The large-scale multi-objective competitive swarm optimization algorithm encounters problems such as a tendency to fall into local optima and the collapse of solution distribution diversity in high-dimensional decision spaces. To address issues, this paper proposes an adaptive fitness landscape competitive swarm optimization algorithm (AFLCSO). It uses a terrain feature extraction mechanism to partition the landscape and assigns multi-strategy synergistic updates to losers, avoiding local optima and finding the optimal solution efficiently. Then, a terrain-aware Gaussian winner guidance (TA-GWG) mechanism is adopted in the elite evolution phase to adaptively inflate the exploration variance with the population stagnation ratio, allowing winners to break free from local traps. Finally, a subregion terrain-aware environmental selection (STA-ES) strategy is introduced to integrate the decision-space terrain repulsion index into the angle penalized distance, enhancing the spatial uniqueness of selected individuals and improving solution distribution uniformity in objective and decision spaces. Tests on standard large-scale benchmark functions show the effectiveness of AFLCSO for complex high-dimensional spaces.

The resource allocation strategy has been well applied in improving the performance of optimization algorithms. However, most resource allocation methods focus on either the objective space or the decision space. Few reports have been published on establishing adaptive resource allocation model that considers the two spaces simultaneously. In this work, a hybrid resource allocation framework is developed to enhance the performance of large-scale multi-objective optimization algorithms and thus fully use the information of the two spaces to build a reasonable resource allocation model. In objective space, shift-based density estimation strategy is used to allocate the individuals to different groups. In decision space, the decision variables are grouped with Pearson correlation coefficient and k-means clustering. This method does not add significant external computational cost to the algorithm, and the new resource allocation model contains more information than single space resource allocation model does. The adaptability of the resource allocation model is enhanced further by designing an adaptive distribution method to determine the reference resources for the population at different evolution phases. The division speed of the population is adaptive modified, and the diversity and evolution time of individuals in the sub-populations are balanced. In addition, a hybrid environmental selection strategy is designed to balance the quality of offspring and computation cost. Three types of experiments on two benchmark functions and a practical experiment are conducted to prove the effectiveness of the new framework. Statistical results indicate that the new framework can enhance the mean inverted generational distance and hypervolume indicator of the algorithms.

The shape of the Pareto Front (PF) has a significant influence on maintaining population diversity, which makes it challenging for most algorithms to effectively balance the population convergence and diversity when solving many-objective optimization problems (MaOPs). Therefore, this paper proposes an indicator-guided many-objective evolutionary algorithm with adaptive mapping distance, termed MaOEA-IAMD. In this algorithm, an adaptive mapping distance calculation strategy is designed to maintain population diversity. This strategy estimates the PF curvature using the Newton–Raphson method, and maps individuals onto the PF based on the obtained curvature information, thereby calculating the adaptive mapping distance between two individuals to measure their similarity in search direction. Meanwhile, we develop a select-replacement strategy by using the binary additive epsilon indicator and the adaptive mapping distance to balance the population convergence and diversity. In addition, a knee points-based classification mutation strategy is designed to generate high-quality offspring, further improving the search efficiency of MaOEA-IAMD. We compare MaOEA-IAMD with six state-of-the-art algorithms on 21 benchmark test problems and 2 real-world optimization problems. The experimental results demonstrate the competitiveness and effectiveness of the proposed algorithm.

Precise segmentation of dermatological lesions is vital for clinical diagnostics and therapeutic planning, yet it faces obstacles such as indistinct boundaries, hair interference, measurement tool artifacts, and image noise, which degrade performance. Annotating medical data is costly and time-intensive, and traditional fully supervised methods require extensive labeled data. Semi-supervised learning effectively harnesses limited annotated samples alongside abundant unannotated data to enhance model training. This paper proposes a semi-supervised Mean Teacher neural network model to improve performance using minimal labeled data. Built on the U-Net architecture, it integrates the Mean Teacher framework with consistency regularization and employs Exponential Moving Average (EMA) to enhance learning from unlabeled data. Validated on ISIC2018 and ISIC2016 datasets, the model achieves a Dice coefficient of 89.34

Competitive swarm optimization has shown impressive performance in addressing large-scale multi-objective optimization problems. However, its competitive mechanism can lead to contention among local optimal solutions within the current population. As the decision space grows more complex, the diversity of the population in competitive swarm optimization often decreases, resulting in “premature convergence”. To address this issue, we propose an adaptive multi-region multi-directional competitive swarm optimization algorithm, named AMRCSO. First, this algorithm introduces an adaptive reference-guided subregion division strategy that segments the objective space based on the population’s distribution, enhancing search diversity and coverage. Additionally, a multi-directional exploration strategy is employed to explore each subregion of the decision space, enabling parallel and multi-directional exploration that accelerates convergence to the global optimal solution. Furthermore, an effective population initialization and tri-stage updating strategy are implemented to improve population diversity and enhance convergence accuracy, respectively. Experimental results on large-scale multi-objective benchmark functions, ranging from 500 to 5000 decision variables, demonstrate that the AMRCSO algorithm achieves superior convergence accuracy and distribution compared to existing algorithms for large-scale multi-objective optimization problems.

