Artificial Intelligence Review最新文献

The cloud-edge-end collaboration system provides a new impetus for the development of intelligent transportation. In order to optimize the quality of service for intelligent transportation system users and improve system resource utilization, A three-tier caching strategy for cloud-edge-end collaboration based on efficiency collaboration task popularity (CSEPCA) was proposed, which exploits server resource characteristics and performs fine-grained cache replacements based on real-time task popularity to address the challenges associated with balancing server cache space and cost. To achieve an optimal balance between server cache space and cost, the problem of determining the availability of server cache space is formulated as a constrained markov decision process (CMDP), and an enhanced deep reinforcement learning algorithm based on soft updating (AT-SAC) was designed to achieve multi-objective optimization of system latency, energy consumption, and resource depletion rate, with the aim of improving service response speed and enhancing user service quality. To address challenges in effectively serving vehicles in areas with weak communication signals from cloud-edge servers, UAV swarms were introduced to assist with vehicle task offloading computations. A comprehensive optimization algorithm (Co-DRL-P) was proposed, which integrates enhanced deep reinforcement Learning (ERDDPG) and improved particle swarm optimization (A-PSO) algorithms to optimize UAV trajectories and communication angles, aiming to deliver superior service quality to users. Finally, we evaluate the performance of the proposed scheme through comprehensive simulation experiments. Specifically, when the number of users is 30, the system latency of the proposed scheme is 17.9%, 11.5%, 2.6%, and 60.2% lower than baseline schemes such as DQN, DDPG, TD3, and collaborative randomized schemes, and the system energy consumption is reduced by 20.6%, 15.9%, 9.4%, and 129.9%. Notably, the overall system cost for drone-assisted user offloading is reduced by approximately 49.6% in areas with weak cloud server signals.

This study introduces the Scientific Approach to Problem Solving-inspired Optimization (SAPSO) algorithm, a novel metaheuristic specifically designed for applications in civil engineering informatics. SAPSO imitates the structured process of scientific inquiry—covering problem review, hypothesis formulation, data collection, and analysis—to systematically explore complex search spaces. This approach enables SAPSO to reliably identify global optima. The algorithm’s performance was extensively tested against eleven leading metaheuristic algorithms using the IEEE Congress on Evolutionary Computation benchmark suites from 2020 (CEC 2020) and 2022 (CEC 2022). The comparison included the Artificial Bee Colony, Cultural Algorithm, Genetic Algorithm, Differential Evolution, Artificial Gorilla Troops Optimizer, Grey Wolf Optimizer, Particle Swarm Optimization, Red Kite Optimization Algorithm, Symbiotic Organisms Search, Teaching–Learning-Based Optimization, and Whale Optimization Algorithm. Statistical analysis with the Wilcoxon rank-sum test confirmed SAPSO’s superior results across these benchmarks. Additionally, this study presents a stacked ensemble machine learning framework called the SAPSO-Weighted Features Stacking System (SAPSO-WFSS), which combines SAPSO with two predictive models: a Radial Basis Function Neural Network and Least Squares Support Vector Regression. SAPSO is used to optimize both feature weights and model hyperparameters. Experiments on five diverse civil engineering case studies show that SAPSO-WFSS provides high accuracy, with Mean Absolute Percentage Error values as low as 2.4%, outperforming traditional methods. These findings demonstrate SAPSO’s potential as a powerful tool for improving prediction reliability in infrastructure maintenance and solving complex optimization problems in civil engineering.

