Autonomous mobile robots (AMRs) must operate in dynamic, unstructured environments where traditional control and reinforcement learning (RL) face adaptability and reward design limitations. This study proposes a hybrid framework combining proximal policy optimization (PPO) with a large language model (LLM) as an adaptive reward designer. Using GPT-4o-mini, the LLM dynamically shapes rewards based on performance logs, improving exploration and stability. Experiments in complex indoor navigation show the LLM-PPO model reduces collisions by 38%, shortens completion time by 21%, and increases rewards by 8% over PPO. Results demonstrate LLM-RL integration enhances safety, efficiency, and consistency, offering a promising paradigm for AMR control.
In education reform, a growing number of courses are being taught in conjunction with online platforms, posing new challenges for predicting students' performance. This study focuses on ideological and political education courses delivered through a massive open online courses platform. This paper gathered student behavior data, screened relevant features, and proposed an improved beetle antennae search-backpropagation neural network (IBAS-BPNN) method to forecast students' grades after learning the course. Moreover, experiments were conducted using the collected dataset. The findings indicated that the features selected based on information gain rate exhibited superior predictive capabilities compared to those selected using the Pearson correlation coefficient. By utilizing the top ten features as inputs to the IBAS-BPNN model, the model achieved a macroP value of 88.86%, a macroR value of 87.52%, and a macroF1-score of 88.18%. These results outperformed other optimized BPNN algorithms and surpassed machine learning approaches such as k-nearest neighbor and decision tree. The outcomes validate the reliability of the proposed method for grade prediction and its potential applicability in practice.
This paper mainly studied the loneliness of international students in China under the influence of social support and mobile network media. Data were collected through a questionnaire survey utilizing the perceived social support scale (PSSS), mobile network media use scale, and University of California, Los Angeles (UCLA) loneliness scale. The data were analyzed using SPSS version 22.0. It was found that the total score of social support was 5.15 +/- 1.12 points, and the school support score was low. The use intensity of mobile network media was 4.06 +/- 0.55 points, and the use behavior was 4.08 +/- 0.56 points, which was above the medium level. The total score of loneliness was 45.84 +/- 7.45 points. No remarkable differences were observed in gender, length of stay in China, and Chinese proficiency (p>0.05). Correlation and regression analyses revealed that social support and mobile network media were significantly correlated with loneliness of international students. The findings confirm that social support and mobile network media have an impact on loneliness, suggesting that loneliness can be alleviated through improving social support and other means.
Early detection of Parkinson's disease (PD) remains a critical challenge in healthcare, as current diagnostic methods often rely on subjective clinical assessments when motor symptoms have already progressed. This paper presents a novel deep learning approach for automated PD detection using digital handwriting analysis. We propose a dual-stream attention-enhanced bidirectional long short-term memory (BiLSTM) network that separately processes spatial trajectory features and dynamic behavioral features through dedicated pathways. Our architecture employs stream-specific attention mechanisms to identify discriminative temporal segments within each feature modality, followed by an attention-based fusion layer that integrates complementary representations from both streams. Extensive experiments on a comprehensive dataset of handwriting samples from both PD patients and healthy controls demonstrate that our approach achieves state-of-the-art performance with 94.7% accuracy and 0.9872 AUC. The attention mechanism provides interpretable insights into which temporal phases of handwriting are most indicative of PD symptoms, offering potential clinical value for understanding disease progression. Our findings suggest that automated analysis of digital handwriting can serve as an effective, non-invasive screening tool for early PD detection.
Plant leaf automatic recognition is a challenging issue in both computer vision and plant taxonomy areas, and its keys are extracting discriminative features effectively from leaf images and matching them fast. To deal with the scale parameter selection problem in multi-scale feature description methods, an effective rule is proposed to generate constant scale parameters for leaf image recognition. Given the current scale, a new scale can be determined through the trisection of the leaf contour. And for each point on the leaf contour, their paired neighbor points under each scale are located using these scale parameters. Angle values calculated by these neighbor points can describe a leaf in a multi-scale manner. The final multi-scale leaf contour descriptor is made compact using fast Fourier transform. In the matching stage, support vector machine and Manhattan distance are used for recognition and retrieval tasks. Leaf recognition and retrieval experiments were conducted with standard evaluation metrics on three well-known datasets, called Swedish, Flavia, and ImageCLEF2012 leaf datasets. The method achieves 97.27% accuracy on Swedish, 93.78% on Flavia, and 0.597 S-score on ImageCLEF2012 dataset. Experimental comparisons with notable methods show that the proposed method achieved higher recognition accuracy and computation speed.