In recent years, crowdsourcing has garnered attention as a work model for a large number of workers. The types of crowdsourcing tasks have also become increasingly diverse, encompassing ride-hailing services offered by platforms such as Didi, online-to-offline (O2O) applications for tutoring services, and online crowdsourcing platforms such as Amazon and Upwork. However, as workers’ aspirations for personal interests and future career development grow and clients seek greater convenience in task posting, traditional crowdsourcing platforms face challenges in satisfying the diverse needs of both workers and clients in task selection and posting. To address this, our study introduces and investigates a new problem, namely the multi-type crowdsourcing (MTC) task recommendation problem, which we model as a multi-objective optimization problem aimed at maximizing both worker and client satisfaction. Given the varying requirements of different task types and the challenges of achieving global optimization with traditional recommendation methods for large-scale data, we propose the Low-Constraint Two-Stage Alternating Multi-Objective Evolutionary Algorithm (LTA-MOEA). The optimization process is divided into two stages. The first stage focuses on finding extreme solutions of the objective function under low-constraint conditions, while the second stage balances these extreme solutions by considering all constraints, thereby approximating the Pareto front. These two stages alternate, with an auxiliary population introduced during the process to facilitate the generation of both extreme and balanced solutions. We validate the efficacy of the proposed algorithm using both real and synthetic datasets. Although no unified multi-type crowdsourcing platform currently exists in the market, we believe that such platforms will inevitably emerge in the future as the demand for personalized, flexible, and intelligent task matching continues to grow. Our study thus provides a forward-looking framework and technical foundation for future research and development in this area.

Vision-language models (VLMs) have demonstrated strong capabilities in zero-shot classification. However, their performance remains highly dependent on the quality of text prompts. Existing approaches still suffer from heavy reliance on manually designed prompts, limited automation, or a lack of natural language interpretability. To address these limitations, this paper proposes Meta-Prompt Optimization based on Differential Evolution (DE-MP), which formulates meta-prompt design as optimization problem. Unlike existing methods that directly optimize text prompt, DE-MP treats meta-prompts as optimization variables and adopts zero-shot Top-1 classification accuracy as the fitness function. This design constructs a closed-loop optimization framework, incorporating downstream task performance feedback. This transforms the prompt optimization from being driven by manual experience to being performance-driven, significantly reducing the cost of manual design. To enable differential evolution in natural language optimization, an LLM-driven semantic differential mutation mechanism is introduced. The proposed mechanism extends differential operations from numerical space to semantic space, enabling semantic mutations to occur under clear semantic guidance, while allowing for controllable semantic exploration. This, in turn, enhances the diversity of the search and the semantic expressiveness. Furthermore, a hybrid mutation strategy is designed for different meta-prompt components by combining semantic differential mutation, index-based differential mutation, and random exploration, thereby balancing search efficiency and population diversity. Extensive experiments on multiple standard zero-shot visual classification benchmarks demonstrate that DE-MP consistently improves classification accuracy without updating the parameters of either the VLM or the LLM. Moreover, it achieves superior performance compared with several state-of-the-art prompt engineering and prompt optimization methods.

Existing attention-based methods for large-scale multi-objective optimization (LMOAM) focus only on decision variables, using their variance to guide search behavior. However, single-space strategies ignore critical information in the objective space and the diversity and search efficiency are often degraded for solving multimodal multi-objective optimization problems (MOPs). To address this problem, a novel large-scale optimization framework that integrates a dual-space attention mechanism is proposed in this paper. Different from building attention only with information in decision space, a dual-space Key matrix that quantifies variable importance by combining decision-variable and objective-space distributions is first designed in the framework to refine the precision of the attention. Subsequently, a cross-space clustering method is adopted to select the representative solutions by analyzing the characteristics of individuals in both spaces to construct the Query matrix. The accuracy of attention allocation is improved. Finally, A linear inverse mapping strategy is used to enhance the diversity of the population by translating promising objective-space solutions back to the decision space. Unlike existing approaches, the characteristics of decision and objective space are linked with a new attention mechanism, and the exploration and exploitation of the population are well balanced. Three types of experiments are designed on two benchmark test sets with 500-dimensional and 1000-dimensional decision variables and the voltage transformer optimization problem to demonstrate the efficacy of the AIDF framework, experimental results indicate that AIDF surpasses comparative algorithms in terms of the average performance of IGD and HV.

