Federated Learning (FL) is a critical enabler for intelligent Internet of Things (IoT) systems, allowing collaborative model training across heterogeneous devices while preserving data privacy. However, FL performance degrades significantly under non-Independent and Identically Distributed (non-IID) data due to weight divergence, and existing mitigation methods often introduce substantial overhead unsuitable for resource-constrained IoT deployments. We identify that weight divergence in non-IID FL stems primarily from insufficient cross-device knowledge exchange before aggregation—a contributing factor that has been underexplored in existing literature—and propose FedCAD, a communication-efficient FL framework that directly addresses this root cause. The core innovation is delayed cross-group aggregation, which strategically postpones global model updates until each model copy has been sequentially trained across all device groups, ensuring comprehensive knowledge integration. This mechanism is supported by dual-feature device clustering that creates meaningful group structures and fairness-aware adaptive selection that ensures representative participation—forming a causally linked pipeline where clustering provides structure, selection provides quality, and delayed aggregation provides the mechanism for thorough knowledge exchange. Extensive experiments on five benchmark datasets across 13 heterogeneous scenarios demonstrate that FedCAD consistently outperforms ten state-of-the-art methods with negligible additional communication overhead.

The deployment of billions of Internet of Things (IoT) devices is driving unprecedented data generation at the network edge, demanding high computational power for real-time deep learning (DL) while raising serious concerns about energy consumption. While edge computing offers a viable paradigm for decentralized DL by preserving data privacy and reducing latency, the substantial energy costs of DL training and inference pose a major challenge for resource-constrained edge devices. This work provides a comprehensive review of state-of-the-art studies that address energy efficiency at the intersection of DL and edge computing. Moving beyond isolated solutions, we analyze the critical need for a co-design approach integrating hardware and software with adaptive resource management to build sustainable systems. The paper systematically examines hardware-level optimizations and software-level techniques for reducing energy consumption while maintaining model accuracy. Furthermore, it investigates how adaptive management of compute, memory, and communication resources is key to dynamic energy savings. Finally, the paper synthesizes recent trends, identifies emerging opportunities, and discusses open challenges, positioning hardware-software co-design as the most promising approach for achieving scalable and energy-efficient deep learning in edge computing.

Traditional continuous deep reinforcement learning (RL) algorithms employ deterministic or unimodal Gaussian actors, which cannot express complex multimodal decision distributions. This limitation can hinder their performance in diversity-critical scenarios. There have been some attempts to design online multimodal RL algorithms based on diffusion or amortized actors. However, these actors are intractable, making existing methods struggle with balancing performance, decision diversity, and efficiency simultaneously. To overcome this challenge, we first reformulate existing intractable multimodal actors within a unified framework, and prove that they can be directly optimized by policy gradient via reparameterization. Then, we propose a distance-based diversity regularization that does not explicitly require decision probabilities. We identify two diversity-critical domains, namely multi-goal achieving and generative RL, to demonstrate the advantages of multimodal policies and our method, particularly in terms of few-shot robustness. In conventional MuJoCo benchmarks, our algorithm also shows competitive performance. Moreover, our experiments highlight that the amortized actor is a promising policy model class with strong multimodal expressivity and high performance. Our code is available at https://github.com/PneuC/DrAC

With the rapid advancement and integration of Internet of Things technology into manufacturing, industrial workshops in computer, communication, and consumer electronics (3C) manufacturing are increasingly confronted with complex computational tasks during production. However, the limited hardware resources and computational capabilities of local devices often hinder efficient task execution. Computational offloading offers a viable solution by allowing complex computational tasks to be processed on either edge or cloud servers, enhancing the efficiency of computational task handling in production environments. A critical challenge lies in optimizing task offloading among local devices, edge servers, and cloud servers to maximize production efficiency while ensuring reasonable task scheduling. To address this challenge, this work proposes a flexible computational offloading strategy based on an edge-cloud architecture in a smartphone manufacturing workshop. First, a framework for edge-cloud workshop manufacturing is constructed, integrating various smartphone production devices. Based on the edge-cloud framework, a constrained optimization problem for computation offloading is formulated, using latency as the objective in industrial production settings. A time consumption model is employed to optimize computational time, and a novel scheduling strategy named Ivy-Genetic Evolution Algorithm (IGEA) is designed to solve the scheduling problem. The IGEA integrates genetic operators into the Ivy Algorithm to introduce a randomness strategy. Experimental results demonstrate that IGEA significantly outperforms state-of-the-art approaches in optimizing production efficiency.