Coordination plays a critical role in basketball training, significantly influencing athletes' performance in actual games. To enhance the overall technical and tactical application level of Liaoning Petrochemical University women's basketball players, this study investigates the evaluation of training coordination. Because traditional backpropagation (BP) neural network evaluation methods exhibit low accuracy, a radial base function (RBF) evaluation prediction model based on cooperative game theory is proposed. An evaluation index system for the training coordination of women's basketball players is established to calculate the weights of each index and rank their importance, and the RBF neural network is utilized for evaluation and prediction. Using actual training data from the university's female basketball players, the study analyzes balance, reaction ability, agility, and processing speed during training. Simulation experiments demonstrate that the proposed method achieves considerable accuracy, validating its effectiveness. This provides a robust technical tool for real-time monitoring of training outcomes.
In response to the current situation where pipeline basic data is poorly digitized, under-visualized, and fragmented, a pipeline data monitoring and integrated management platform is being developed to promote the digitalization of pipeline management information. This will enable dynamic updates and real-time querying of basic data, improving the ability to control and utilize the data, which is very important for pipeline projects. The design plans to use the Hefei Gas Transmission Branch as the pilot unit for data entry and application. This study explains the platform's overall design goals, framework, database structure, and data weighting analysis methods. Based on Internet-of-Things technology, this platform manages data for the entire lifecycle of pipelines. It includes real-time data entry and querying across ten key categories: pipeline centerline, pipeline body, auxiliary facilities, third-party facilities, environment, risk, monitoring and maintenance, protection, disaster prevention, and emergency management. The integrated query function allows data association and traceability, while geographic information system (GIS) and mobile apps support data sharing and online viewing through spatial information. It enables dynamic updates of basic data, supports real-time querying, and enhances the ability to control and utilize basic data effectively.
With the proliferation of the Internet of Things (IoT), the exponential growth in wireless network traffic has made efficient resource management crucial. A key challenge is to simultaneously achieve high throughput and energy efficiency, particularly because most IoT devices have limited battery capacities. Previous studies have focused on optimizing throughput but have often overlooked the additional energy consumption incurred by their proposed methods. To address this limitation, this study proposes a traffic-aware clustering and powersaving mechanism (TAC-PSM) and a novel scheduling scheme based on an aggregated medium access control protocol data unit (A-MPDU). The proposed technique clusters nodes based on the channel load. It minimizes unnecessary power consumption by transitioning idle nodes to sleep mode, while maximizing throughput using aggregated data transmission. The performance evaluation results show that the TAC-PSM improved the average network throughput by 799.73% compared to the conventional model and reduced the average energy consumption by 91.45%. This leads to a 15.68 times increase in the overall efficiency, demonstrating that the proposed scheme is an effective solution for large-scale IoT environments.
With the rapid development of intelligent technology, the teaching system is gradually transitioning towards personalization and adaptability. In this context, building an intelligent English teaching platform based on fuzzy logic and semantic analysis is crucial for improving learning efficiency and teaching quality. An English intelligent teaching platform that combines a fuzzy clustering system and semantic entity analysis is proposed. The platform classifies users through fuzzy clustering analysis and recommends personalized exercise recommendations based on user feedback dynamics. The results indicated that the model significantly improved learning performance, with 86.67% of students showing better performance. The proposed model increased the number of students by 13.33%-20.01% compared with the two comparison algorithms. This result is of great significance for the research and practice of English intelligent teaching platforms. Overall, the teaching platform not only provides innovative solutions at the technical level, but also has practical benefits in improving teaching efficiency and promoting students' academic performance in practical applications, providing strong support for intelligent teaching.
A new method for non-local random denoising using slice sampling is introduced. This method can significantly enhance the efficiency of non-local image denoising algorithms. The proposed algorithm consists of two stages: first, similar image patches are searched using slice sampling, and then a denoising algorithm is designed to reconstruct the original image using these similar patches. Low-rank matrix approximation methods are used to obtain estimates of clean patches, and a clean denoised image is generated through superposition. The theoretical analysis and experimental tests demonstrate that this algorithm can overcome the dependence on proposal distributions in traditional random algorithms. The experimental results on benchmark images with additive Gaussian noise show that the proposed method can achieve good performance compared to state-ofthe-art methods such as BM3D. Specifically, for the test image "Lena" with a noise standard deviation of 20, this method can achieve a peak signal-to-noise ratio (PSNR) of 32.82 dB and a structural similarity index measure (SSIM) of 0.87. For the "Barbara" image with the same noise level, the PSNR is 31.50 dB and the SSIM is 0.89. These results confirm the effectiveness of the algorithm in denoising performance and edge preservation.