In recent years, neural heuristics leveraging deep reinforcement learning have exhibited considerable promise in addressing multi-objective combinatorial optimization problems (MOCOPs). Nonetheless, challenges persist in attaining both high learning efficiency and optimal solution quality. To address this issue, we propose a novel multi-objective optimization algorithm grounded in information geometry and machine learning principles, which integrates adaptive gradient descent with meta-reinforcement learning techniques to effectively tackle MOCOPs. In this paper, we present a meta-learning framework aimed at enhancing model performance in multi-objective combinatorial optimization through tensor remodeling, preconditioned gradient descent, and entropy regularization strategies. Experimental results demonstrate that the proposed method yields significant performance improvements across several classic multi-objective combinatorial optimization challenges, including the Multi-objective Traveling Salesman Problem (MOTSP), Multi-objective Vehicle Routing Problem (MOCVRP), and Multi-objective Knapsack Problem (MOKP).

Convolutional neural networks (CNNs) are used in diverse domains, such as computer vision and natural language processing. However, the increasing number of parameters and memory requirements pose challenges for their deployment on edge devices. Studies have demonstrated that CNNs have redundant parameters, the criteria for importance-based pruning methods are cumbersome and existing automatic pruning algorithms struggle to effectively capture global features. To address the drawbacks of such methods, an automatic channel pruning method using improved the poplar optimisation algorithm (ACPPO) is proposed herein. This study aimed to search for the optimal network structure using clustering and optimisation algorithms. First, hierarchical density–based clustering is employed to group the channels of each convolutional layer based on similarity, thereby forming a preliminary compression structure. Subsequently, the improved poplar optimisation algorithm (POA) is used to determine the optimal compact network structure with an iterative optimisation search. During the asexual propagation of the POA, the local optimum is used instead of the original random historical information to guide the search direction of individuals, thereby enhancing the diversity of the population and improving search efficiency and global search capability. Herein, the effectiveness of ACPPO is verified on two commonly used image classification datasets: CIFAR-10 and ILSVRC-2012. On CIFAR-10, ACPPO reduces the GoogLeNet network by 69.37% of the parameters and 75.33% of the FLOPs, with an accuracy improvement of 0.13% over the benchmark.

The Particle Swarm Optimization (PSO) algorithm is widely applied in multitask optimization because of its simplicity and rapid convergence. Nevertheless, the original Multitask PSO (MTPSO) algorithm rarely utilizes local similarity for dissimilar or less similar tasks and lacks mechanisms for information exchange (IE) among variables of different dimensions. This study presents a novel MTPSO based on variable chunking and local meta-knowledge transfer (MKT) to leverage the local information of individuals and enable IE among variables of varying dimensions. First, a construction-assisted transfer individual strategy is proposed. Using variable chunking and Latin hypercube sampling, an auxiliary transfer individual is constructed for each task. Using this individual to guide population evolution can promote IE among individuals with different dimensions and effectively enhance individual diversity. Subsequently, the populations are clustered to assess the local similarities between tasks. Based on these similarities, the MKT strategy is designed to promote mutual learning opportunities among locally similar populations. On the adaptive side, an adaptive matching probability strategy is proposed to help the algorithm dynamically adjust the transfer probability according to the task similarities, effectively reducing the occurrence of negative transfers. Finally, the algorithm is evaluated on the CEC 2017 problem set and two real-world multitask optimization problems, and its performance is compared with 12 other typical multitask optimization algorithms. The results show that the proposed algorithm outperforms most of the compared algorithms both in terms of convergence speed and accuracy. Meanwhile, variant experiments demonstrate the effectiveness of the proposed strategies.

