutionary models (GP, ANN, and ANFIS) showed the best performance in this sample, with substantial to high agreement (κ = 0.81–1.00) and high diagnostic accuracy (AUC = 0.92–1.00) with the subjective malnutrition risk assessment. The GP model demonstrated the highest apparent accuracy in this dataset, but also higher complexity, whereas simpler models such as decision trees showed lower accuracy but greater interpretability and feasibility for routine clinical use. However, validation was performed on a relatively small independent sample, and no external validation was conducted, which may limit the generalizability of the findings. Conclusions: While complex models may serve as digital assessment instruments, simpler models are rapid and more suitable for bedside screening. All developed models are non-invasive and cost-effective and show potential for supportive approaches for early detection of malnutrition risk at hospital admission. However, given the limited validation sample and the absence of external validation, these findings should be interpreted with caution, and further large-scale, multicenter studies are required to confirm generalizability and clinical applicability.

Moving towards a sustainable future, optimising energy systems and resources involves increasingly sophisticated decision-making processes that require balancing technical efficiency with societal, economic and environmental constraints. However, traditional decision-making methods often rely on unidimensional cost metrics and are prone to rank reversal and sensitive to normalisation, leading to unstable or biased outcomes. Driven by these methodological and domain-specific research gaps, this article proposes a novel algorithm, “Multi-Criteria Evaluation via Gradual-Weighting and Aggregation of Normalised Distance Matrices (MEGAN).” Rather than using fixed-weight vectors as in conventional methods, MEGAN adjusts the weight vectors gradually to aggregate rankings across scenarios. Moreover, it utilises dynamic thresholds based on dispersion measures to improve decision precision. The proposed algorithm was verified using a synthetic dataset of ten grid alternatives evaluated across fourteen criteria (economic, societal, environmental, and technical) based on real-world expert priorities. Experimental analyses demonstrated that MEGAN achieves superior ranking stability, remaining insensitive to variations in normalisation techniques compared to conventional decision-making methods. In addition, according to the Institute of Electrical and Electronics Engineers (IEEE) Congress on Evolutionary Computation (CEC) complexity protocol, it achieved a lower computational complexity score than complex outranking methods, thereby offering scalability for large-scale problems. In summary, MEGAN provides a robust, insensitive, and computationally efficient decision-support tool capable of managing the multidimensional trade-offs essential to sustainable energy planning.

Evolutionary multitasking (EMT) has emerged as a popular topic of evolutionary computation over the past decade. It aims to concurrently address multiple optimization tasks within limited computing resources, leveraging inter-task knowledge transfer techniques. Despite the abundance of multitask evolutionary algorithms (MTEAs) proposed for multitask optimization (MTO), there remains a need for a comprehensive software platform to help researchers evaluate MTEA performance on benchmark MTO problems as well as explore real-world applications. To bridge this gap, we introduce the first open-source benchmarking platform, named MToP, for EMT. MToP incorporates over 50 MTEAs, more than 200 MTO problem cases with real-world applications, and over 20 performance metrics. Based on these, we provide benchmarking recommendations tailored for different MTO scenarios. Moreover, to facilitate comparative analyses between MTEAs and traditional evolutionary algorithms, we adapted over 50 popular single-task evolutionary algorithms to address MTO problems. Notably, we release extensive pre-run experimental data on benchmark suites to enhance reproducibility and reduce computational overhead for researchers. MToP features a user-friendly graphical interface, facilitating results analysis, data export, and schematic visualization. More importantly, MToP is designed with extensibility in mind, allowing users to develop new algorithms and tackle emerging problem domains. The source code of MToP is available at: https://github.com/intLyc/MTO-Platform

De novo protein design has emerged as a transformative approach for generating functional proteins without relying on naturally occurring templates, offering vast potential in biomedical and industrial applications. The integration of generative deep learning and evolutionary computation has opened new avenues in synthetic protein engineering. In this study, a hybrid framework, termed GAN-GA (Generative Adversarial Networks + Genetic Algorithm), has been proposed to address the limitations of traditional sequence based generation methods. Specifically, Wasserstein GANs have been employed to produce novel protein sequences that resemble natural proteins. However, their outputs have often lacked functional and physicochemical validation. To improve the biological relevance of these sequences, a multi-objective genetic algorithm, namely the Non-dominated Sorting Genetic Algorithm II (NSGA-II), has been incorporated as a post-generation refinement stage. Through this optimization, sequences satisfying essential physicochemical criteria have been obtained. The proposed framework has been applied to design synthetic Heat Shock Proteins(HSPs), particularly HSP70 and HSP90. Two configurations of the framework have been developed; one focusing on dual-objective optimization (GAN-GA$_{2}$) and another extending the optimization to six key physicochemical attributes (GAN-GA$_{6}$). The resulting sequences have been evaluated and compared with those produced by other state of the art algorithms such as GAN, ProtBert, ESM, Variational Autoencoder (VAE), and Reinforcement Learning (RL) approaches. The optimization in the second stage ensures both structural soundness and biological plausibility of the generated protein sequences. This establishes GAN-GA as a promising strategy for de novo protein engineering in therapeutic and synthetic biology applications.