Medical image segmentation challenges stem from complex lesion morphologies and real-time clinical needs. Here we present MAR-Net, a novel framework that integrates adaptive attention mechanisms, hierarchical contextual modeling, and efficient training strategies to address these challenges. The architecture employs a CBAM-based dual-attention module to dynamically enhance discriminative features while suppressing redundant information, improving lesion boundary localization. Cascaded dilated convolutions expand the receptive field for global context capture, complemented by a multi-scale decoder that integrates deep semantic and shallow spatial features. A multi-scale training strategy with hierarchical loss supervision optimizes model adaptability without compromising inference efficiency. Experimental validation on ISIC and CholecSeg8K datasets demonstrates MAR-Net’s superiority: it outperforms mainstream methods across segmentation accuracy, recall rate, and other metrics, achieving notable improvements for complex lesions of varying sizes. Notably, MAR-Net maintains high performance on both dermoscopic and laparoscopic images, showcasing its broad applicability. These results confirm MAR-Net as a robust medical segmentation solution, balancing precision and efficiency for clinical use.

The Dynamic Job Shop Scheduling Problem (DJSP), a critical challenge in 3C manufacturing, requires efficient resource allocation under dynamically changing production conditions where jobs arrive unpredictably. Traditional optimization methods struggle to provide scalable solutions due to the high computational cost of searching for optimal schedules in large and complex environments. To solve this problem, this work proposes the Dual-Graph convolutional networks assisted Dynamic Cooperative Hunting Optimizer (DG-DCHO), which integrates graph-convolutional networks (GCN) with metaheuristic optimization to generate high-quality schedules while significantly improving computational efficiency. GCN generator processes graph representations of the job-shop environment and captures complex dependencies among jobs and machines to construct high-quality initial schedules that serve as initial solutions for the optimization process. GCN evaluator estimates makespan values directly from schedule representations and replaces costly fitness evaluation that minimizes computational overhead and improves optimization speed. Dynamic Cooperative Hunting Optimizer (DCHO) serves as the base optimizer and generates scheduling solutions by balancing global exploration with local exploitation through an adaptive search strategy. Experimental results across various DJSP instances demonstrate that DG-DCHO consistently outperforms state-of-the-art scheduling algorithms by producing superior solutions while requiring fewer computational resources, making itself a scalable and effective framework for real-time dynamic scheduling in large-scale 3C manufacturing systems.

Neurological dysfunction encompasses a variety of diseases resulting from neural damage. Accurate assessment of neurological function is critical for diagnosis and the development of effective treatment plans. A significant number of patients with neurological disorders exhibit ocular abnormalities. Analyzing ocular status through eye movement capture plays a pivotal role in understanding various neurological dysfunctions. However, current methods of analyzing ocular status for neurological function assessments lack precision and objectivity, often relying heavily on physicians’ subjective judgment. This work proposes the Ocular-enhanced Face Keypoints Network (OFKNet), a facial keypoint detection model based on deep convolutional neural networks. OFKNet employs ConvNeXt as its backbone network and introduces a multi-scale input enhancement strategy. Additionally, a region enhancement module based on MobileNetV3 is designed to optimize features in the canthus area. Multiscale feature fusion and channel weighting are achieved through an improved Path Aggregation Network and Squeeze-and-Excitation modules. To validate OFKNet’s accuracy, we compared it with state-of-the-art models, including MediaPipe FaceLandmarker, InsightFace, Dlib68, and Dlib81, using a patient dataset we collected. Experimental results demonstrate that OFKNet outperforms existing models, particularly in calibration accuracy around the eyes. By monitoring eye movements in real-time, OFKNet ensures high-precision extraction of key points in each frame, accurately reflecting changes in patients’ ocular movements.