The development of intelligent optimization methods has become a highly active research area in recent decades. This paper introduces a philosophy-inspired optimization algorithm called the Philosophical Proposition Optimizer (ΦPO), which models knowledge acquisition based on philosophical propositions in epistemology. In the proposed philosophical model, three developmental states for philosophical propositions, Justified True Belief (JTB), Possibly False Belief (PFB), and Unjustified True Belief (UTB), are iteratively refined using three specialized operators: Providing Justification (PJ), Raising Metaphysical Skepticism (RMS), and Raising Epistemic Skepticism (RES). To evaluate the performance of ΦPO on challenging optimization problems, it is applied to the single-objective bound-constrained benchmark problems of the IEEE Congress on Evolutionary Computation 2014 and 2024 (CEC 2014 and 2024), as well as to benchmark engineering problems. The performance of ΦPO is compared against five categories of algorithms: (1) widely used classical methods, (2) established post-2019 methods, (3) advanced PSO- and DE-based methods, (4) winners of CEC competitions, and (5) well-studied methods for solving engineering design problems. Two established non-parametric statistical methods, the Friedman test and the Wilcoxon signed-rank test, are used to analyze performance. The findings highlight the advantages of ΦPO across a range of numerical optimization problems, underscoring its competitiveness and potential in the field. Importantly, ΦPO was intentionally designed to be simple, interpretable, and parameter-free, avoiding complex adaptive strategies and extensive parameter tuning. It consistently delivers stable, high-quality solutions and exhibits fast convergence in many cases. The results demonstrate that ΦPO performs competitively across multiple benchmark suites, often ranking among the top-performing algorithms and outperforming several state-of-the-art methods, including recent CEC competition winners and engineering-specific optimizers. Its unique epistemic approach to solution refinement further enhances robustness, distinguishing it in both numerical and engineering optimization tasks.

A style change detection (SCD) task finds the locations of writing style changes within multi-authored documents. In 2017, the SCD task was introduced to assist in various applications, including cybercrime and literary analysis. This paper examines several document representation and prediction methods, including statistical, deep neural network (DNN), classical machine learning (ML), and hybrid methods. An analysis of existing datasets and the performance of SCD solutions is also included in this review. The findings demonstrate that the best method for attaining high performance is supervised ML, especially feed-forward neural networks with pretrained-based representations. Even though DNN models work well for other tasks, they are less frequently developed for this task, and their results need to be improved. This study details a set of challenges that are related to selecting features and adopting pretrained models. Additionally, the available literature is fairly limited in this task compared to others. For further investigation, several research directions are highlighted, including the design of SCD datasets with more realistic styles and the exploration of various learning techniques. This work encourages researchers to foster the growth and development of this research area.

The digitisation of hundreds of millions of herbarium specimen images and their labels has created an unprecedented resource for taxonomy, ecology, and conservation, motivating the development of artificial intelligence (AI) solutions. Automated analysis of these high-resolution scans faces significant challenges, including data imbalance, information loss, model interpretability and explainability, and scalable Open-Set Recognition (OSR). This paper provides an in-depth algorithm-level review of AI methodologies for herbarium image classification, tracing the development from classical classification models like Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to cutting-edge multimodal frameworks. In addition to classification, the review further investigates vision-based analytical tasks critical to herbarium image analysis, including specimen image segmentation, label text identification using Large Language Models (LLMs), and Human-in-the-Loop (HITL) quality assurance strategies. Furthermore, this review reveals practical challenges in specimen image analysis along with their promising solutions and potential future directions.

This paper embarks on an exploration into the large language model (LLM) datasets, which play a crucial role in the remarkable advancements of LLMs. The datasets serve as the foundational infrastructure analogous to a root system that sustains and nurtures the development of LLMs. Consequently, examination of these datasets emerges as a critical topic in research. In order to address the current lack of a comprehensive overview and thorough analysis of LLM datasets, and to gain insights into their current status and future trends, this survey consolidates and categorizes the fundamental aspects of LLM datasets from four perspectives: (a) pre-training corpora; (b) instruction fine-tuning datasets; (c) preference datasets; (d) evaluation datasets. The survey sheds light on the prevailing challenges and points out potential avenues for future investigation. Additionally, a comprehensive review of the existing available dataset resources is also provided, including statistics from 303 datasets, covering 8 language categories and spanning 32 domains. Information from 20 dimensions is incorporated into the dataset statistics. The total data size surveyed surpasses 774.5 TB for pre-training corpora and 700 M instances for other datasets. We aim to present the entire landscape of LLM text datasets, serving as a comprehensive reference for researchers in this field and contributing to future studies. Related resources are available at: https://github.com/lmmlzn/Awesome-LLMs-Datasets.