Abstract Artificial Evolution is a computational approach inspired by biological evolution that is increasingly being applied in the field of computer graphics. It uses evolutionary algorithms such as genetic algorithms, neuro evolution, novelty search, and interactive evolutionary computation to automatically generate visual content. These methods enable systems to explore large and complex design spaces, producing diverse and creative graphical outputs such as images, textures, terrains, animations, and shaders. Recent research from 2021 to 2025 highlights the growing importance of artificial evolution in creative applications where traditional rule-based design is limited. Many studies have also combined evolutionary techniques with deep learning models, including Generative Adversarial Networks (GANs), to enhance realism and control in generated visuals. Additionally, interactive evolutionary systems allow users to guide the generation process based on subjective preferences, improving artistic outcomes. This paper reviews thirty research papers related to artificial evolution in computer graphics, analyzes their objectives and methodologies, and compares key contributions. The study demonstrates that artificial evolution is a powerful tool for automated creativity and innovation in modern graphics systems. Keywords Artificial Evolution, Evolutionary Algorithms, Computer Graphics, Procedural Generation, Generative Art.

Traditional computational protein design heavily relies on expert-level biological inputs to define structural and functional constraints, posing significant barriers in terms of technical implementation and workflow construction. To address this gap, we capitalize on recent advancements in large language models (LLMs)─which excel at complex reasoning in specialized domains by leveraging knowledge bases to generate expert-grade outputs. In this study, we first propose the protein evolutionary paradigm, a design paradigm that emulates the core logic of natural protein evolution by taking biological function as the ultimate target, achieving progressive optimization of protein sequences under explicit functional and structural constraints through iterative evolutionary refinement. Guided by this paradigm, we present MAESD (Multiagent Evolutionary Framework for Protein Sequence Design), a unified computational framework for function- and structure-constrained evolutionary protein design guided by natural language instructions. This paradigm integrates multiagent collaborative reasoning to bridge the semantic gap between natural language descriptions and biological constraints, while adopting an iterative evolutionary optimization mechanism to ensure the biological plausibility of designed sequences at each iteration. MAESD operates through two core collaborative modules for sequence generation: (1) A semantic-to-biological translation module, which employs LLMs and biological databases to interpret user-provided natural language biological requirements and extract actionable protein design constraints; (2) an evolutionary loop module, which realizes iterative sequence refinement via a ″generation-validation″ cycle─utilizing ProGen2 and ProteinMPNN for sequence generation and integrating structural, energetic, and functional verification to filter and optimize sequences. By fusing natural language understanding with evolutionary computation, MAESD reduces the engineering and implementation burden of protein design workflows by automating pipeline integration and parameter adaptation, while expert biological judgment remains necessary for interpreting results and guiding experimental decisions.

Abstract—Cloud computing costs have become a critical concern for organizations deploying workloads on public cloud platforms. Manual cost optimization approaches are inherently limited by the combinatorial complexity of resource allocation decisions across multiple services and time periods. This paper presents a research-grade Cloud Cost Optimization System built for Microsoft Azure that employs two nature-inspired evolutionary algorithms — the Genetic Algorithm (GA) and the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) — to automate and mathematically validate cost reduction strategies while preserving service performance. The system is implemented as a full-stack Python Flask web application that ingests real-world Azure billing export data, performs statistical anomaly detection using Z-score analysis, conducts linear regression trend forecasting, generates rule-based recommendations, and runs both single-objective (GA) and multi-objective (NSGA-II) evolutionary optimization. Experimental results on a 12-month dataset of 489 records across 10 Azure services demonstrate consistent cost reductions of 20–50% while maintaining service performance above user-specified thresholds. NSGA-II further produces a Pareto-optimal front of 20–50 non-dominated solutions per run, enabling decision-makers to select cost-performance trade-offs aligned with their specific business priorities. The system introduces a dynamic optimization engine that allows users to compare algorithm outputs on both the original dataset and custom-defined service costs simultaneously, enabling real-world deployment scenarios beyond the training data. This work demonstrates that evolutionary computation techniques are highly effective for cloud financial management (FinOps) and provides a reproducible, open-architecture framework for researchers and practitioners to further extend and validate cloud cost optimization techniques. Keywords—Cloud Computing, Cloud Cost Optimization, FinOps, Genetic Algorithm (GA), NSGA-II, Multi-Objective Optimization, Evolutionary Algorithms, Resource Allocation, Waste Detection, Z-score Analysis, Linear Regression, Cost Forecasting, Microsoft Azure, Cloud Analytics, Performance Optimization