Water quality anomaly detection refers to the identification of abnormal changes in water parameters, which is crucial for ensuring environmental safety and preventing contamination events. With the growing volume of water environment sensing data and increasing demand for intelligent, transparent water quality management systems, achieving accurate, rapid, and interpretable anomaly detection has become a critical challenge in early warning systems. To tackle this challenge, this work proposes an anomaly detection model named Shapelet Temporal Evolution Graph Network (STEG), which constructs time-aware Shapelets and adopts graph attention networks to build Shapelets evolution graphs, learning multidimensional dynamic relationships within and between time segments. By incorporating both local and global temporal evolution factors, the approach ensures the interpretability of both the detection process and its resulting outputs. Experiments on two real-world datasets show that STEG outperforms state-of-the-art methods in terms of anomaly detection accuracy and generalization. Moreover, it provides clear and transparent reasoning for water quality anomaly detection.

Accurate estimated time of arrival (ETA) prediction for long vehicular trips remains challenging in intelligent transportation systems (ITS) due to heterogeneous traffic patterns and limited local data availability. While federated learning (FL) addresses privacy concerns by decentralizing data training, traditional FL frameworks often struggle with high computational costs and poor adaptability to multi-task scenarios. To overcome these limitations, this paper proposes a Lightweight Multi-task Federated Learning (LMFL) framework for efficient and privacy-preserving ETA prediction. LMFL integrates a novel SE-CIFG, combining a Squeeze-Excitation (SE) attention module to prioritize critical spatio-temporal features and a Coupled Input and Forget Gate (CIFG) to simplify long-term traffic dependency modeling. Additionally, LMFL employs a Federated Gradient Compression Algorithm (FedGCA) to reduce communication overhead between edge and cloud using adaptive thresholding and sparse tensor encoding. Real-world traffic simulation dataset demonstrates that LMFL achieves significantly higher predictive accuracy compared to existing methods, achieving an average 17.1% improvement in prediction precision while reducing training time by 4.1%.

Water quality prediction methods forecast the short- or long-term trends of its changes, providing proactive advice for preventing and controlling water pollution. Existing water quality prediction methods typically fail to capture water quality’s nonlinear characteristics accurately and only consider historical time series data. However, meteorology and other factors also significantly impact water quality indicators. Therefore, considering only historical data of water quality time series is not feasible. To solve this problem, this work proposes a hybrid water quality prediction model called CMLIP, which integrates convNeXt V2, multimodal bottleneck transformer, low-rank multimodal fusion, iTransformer, and PatchTST. CMLIP inputs water quality time series and meteorological remotely sensed rainfall images into a multimodal fusion module before prediction. Specifically, CMLIP integrates the model of ConvNeXt V2 to extract image features. Its multimodal fusion module combines a multimodal bottleneck transformer and the low-rank multimodal fusion to fuse the time series and images. Furthermore, CMLIP combines iTransformer and PatchTST to form an improved prediction module that realizes the prediction of fused features. Experimental results with real-life water quality time series and remotely sensed rainfall images demonstrate that CMLIP when fusing meteorological data, achieves an average improvement of 17% in water quality forecasting accuracy compared to forecasts using only water quality time series. Moreover, CMLIP outperforms other state-of-the-art algorithms in both data fusion and prediction, with an average enhancement of 6% in fusion effectiveness and an average improvement of 22% in prediction accuracy.

Water quality forecasting is a time series analysis task involving estimating future water conditions, vital in environmental management and pollution control. However, existing time series analysis methods focus only on historical observational data, neglecting information from other modalities, leading to incomplete feature extraction and affecting forecasting accuracy and robustness. In addition, the complex spatial dependencies between water quality monitoring stations and the nonlinear fluctuations in water quality indicators caused by meteorological factors present additional challenges. This work proposes a Spatio-Temporal Multimodal Fusion architecture for long-term water quality forecasting, named STMF, to address these issues. It first captures spatio-temporal dependencies by integrating temporal features with upstream-downstream relationships among monitoring stations. Then, STMF further designs a Low-rank Cross-modal Interaction Fusion (LRCIF) method, which fuses spatio-temporal features with precipitation features from the remote sensing image, as an additional modality, effectively leveraging complementary information from multiple data sources to enhance the accuracy and stability of water quality forecasting. Experimental results on real-world water quality datasets demonstrate that the proposed STMF significantly outperforms existing state-of-the-art methods in prediction accuracy. In particular, for long-term forecasting tasks with a 192-step horizon, STMF improves MSE and MAE by 14% and 12%, respectively, compared to unimodal models. It further validates the effectiveness of the multimodal fusion strategy. Overall, STMF offers an effective solution for water quality monitoring and management.