The unconstrained two-dimensional cutting stock problem is a typical NP-hard problem with high complexity. When generating the layout, both the utilization rate of the sheet and simplification of the cutting process must be considered. This study presents an algorithm for generating a same-size T-shape layout. The same-size Tshape layout is suitable for the cutting and punching processes in actual production. It includes two segments in different directions, each consisting of strips of only one size and direction. First, the algorithm determines the optimal same-size strip through dynamic programming; subsequently, it determines the layout of the same-size strip in the composite strip and the composite strip in the segment by solving the knapsack problem. Finally, two segments are selected to generate the layout with the highest piece value. Using 37 benchmark test problems from the literature, the proposed algorithm was compared with four advanced and effective layout algorithms. Our algorithm achieved optimal results for 16 test problems, and the ratio of the calculated results to the optimized results for the remaining test problems reached 99.9%. The average calculation time for each test problem was only 1.3 seconds. The experimental results indicate that the proposed algorithm offers advantages in terms of the computation time and sheet utilization rate. The algorithm not only achieves good optimization results within a reasonable time but also simplifies the cutting process while meeting the engineering requirements.
This paper presents a concise introduction to the cultural and creative packaging design and the "Four Symbols" motif, followed by an analysis of two designs. In the analysis process, a convolutional neural network (CNN) algorithm was applied to generate preliminary design scores, and then further evaluation was conducted according to the scores. It was found that the CNN algorithm could be used for preliminary scoring of the packaging design. The CNN-generated scores for both designs showed strong alignment with manual evaluation scores. The graphic and color matching of the "Four Symbols" motif in the two designs enhanced the cultural heritage of the products and effectively engaged consumer interest.
Machine reading comprehension (MRC) is a fundamental task in natural language processing (NLP), with existing models struggling to capture long-range dependencies and handle complex semantic nuances, particularly in Chinese. This paper proposes the Collaborative Semantic Reader (C-S Reader), a novel model that combines RoBERTa_wwm_ext pre-training and multi-level attention mechanisms to enhance semantic understanding. Experiments on the DuReader2 dataset show that C-S Reader significantly outperforms baseline models in both the Rouge-L and BLEU-4 scores, demonstrating its effectiveness in processing long documents and capturing complex semantic relationships. Our work provides a scalable solution for Chinese MRC tasks and highlights future challenges, including long-range dependency modeling and ambiguity in complex questions.
Topic modeling has evolved from statistical methods such as latent Dirichlet allocation (LDA) to neural hybrid models including BERTopic, which utilize bidirectional encoder representations from transformers (BERT) embeddings. However, traditional statistical evaluation metrics overlook the semantic richness of these neural representations, limiting model assessment capabilities. This paper introduces semantic-based evaluation metrics that leverage deep learning embeddings and validates them through both statistical comparison and large language model (LLM)-based assessment. This study evaluated three synthetic datasets with systematically varying topic overlap and one public dataset (20 Newsgroups). Analysis across 9,608 synthetic documents with 45 topics and a stratified sample of 1,000 documents from 20 Newsgroups shows that semantic metrics achieve improved discrimination compared to statistical baselines. Specifically, semantic coherence shows a 38.1% discriminative range versus 5.0% for statistical measures, representing a 7.62 & times; improvement. Semantic distinctiveness achieves 1.57 & times; higher discrimination than statistical methods. Semantic methods also maintain consistent discrimination quality for diversity metrics, with stable progression across similarity levels. LLM assessments, serving as proxies for human judgment, demonstrate inter-model agreement through a weighted three-model ensemble (mean pairwise Spearman rho=0.937) and positive correlation with semantic metrics on public datasets (rho=0.632-0.671). Domain-specific validation and multilingual extension constitute future work.
Bullet-screen videos contain rich user-generated data. It is of great practical significance to utilize sentiment analysis technology and topic models to identify video topics by fusing multi-dimensional features. A novel video recommendation method (MSSA) based on multi-source sentiment analysis is proposed by fusing the sentiment features of bullet screens and the topic features of video subtitles. Firstly, the method performs sentiment analysis on the bullet screens and subtitles of videos, and constructs a user sentiment feature matrix and a video sentiment feature matrix. Secondly, the method clusters user groups with similar sentiment tendencies by extracting the temporal information of bullet screens posted by users. Next, the topic feature vectors of subtitle texts in video clips are calculated to obtain the topic similarity matrix among videos by fusing with the video label information. Afterwards, a sentiment-oriented video set is generated according to the differences in sentiment polarity between bullet screens and subtitles. Finally, online recommendation of bullet-screen videos is achieved by introducing a recommendation heat index. The MSSA method is validated on real-world datasets, and it conducts comparative experiments with other state-of-the-art methods to assess its recommendation coverage and accuracy. The experimental results show that the MSSA method can effectively enhance the performance of user sentiment clustering and the semantic alignment of video content topics, enabling it to effectively explore user interest characteristics, and optimize the quality of personalized video recommendation services.