Heterogeneous three-dimensional (3D) wireless sensor network (WSN) deployment requires balancing sensing coverage, communication connectivity, and deployment cost under coupled K-coverage and C-connectivity constraints. This setting yields a constrained mixed discrete optimization landscape where many conventional multi-objective methods lose diversity or handle feasibility inconsistently. We formulate heterogeneous 3D WSN deployment as a constrained multi-objective problem and propose the Enhanced Multi-Objective Genghis Khan Shark Optimizer (EnMOGKSO). The core novelty is the integration of leader-pursuit dynamics with (i) dual archive-guided selection (elite and neighborhood memories), (ii) bounded external archive diversity control, and (iii) feasibility-first environmental selection for fragmented feasible regions. On the Congress on Evolutionary Computation (CEC) 2020 suite, EnMOGKSO obtains the best Friedman mean ranks in hypervolume (HV) (2.04) and inverted generational distance (IGD) (2.38), with statistically significant differences against most competitors ([Formula: see text], Wilcoxon/Friedman). In heterogeneous 3D WSN deployment, EnMOGKSO yields higher coverage/connectivity values (typically coverage means around 11-12 and connectivity around 7) than weaker baselines (often coverage 5-7 and connectivity 4-5), with higher but stable deployment cost. Overall, the results indicate a stronger convergence-diversity balance and more reliable feasibility-aware search under tight constraints, with practical applicability to 3D monitoring tasks such as industrial facilities, smart buildings, and environmental sensing.

Abstract Researcher evaluation remains a central challenge in scientometrics, where reliable, transparent, and context-sensitive methods are required for decisions related to recruitment, promotion, and awards. Traditional assessment approaches rely heavily on bibliometric indices such as the h -index and its variants; however, these measures often neglect publication venue quality and provide limited discriminatory capability. This study proposes an integrated framework that combines retrospective bibliometric analysis, venue-aware modeling, and Genetic Programming (GP)-based symbolic regression for interpretable researcher evaluation. A balanced dataset of 1,200 computer science authors, equally divided between awardees and non-awardees, was constructed using Google Scholar and Publish or Perish . First, sixty-four bibliometric indices were computed to establish a retrospective ranking baseline, where the $$h_2$$ upper-index achieved the strongest performance by identifying 73% of awardees within top-ranked positions. Second, a venue-aware dataset incorporating journals, conferences, books, and patents was developed, and neural-network–estimated venue contributions indicated journals as the dominant factor (84%). Finally, GP-based symbolic regression was applied to evolve interpretable closed-form equations integrating venue-level features. The best GP-derived model identified 91% of awardees, outperforming both the retrospective baseline and the venue-aware linear model while maintaining interpretability. These findings demonstrate that combining venue-aware modeling with interpretable evolutionary computation provides a more accurate, transparent, and equitable framework for researcher evaluation, with practical implications for academic institutions and research policy design.

K-mer counts are fundamental in many genomic data analysis tasks, providing valuable information for genome assembly, error correction, and variant detection. State-of-the-art k-mer counting tools employ various techniques, such as parallelism, probabilistic data structures, and disk utilization, to efficiently extract k-mer frequencies from large datasets. The distribution of k-mer counts in raw sequencing reads reveals key genomic characteristics such as genome size, heterozygosity, and basecalling quality. The number of reads containing a k-mer has also shown application in genome assembly and sequence analysis. We present ntStat, a toolkit that employs succinct Bloom filter data structures to track both k-mer count and depth information and use in downstream applications. ntStat models the k-mer count histogram using evolutionary computation, and infers valuable insights about the genome, sequencing data, and individual k-mers, de novo. ntStat consistently ran faster than DSK, BFCounter, hackgap, and Squeakr in all of our tests. Jellyfish performed faster than ntStat for human data with k = 25 but fell behind with k = 64. KMC3 was faster overall but at a high disk usage and memory cost. ntStat also used less memory than other non-disk-based k-mer counters and typically, 99.5-99.9% of the k-mers processed by ntStat are counted correctly. ntStat’s histogram analysis module detected heterozygosity percentages and k-mer coverage for long-read datasets simulated from a diploid human genome with less than 1% and 0.5-fold difference to the ground truth. The analysis of simulated long read datasets showed an average error of just 2% in k-mer robustness estimates.