The advent of smart manufacturing in Industry 4.0 signifies the era of connections. As a communication protocol, Object linking and embedding for Process Control Unified Architecture (OPC UA) can address most semantic heterogeneity issues. However, its semantics are not formally defined at the application layer. To address the information silo problem caused by semantic heterogeneity, an integration framework named Querying of Ontology Mapping-based OPC UA (QOMOU) is proposed. QOMOU extracts information models of OPC UA servers into resource description framework triples and utilizes web ontology language for semantic enrichment and inference. Then, an Event Class Semantic Similarity Calculation (ECSSC) method is proposed for device type identification, enabling the classification of semantically heterogeneous OPC UA devices. The effectiveness of ECSSC is validated through queries with the RDF query language (SPARQL) protocol in Apache Jena. Experimental results demonstrate that ECSSC improves the accuracy of device identification by approximately 7% compared to benchmark device identification models. Specifically, compared with graph embedding-based methods, QOMOU’s query performance is approximately 13% higher, and its query efficiency is 5% higher on average compared to both structured query and extensible markup languages. Moreover, by employing a keyword-matching algorithm, the query accuracy of the existing heterogeneous data integration scheme is improved by 4% on average. This enhancement can boost the operational efficiency of Internet of Things systems based on the OPC UA architecture.

ABSTRACT Background In recent years, Large Models (LMs) have been rapidly developed, including large language models, visual foundation models, and multimodal LMs. They are updated and iterated at a very fast pace. These LMs can accomplish many tasks, e.g., daily work assistant, intelligent customer service, and intelligent factory scheduling. Their development has contributed to various industries in human society. Aims The architectural flaws of LMs lead to several problems, including illusions and difficulty in locating errors, limiting their performance. Solving these problems properly can facilitate their further development. Methods This work first introduces the development of LMs and identifies their current problems, including data and energy consumption, catastrophic forgetting, reasoning ability, localization fault, and ethical problems. Then, potential solutions to these problems are provided, including increase data and computation capability, neural‐symbolic synergy, and data orientation to human pattern. Discussion This work discusses developing vertical domain LMs on top of some base LMs. In addition, this work introduces three typical real‐world applications of LMs, including autonomous driving, smart industrial productions, and intelligent medical assistance. Conclusion By embracing the advantages of LMs and solving their fundamental problems, many industries are expected to achieve promising prospects in the future.

Over the past decades of rapid development, the global water pollution problem became prominent. Accurate water quality prediction can detect the trend and anomaly of water quality changes in advance, thereby taking timely measures to avoid water quality problems. Traditional statistical methods for water quality prediction tend to fail to capture the complex relationship among multiple water quality variables. Deep learning models face a challenge to capture both temporal dependence and spatial correlation of the water quality series data. To solve the above problems, this work proposes an adaptive and dynamic graph fusion water quality prediction model based on a spatiotemporal attention mechanism namedSpatial-TemporalGraphFusionTransformer (STGFT). It integrates a spatial attention encoder, a temporal attention encoder, an adaptive dynamic adjacency matrix generator, and a multi-graph fusion layer. Among them, the first two are adopted to capture the spatial correlations and temporal characteristics among different water quality monitoring stations, respectively. The generator can produce adaptive and dynamic adjacency matrices to reflect potential spatial relationships in a river network. Experimental results with real-life water quality datasets reveal that the prediction accuracy of STGFT outperforms the existing state-of-the-art models. Note to Practitioners—This paper is motivated by the problem of long-term water quality prediction. The highly volatile water quality data and the nonlinear characteristics of the time series greatly affect the accuracy of the forecasting task. Existing approaches fail to simultaneously capture spatial correlations and temporal characteristics among different water quality monitoring stations, affecting the accuracy of water quality predictions. This work proposes a water quality prediction method that captures the spatial correlations and temporal characteristics among different water quality monitoring stations. Moreover, it produces adaptive and dynamic adjacency matrices to reflect potential spatial relationships in a river network. Experimental results from three real-world datasets show that this approach is feasible and obtains more accurate prediction results. Furthermore, this method can also be applied to other areas of time series prediction, including finance, traffic, and smart manufacturing.