Most many-objective evolutionary algorithms balance convergence and diversity by improving environmental selection strategies, with less attention paid to the contribution of offspring to the algorithm performance. This study proposes a many-objective evolutionary algorithm based on a decision variable classification mutation and indicator (MaOEA-DI) that improves the quality of offspring while balancing convergence and diversity to enhance the algorithm’s performance. In MaOEA-DI, a dual-archive guided decision variable classification mutation strategy is designed. This strategy utilizes the decision variable information of elite individuals from the convergence and diversity elite archives, which are updated every generation, to guide the mutation process in generating high-quality offspring. To enhance the quality of the candidate solution set, the newly generated offspring are filtered using the current population information. In addition, an indicator- and density-based environmental selection strategy is developed to balance convergence and diversity. Experimental results on 27 benchmark problems, two real-world optimization problems, and a multiline distance minimization problem show that MaOEA-DI outperforms six advanced algorithms.

ABSTRACTDepression, a prevalent mental disorder in modern society, significantly impacts people's daily lives. Recently, there have been advancements in developing automated diagnosis models for detecting depression. However, data scarcity, primarily due to privacy concerns, has posed a challenge. Traditional speech features have limitations in representing knowledge for depression diagnosis, and the complexity of deep learning algorithms necessitates substantial data support. Furthermore, existing multimodal methods based on neural networks overlook the heterogeneity gap between different modalities, potentially resulting in redundant information. To address these issues, we propose a multimodal depression detection model based on the Enhanced Cross‐Attention (ECA) Mechanism. This model effectively explores text‐speech interactions while considering modality heterogeneity. Data scarcity has been mitigated by fine‐tuning pre‐trained models. Additionally, we design a modal fusion module based on ECA, which emphasizes similarity responses and updates the weight of each modal feature based on the similarity information between modal features. Furthermore, for speech feature extraction, we have reduced the computational complexity of the model by integrating a multi‐window self‐attention mechanism with the Fourier transform. The proposed model is evaluated on the public dataset, DAIC‐WOZ, achieving an accuracy of 80.0% and an average F1 value improvement of 4.3% compared with relevant methods.

Software crowdsourcing has become a cornerstone of the Internet economy because of its unique capacity for selecting optimal workers to complete specific tasks. However, new workers face limited task opportunities compared to experienced workers, which negatively impacts their motivation and decreases overall activity on crowdsourcing platforms. This reduced activity can harm platform reputation. To encourage the active participation of new workers, this study introduces a novel method to identify and match worker–task preferences. Our approach categorizes preferred tasks based on golden tasks, historical data, and worker interests. We then present the Multi-Objective Worker–Task Recommendation (MOWTR) algorithm, built upon the Non-Dominated Sorting Genetic Algorithm II (NSGA-II). The MOWTR algorithm allocates tasks by considering worker preferences, salaries, and capabilities, aiming to optimize collective team performance while minimizing team costs, especially for new workers. New crossover and two-stage mutation operators are incorporated to increase algorithm efficiency. Experimental evaluations on four real and synthetic datasets demonstrate that MOWTR outperforms four advanced baseline methods, confirming its effectiveness.

Multitask optimization algorithms can simultaneously derive the best solution for different tasks; however, the convergence speed of such algorithms is slow when frequent negative transfers occur. This is primarily because the similarity function of different tasks is designed only in the decision or target space. Moreover, an algorithm is prone to fall into local optima when population diversity is lost. To reduce negative migration and balance diversity and enhance convergence of multitask optimization algorithms, a multitask particle swarm optimization algorithm based on dual spatial similarity (MTPSO-DSS) is developed in this study. A new similarity function is built into the algorithm for the different tasks based on both decision and target spaces, whereby the transfer probability is adaptively adjusted. The new similarity function, which is more rigorous and accurate, can reduce the probability of negative migration and maintain the convergence speed. Furthermore, a new updating method is designed to handle negative migration and increase diversity of search directions. Adaptive mutation and non-allelic gene crossover strategies are designed to increase the diversity of the algorithm and help it escape from local optima. To verify the performance of the proposed algorithm, nine general multitasking optimization test functions are tested via the proposed algorithm, and the results are compared with other eight multitasking algorithms. The proposed algorithm outperformed the other algorithms for most functions in terms of convergence accuracy and speed, and the average improvement in the convergence accuracy compared with the other eight algorithms is between 23.35 and 99.99%.