Following the introduction of data privacy regulations and “the right to be forgotten”, large language models (LLMs) unlearning has emerged as a promising data removal solution for compliance purposes, while also facilitating a diverse range of applications, including copyright protection, model detoxification and correction, and jailbreaking defence. In this survey, we present the taxonomy of existing LLMs unlearning algorithms, summarise unlearning evaluation methods including specialised benchmarks and threat models, and explore the applications of unlearning to provide a broad overview of the current state-of-the-art. We propose a novel problem formulation of LLMs unlearning with the additional unlearning objective: “robustness” to reflect the growing research interest in not only effectively and efficiently eliminating unwanted data, but also ensuring the process is performed safely and securely. To the best of our knowledge, we are the first to examine the robustness of unlearning algorithms as well as threat models for robustness evaluation, aspects that have not been assessed in past surveys. We also identify the limitations of the current approaches, including limited applicability to black-box models, vulnerability to adversarial attacks and knowledge leakage, and inefficiency, all of which require further improvement in future works. Furthermore, our survey highlights future directions for LLMs unlearning research, such as the development of comprehensive evaluation benchmarks, the movement towards robust unlearning and explainable AI for unlearning mechanisms, and addressing potential ethical dilemmas in unlearning governance.

In the field of anomaly detection in time series, remarkable advances based on deep learning methodologies and, more specifically, reconstruction-based methods have been proposed. These methods are particularly valuable, as they can capture the fundamental structure of the data and enable the detection of subtle anomalies that traditional techniques might overlook. Reviews of the existing literature discuss anomaly detection from a general perspective and consider a specific anomaly type, hindering the process of proposing new and better algorithms. This paper focuses on reconstruction-based methods in isolation, as they have been demonstrated to present the best performance of the three main groups in deep anomaly detection models described so far. Thus, it intends to extend the literature by addressing in detail reconstruction-based methods for anomaly detection in multivariate time series, to provide richer information about these methods, and to include extensive experimentation, not usually performed in existing surveys on the topic. Finally, the paper presents useful insights (strengths and weaknesses of the methods) extracted from the experimental study, and significant challenges and future research directions related to the potential of these methods in anomaly detection.

An original mathematical theory-inspired optimization algorithm, named multi-sequence discrete recursive descent optimizer (MDRDO) is proposed for numerical optimization and engineering issues. The main inspiration is derived from the deterministic recursive descent method, composed of descent directions and step lengths. Firstly, by adopting multiple initial sequences, a multiple-sequence searching framework (MSF) is constructed to explore the feasible space fully. Then, according to the Taylor expansion, the approximate discrete gradient is calculated to form a discrete quasi-Newton direction (BFGS) and the imprecise search step length for each sequence. Furthermore, to enhance the ability to escape local optimums, two random searching strategies are employed to keep the diversity of candidate solutions. Hence, the MDRDO is constructed by merging the quasi-Newton direction, imprecise step length and random strategies, in which suitable descent directions and step lengths help the MDRDO to explore a promising region and exploit the optimal solutions. The effectiveness of the MDRDO is verified by the comparison experiments with other 10 excellent algorithms on two typical test suites. Results reveal that the MDRDO outperforms others for 62.07% among 29 functions of CEC2017 test suite and 80% among 10 functions of CEC2020 test suite. Average convergence curves prove the ability of MDRDO to approach better solutions quickly. The applicability of the MDRDO is tested on 5 engineering problems with the other 9 widely used algorithms, in which the MDRDO ranks first for all problems with excellent comprehensive performance. Overall, the test results demonstrate that MDRDO is a promising tool for addressing practical optimization problems.