Computing ecosystems are shifting toward a computing continuum paradigm designed to handle the diverse and dynamic nature of computing resources spread across various locations. It demonstrates significant potential in providing high-bandwidth and low-latency services for users. However, as a large number of users request services from distributed computing continuum systems, it is critical to schedule numerous delay-sensitive, fractional workloads and maximum parallelism-bound jobs to appropriate backend resources,e.g., cloud container instances. In addition, the scheduling strategy also needs to maximize the social welfare that incorporates the utilities of jobs and the revenue of service providers. However, current workload scheduling algorithms are based on simple heuristics and lack performance guarantees. Due to the unpredictability of online requests, the distribution of requests should not be assumed. Therefore, designing an online workload scheduling strategy without assumptions on request distributions is essential for balancing the online workload. This work first establishes a spatiotemporal integrated resource pool to reflect the computational resources provided by distributed computing continuum systems. Then, several pseudo-social welfare functions and marginal cost functions are constructed, where the latter is used to estimate the marginal cost of provisioning services to each newly arrived job based on the current resource surplus. We propose an online workload scheduling strategy namedOnSocMaxto solve the above problems. It operates by following the solutions to several convex pseudo-social welfare maximization problems and is proven to be$\alpha$-competitive for some$\alpha$with a value of at least 2. The evaluation results demonstrate thatOnSocMaxoutperforms several benchmark strategies in maximizing social welfare.

With the rapid development of the Internet, various network invasive behaviors are increasing rapidly. This seriously threatens the economic development of individuals, enterprises, and society. Network intrusion detection is important in network security systems, which can be regarded as a classification problem. It aims to distinguish between the specific categories of various network behaviors and determine whether the behavior belongs to network intrusion. However, network intrusions present a diverse and fast-changing trend, making categorizing difficult. Due to feature redundancy, uneven distribution of sample numbers, and inefficient parameter optimization, traditional rule-based approaches fail to achieve satisfying classification accuracy. This work proposes a multi-classification intrusion detection model based on Stacked Sparse Shrink AutoEncoder (SSSAE), Genetic Simulated annealing-based particle swarm optimization optimized Tabnet classifier (GS-Tabnet), and Decision Fusion (DF), called for SGTD short. Among them, SSSAE extracts multiple feature sets from the input data. Then GS-Tabnet trains a classifier for each feature set. Finally, the decision fusion fuses the results from these classifiers to obtain the final classification result. SGTD is compared with eight multi-classification benchmark models, and its intrusion detection accuracy is superior to its peers.

Water quality prediction methods forecast the future short or long-term trends of its changes, providing proactive advice for water pollution prevention and control. Existing water quality prediction methods only consider the historical data of single-type or multi-type water quality. However, meteorology and other factors also have a significant impact on water quality indicators. Therefore, only considering the historical data of water quality is not feasible. Unlike existing studies, this work proposes a hybrid water quality prediction model called CMI to solve the above problem. Before prediction, CMI incorporates a multimodal fusion mechanism of water quality time series and remote sensing images of meteorological rainfall. Moreover, CMI integrates the model of ConvNeXt V2 and a multimodal bottleneck transformer to extract image features for fusing the time series and images. Furthermore, it utilizes an emerging model of iTransformer to realize prediction with the fused features. Experimental results with real-life water quality time series and remotely sensed rainfall images demonstrate that CMI outperforms other state-of-the-art fusion algorithms, and the water quality prediction accuracy with fused meteorological data is 13% higher on average than that with only water quality time series.