This paper introduces D-MIX, a novel blockchain-based platform to decentralize mixology by enabling sharing, execution, and monetizing original drink recipes. D-MIX enables creators (baristas and bartenders) to securely upload recipes, which then get autonomously reproduced using smart machines called D-MIXers. These drink dispensing machines create drinks according to recipes with precision and consistent quality aided by the DMIX AI engine. When recipes are used, fair royalty distribution is guaranteed via smart contracts. By integrating blockchain technology with AI-assisted barista/bartender machines, D-MIX bridges the physical and digital realms, offering a solution that can democratize access to exclusive recipes that are only available to locals. Extensive evaluation in the testbed and pilot trials validate the system's performance and effectiveness that it can achieve high precision in recipe reproduction and recreate real-world recipes.
Currently, most Internet of Things (IoT) resource-allocation solutions are based on centralized management, rendering it difficult to successfully establish diverse and dynamic IoT networks. Consequently, the fields of reinforcement learning and distributed computing require additional technological advancements. We propose an innovative approach for slot scheduling in IoT networks. This approach focuses on the utilization of distributed resource blocks. The purpose of this work is to demonstrate, via simulations, the influence of distributed slot assignment on the signal-to-interference ratio (SIR) and the probability of accidents occurring. The results of this research suggest that the proposed approach, in which each device in the IoT network is provided with an appropriate slot that possesses acceptable SIR levels, was successful. As the process of distributed slot allocation progresses, it is beneficial to build network convergence by utilizing the learning capabilities of each device. This was accomplished using a distributed slot-allocation process. Therefore, the existing bandwidth can be utilized more efficiently.
Simultaneous localization and mapping (SLAM) is a core technology in robotics and autonomous navigation. Visual SLAM has gained attention through advances in object recognition enabling landmark-based mapping. However, a critical gap exists in handling inconsistencies arising from repeated text recognition processes, a challenge that existing studies have largely overlooked. To address this, we propose an innovative algorithm that enhances text recognition accuracy for mobile platforms. Our method excels in tracking objects and consistently recognizing and determining textual content in recurrent scenes. Through rigorous experimentation, we demonstrate that our algorithm significantly improves real-time text recognition accuracy by mitigating errors inherent in conventional approaches. This advancement not only refines the reliability of text-based Visual SLAM but also broadens its applicability in dynamic, text-rich environments. Our work paves the way for more robust and efficient autonomous navigation systems, particularly in urban landscapes where textual cues are abundant.
This paper presents a study that integrates two primary methodologies for investigating automated proofreading: one employing a state-of-the-art syntactic analyzer and another based on a sophisticated semantic hierarchical model. The analyzer meticulously processed the input text, extracting part-of-speech tags and intricate syntactic dependencies, while the semantic hierarchical model was used to perform an innovative analysis. The resulting integration scheme of syntactic rigor and semantic depth represents a paradigm shift in error detection, outperforming the benchmarks established in the literature and state-of-the-art software systems. The integration of the syntactic structure with semantic understanding resulted in a marked increase in error detection accuracy. In particular, the study demonstrated a substantial enhancement in the average Fmeasure, surpassing Kingsoft WPS (2024) by 39% and Microsoft Word (2024) by 55%. It is worth noting that for error types that have historically been difficult to improve F-measure, particularly word ambiguity, the study achieved a 65% increase in detection accuracy.
Knowledge graphs are crucial for numerous applications, but their frequent incompleteness limits their utility. Few-shot knowledge graph completion (FKGC) addresses this by learning to infer new facts from only a handful of examples. However, existing FKGC methods are highly vulnerable to noisy or inconsistent reference examples, which can severely degrade model performance. To overcome this critical challenge, we introduce ATMR, an attention-based meta-relational learning framework. ATMR incorporates a novel attention mechanism that strategically identifies and upweights the most informative reference triples while diminishing the influence of potential noise. This allows for the construction of more robust and accurate relation representations. Rigorous experiments on two public datasets demonstrate that ATMR consistently outperforms baselines. Notably, it achieves an 8.5% improvement in the Hits@10 metric for 5-shot completion on the NELL-One dataset, validating its superior ability to handle noise in few-shot scenarios.