• Multi-objective algorithm based on decomposition for the tagSNP selection problem • Implementation of five new problem-aware operators to explore the search space • Comparative study of the proposed method with six different alternative approaches • Experimentation with five highly relevant datasets from 1000 Genomes Project (1KGP) • Competitive approach in terms of runtime and results quality after the comparisons Nowadays multiple bioinformatics issues can be solved by using evolutionary computation due to its potential to address complex optimization problems. TagSNP selection lies within this class of challenging problems, since genotyping all the Single Nucleotide Polymorphisms (SNPs) for haplotype identification is economically costly and time-consuming. If a reduced number of tagSNPs is chosen instead, the classification of haplotypes will accordingly show a worsening. As a result, tagSNP selection can be considered as a multi-objective optimization problem, in which the aim is to optimize haplotype dissimilarity while minimizing the number of selected tagSNPs. We propose and detail an approach based on the Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) for accurately selecting tagSNPs attending to these two objectives. The proposed method includes novel problem-aware operators for the initialization, crossover, and mutation to boost optimization capabilities. The proposal is experimentally compared with six approaches from the literature on five real datasets, using in the evaluation three quality metrics and their corresponding statistical analyses. The attained results denote that our algorithm provides statistically-significant improvements over previous methods with competitive runtimes, thus highlighting the relevance of the proposed multi-objective approach.

We review major developments in multi-objective optimization over the past decades. Although mathematical foundations and basic concepts have been established earlier, substantial progress in methods for constructing and identifying preferred solutions started in the late 1950s. We classify these approaches into two broad categories: mathematical programming-based and population-based. The former originated in the late 1950s, and its growth accelerated from the 1970s onward. We differentiate between approaches dealing with problems that operate in a continuous solution space and combinatorial problems where some variables are restricted to integer values. Population-based approaches flourished in the 1990s. Our focus is on evolutionary computation techniques that either aim to discover the entire Pareto front or incorporate the decision maker’s preferences to select the most favorable solution(s) or bias the search toward preferred regions. For all categories, we discuss those approaches that, in our opinion, have made major impacts. We examine current research trends and speculate on future directions in the field. • Major developments in the last 50 years in multiobjective optimization. • Mathematical programming-based approaches in multiobjective optimization. • Population-based approaches in multiobjective optimization. • Preference incorporation in multiobjective optimization.

This research presents the development of a universal genetic optimizer (UGO) aimed at solving optimization problems across a wide range of test functions, focusing on achieving reliable global extrema for both theoretical and practical applications. The study utilizes a genetic algorithm (GA) enhanced with neural network-based approaches, implementing key genetic operators such as crossover, mutation, and elitism. The algorithm was tested on benchmark functions including Rosenbrock, De Jong, and Griewank, among others, with statistical analysis identifying optimal parameter settings. The results demonstrate superior performance compared to traditional tools like Excel Solver and GAMS, particularly when elitism is included. The proposed framework highlights the adaptability of evolutionary computation when combined with machine learning techniques. Moreover, it opens perspectives for applying UGO in real-world optimization challenges such as logistics, energy systems, and engineering design.

Multi-objective optimization problems (MOPs) require the simultaneous optimization of conflicting objectives. Real-world MOPs often exhibit complex characteristics, including high-dimensional decision spaces, many objectives, or computationally expensive evaluations. While population-based evolutionary computation has shown promise in addressing diverse MOPs through problem-specific adaptations, existing approaches frequently lack generalizability across distinct problem classes. Inspired by pre-training paradigms in machine learning, we propose a Population Pre-trained Model (PPM) that leverages historical optimization knowledge to solve complex MOPs within a unified framework efficiently. PPM models evolutionary patterns via population modeling, addressing two key challenges: (1) handling diverse decision spaces across problems and (2) capturing the interdependency between objective and decision spaces during evolution. To this end, we develop a population transformer architecture that embeds decision spaces of varying scales into a common latent space, enabling knowledge transfer across diverse problems. Furthermore, our architecture integrates objective-space features through objective fusion to enhance population prediction accuracy for complex MOPs. Our approach achieves robust generalization to downstream optimization tasks with up to 5,000 dimensions-five times the training scale and 200 times greater than prior work. Extensive evaluations on standardized benchmarks and out-of-training real-world applications demonstrate the consistent superiority of our method over state-of-the-art algorithms tailored to specific problem classes, improving the performance and generalization of evolutionary computation in solving MOPs.