The advent of smart manufacturing in Industry 4.0 signifies the arrival of the era of connections. As an excellent communication protocol, Object linking and embedding for Process Control Unified Architecture (OPC UA) can address most semantic heterogeneity issues. However, its semantics are not formally defined at the application layer. To address the information silo problem caused by semantic heterogeneity, a method named Querying of Ontology Mapping-based OPC UA (QOMOU) is proposed. It extracts the information models of OPC UA servers into resource description framework triples, utilizes web ontology language for semantic enrichment and inference, and employs a semantic similarity model for event ontology mapping to improve query efficiency. The method's effectiveness is validated through functional queries using the SPARQL protocol in Apache Jena. The query efficiency is 5% higher on average compared to both structured query and extensible markup languages. Moreover, by employing a keyword-matching algorithm, the query accuracy of the existing heterogeneous data integration scheme is improved by 4% on average. This enhancement can boost the operational efficiency of Internet of Things systems based on the OPC UA architecture.

Swarm intelligence and evolutionary algorithms are widely applied in industrial scheduling, mobile edge computing, etc due to their strong robustness and fast optimization speed. However, some real-world industrial optimization problems involve numerous decision variables, known as high-dimensional problems. Current algorithms often require considerable computational resources to evaluate objective function values because of high-dimensional decision spaces. Moreover, they are also prone to be trapped in local optima. To solve the above problems, this work proposes an improved algorithm named Surrogate-assisted Multi-class Collaborative Teaching and learning optimizer (SMCT). A multi-class collaborative teaching and learning optimizer is proposed as a base optimizer to improve exploration and exploitation abilities. Furthermore, an autoencoder-assisted radial basis function is proposed as the surrogate model to replace true function evaluations, thereby saving computational resources and balancing the complexity and accuracy in fitting true models. Finally, experimental results demonstrate that SMCT surpasses its existing peers in both search accuracy and convergence speed across eight high-dimensional benchmark functions.

Water pollution is continuously increasing in water ecosystems across all continents. Surface water sensors can record data on water quality indicators at regular intervals, and the associated water quality sequences show abnormal trends when extreme weather or unusual industrial discharges occur. Therefore, governments can take timely actions to minimize damage and protect the water environment by detecting these abnormal trends promptly. However, current methods make it difficult to interpret different correlations among water quality parameters effectively. To solve this problem, this work proposes a parameter correlation-aware anomaly detection model, which integrates Dual sliding windows, Convolutional LSTM, and a Deep neural network with dropout, called for DCLD short. First, DCLD designs dual sliding windows to capture local and global patterns within the sequence of water quality. Second, DCLD adopts a stacked long short-term memory with a convolutional neural network to capture complex features and long-term dependencies in the time series. Third, DCLD uses a deep neural network incorporating the dropout algorithm to extract abstract features. DCLD mitigates overfitting risks and enhances the model's generalization capacity. Finally, DCLD is evaluated with two real-world water quality datasets, and its anomaly detection accuracy is improved by 5.41% and 0.79% on average over its peers.

Smart mobile devices (SMDs) are integral for running advanced applications that demand significant computing resources and quick response time, e.g., immersive gaming and advanced image editing. However, SMDs often face constraints in computational capacity and battery duration, restricting their ability to process these tasks instantaneously. Cloud computing can circumvent these limitations by computation offloading, but cloud data centers (CDCs) are often deployed at long distances from users, which results in longer computational latency. To address the latency issue, the incorporation of small base stations (SBSs) in the vicinity of the user provides services with high bandwidth and low latency. The primary challenge lies in balancing the economics of the system consisting of different SMDs, SBSs, and a CDC, i.e., minimizing cost while still meeting the latency requirements of applications. In this work, a cost-minimized computation offloading framework is formulated and solved by a two-stage optimization algorithm named Lévy flight and Simulated Annealing-based Grey wolf optimizer (LSAG). The optimal edge selection strategy is defined in the first stage for dealing with the case of several available SBSs. The second stage coordinates task scheduling and optimizes the allocation of resources among SMDs, SBSs, and CDC. LSAG integrates the extended search property of Lévy flight and the individual selection strategy of simulated annealing in the grey wolf optimizer, which reduces the risk of falling into local optima and finds the global optimum. Experimental results of executing real-life tasks show that LSAG outperforms its state-of-the-art peers in terms of cost and speed of convergence.