Dynamic multi-objective optimization problems (DMOPs) present a significant challenge due to the need to rapidly track optimal solutions after environmental changes. Most existing algorithms overlook the dynamic nature of environmental changes and rely on a single response strategy, which can often lead to ineffective responses. Therefore, in this paper, an ensemble method with a self-adaptive response mechanism, called EM-SARM, was developed, which incorporates multiple environmental response strategies designed to account for dynamic characteristics. In the proposed EM-SARM, a temporary auxiliary population (TAP) was innovatively constructed, and a novel type of Pareto-optimal set (PS) change detection strategy was designed to identify the type of PS change, which ensures the appropriate response is applied, avoiding ineffective responses and enabling the generation of high-quality initial populations when dealing with dissimilar environmental changes. Specifically, upon detecting dissimilar environmental changes, the first step involves identifying the PS change type, followed by activating the dissimilar environment response strategy. Instead, the historical environmental change data is utilized to guide the population evolution. EM-SARM was evaluated on 19 test instances and compared with three dynamic multi-objective evolutionary algorithms. The experimental results show that the proposed EM-SARM is effective in solving various DMOPs.

Weighted optimization framework (WOF) achieves variable dimensionality reduction by grouping variables and optimizing weights, playing an important role in large-scale multi-objective optimization problems. However, because of possible problems such as duplicate weight vectors in the selection process and loss of population diversity, the algorithm is susceptible to local optimization. Therefore, this paper develops an algorithm framework called multi-population multi-stage adaptive weighted optimization (MPSOF) to improve the performance of WOF in two aspects. First, the method of using multi-population is employed to address the issue of insufficient algorithmic diversity, while simultaneously reducing the likelihood of converging towards local optima. Secondly, a processing stage is incorporated into MPSOF, where a certain number of individuals are adaptively selected for updating based on the weight information and evolutionary status of different subpopulations, targeting different types of weights. This approach alleviates the impact of repetitive weights on the diversity of newly generated individuals, avoids the drawback of easily converging to local optima when using a single type of weight for updating, and effectively balances the diversity and convergence of subpopulations. Experiments of three types designed on several typical function sets demonstrate that MPSOF exceeds the comparison algorithms in the three metrics for Inverse Generation Distance, Hypervolume and Spacing.

It is still a huge challenge for traditional Pareto-dominated many-objective optimization algorithms to solve manyobjective optimization problems because these algorithms hardly maintain the balance between convergence and diversity and can only find a group of solutions focused on a small area on the Pareto front, resulting in poor performance of those algorithms.For this reason, we propose a reference vector-assisted algorithm with an adaptive niche dominance relation, for short MaOEA-AR.The new dominance relation forms a niche based on the angle between candidate solutions.By comparing these solutions, the solution with the best convergence is found to be the non-dominated solution to improve the selection pressure.In reproduction, a mutation strategy of k-bit crossover and hybrid mutation is used to generate high-quality offspring.On 23 test problems with up to 15-objective, we compared the proposed algorithm with five state-of-the-art algorithms.The experimental results verified that the proposed algorithm is competitive.

Fine-grained visual classification (FGVC) aims to identify subcategories of objects within the same superclass. This task is challenging owing to high intra-class variance and low inter-class variance. The most recent methods focus on locating discriminative areas and then training the classification network to further capture the subtle differences among them. On the one hand, the detection network often obtains an entire part of the object, and positioning errors occur. On the other hand, these methods ignore the correlations between the extracted regions. We propose a novel highly scalable approach, called TransFGVC, that cleverly combines Swin Transformers with long short-term memory (LSTM) networks to address the above problems. The Swin Transformer is used to obtain remarkable visual tokens through self-attention layer stacking, and LSTM is used to model them globally, which not only accurately locates the discriminative region but also further introduces global information that is important for FGVC. The proposed method achieves competitive performance with accuracy rates of 92.7

Competitive swarm optimizer (CSO) based on multidirectional search plays a crucial role in addressing large-scale multiobjective optimization problems (LSMOPs). However, relying solely on uniform or cluster partitioning of the objective space for sampling, along with two search directions constructed with upper and lower boundaries of global variables, sometimes lacks consideration of regional information. This results in an inefficient search and hinders the global convergence of the algorithm. To solve these problems, this study proposes a large-scale multiobjective competitive swarm optimizer algorithm based on regional multidirectional search (AMSLMOEA). Firstly, an adaptive objective space partitioning method based on the evolutionary state of the population is designed to enhance the adaptability of partitioning. Secondly, an individual multidirectional search strategy is introduced. Considering the algorithm’s computational complexity, the strategy selects the optimal individual within each subregion and constructs four-directional search vectors based on the lower limit of the global decision variables and the upper limit of the individual decision variables within the subregion. To validate the effectiveness of AMSLMOEA, the performance is tested on four benchmark function sets. The results demonstrate that AMSLMOEA outperforms the vast majority of the compared algorithms in terms of the IGD and HV metrics.