Evolutionary algorithms are commonly used to solve many complex optimization problems in such fields as robotics, industrial automation, and complex system design. Yet, their performance is limited when dealing with high-dimensional complex problems because they often require enormous computational resources to yield desired solutions, and they may easily trap into local optima. To solve this problem, this work proposes a Self-adaptive Teaching-learning-based Optimizer with an improved Radial basis function model and a sparse Autoencoder (STORA). In STORA, a Self-adaptive Teaching-learning-based Optimizer is designed to dynamically adjust parameters for balancing exploration and exploitation during its solution process. Then, a sparse autoencoder (SAE) is adopted as a dimension reduction method to compress search space into lower-dimensional one for more efficiently guiding population to converge towards global optima. Besides, an Improved Radial Basis Function model (IRBF) is designed as a surrogate model to balance training time and prediction accuracy. It is adopted to save computational resources for improving overall performance. In addition, a dynamic population allocation strategy is adopted to well integrate SAE and IRBF in STORA. We evaluate it by comparing it with several state-of-the-art algorithms through six benchmark functions. We further test it by applying it to solve a real-world computational offloading problem.

High-dimensional expensive problems are often encountered in the design and optimization of complex robotic and automated systems and distributed computing systems, and they suffer from a time-consuming fitness evaluation process. It is extremely challenging and difficult to produce promising solutions in a high-dimensional search space. This work proposes an evolutionary optimization framework with embedded autoencoders that effectively solve optimization problems with high-dimensional search space. Autoencoders provide strong dimension reduction and feature extraction abilities that compress a high-dimensional space to an informative low-dimensional one. Search operations are performed in a low-dimensional space, thereby guiding whole population to converge to the optimal solution more efficiently. Multiple subpopulations coevolve iteratively in a distributed manner. One subpopulation is embedded by an autoencoder, and the other one is guided by a newly proposed Multi-swarm Gray-wolf-optimizer based on Genetic-learning (MGG). Thus, the proposed multi-swarm framework is named Autoencoder-based MGG (AMGG). AMGG consists of three proposed strategies that balance exploration and exploitation abilities, i.e., a dynamic subgroup number strategy for reducing the number of subpopulations, a subpopulation reorganization strategy for sharing useful information about each subpopulation, and a purposeful detection strategy for escaping from local optima and improving exploration ability. AMGG is compared with several widely used algorithms by solving benchmark problems and a real-life optimization one. The results well verify that AMGG outperforms its peers in terms of search accuracy and convergence efficiency.

Nowadays, smart mobile devices (SMDs) support various computation-intensive and delay-sensitive applications, e.g., online games, and figure compression. However, SMDs have limited computing resources and battery energy and cannot execute all tasks of the above applications in a real-time manner. Cloud computing provides enormous computing resources and energy that can easily execute tasks offloaded from SMDs. However, could data centers (CDCs) are often located in remote sites, which leads to long transmission time. Small base stations (SBSs) offer high-bandwidth and low-latency services for SMDs, which solves the problem of cloud computing. However, it becomes a challenge to achieve the lowest cost in such a heterogeneous architecture including multiple SMDs, SBSs, and the CDC while meeting delay requirements of tasks. This work proposes a cost-minimized computation offloading strategy to minimize the total cost of the system. A constrained optimization problem is first formulated based on the hybrid architecture. Afterward, a two-stage optimization algorithm called a Lévy flights and Simulated Annealing-based Grey wolf optimizer (LSAG) is developed to optimize the total cost of the system. In the first stage, the optimal edge selection policy is determined given multiple available SBSs. In the second stage, task offloading and resource allocation among SMDs, SBSs, and the cloud are determined. Experiments with real-life tasks prove that LSAG significantly achieves lower cost with faster convergence speed than state-of-the-art peers.