Applied AI letters最新文献

Accurately predicting stock prices remains a major challenge in financial analytics due to the complexity and noise inherent in market data. Feature selection plays a critical role in improving both computational efficiency and predictive performance. In this study, we introduce a novel hybrid framework that integrates metaheuristic feature-selection algorithms with an enhanced transformer-based prediction model fine-tuned using temporal embedding and adaptive attention pruning. We evaluate and compare the effectiveness of three nature-inspired metaheuristic algorithms: bat algorithm (BAT), gray wolf optimization (GWO), and beluga whale optimization (BWO) for selecting the most informative features from a time-series stock dataset. After feature selection, the optimal subsets are fed into our modified transformer equipped with temporal embeddings and adaptive attention pruning. Extensive experiments conducted on the Bharat Heavy Electricals Limited (BHEL) dataset show that the proposed hybrid framework outperforms traditional methods in terms of predictive accuracy. Among the evaluated approaches, the combination of BWO and the fine-tuned transformer achieves the best performance, yielding a Test RMSE of 0.0030 and a Test MAPE of 0.0108, demonstrating the superiority of BWO in identifying relevant features. This work provides a comprehensive comparative analysis of hybrid metaheuristic–deep learning models for stock price prediction and offers a foundation for integrating more explainable and scalable AI techniques into financial forecasting.

Artificial Intelligence (AI) is playing an increasingly vital role in the Industrial Internet of Things (IIoT), enabling predictive analytics, real-time monitoring, and autonomous operations across industries such as manufacturing, logistics, and energy. However, widespread adoption is hindered by technological, organizational, and infrastructural challenges. This paper examines the adoption, application, and challenges of AI–IIoT environments, focusing on implementation domains, adoption drivers, enabling technologies, and key barriers. We conducted a Systematic Literature Review (SLR using PRISMA). Peer-reviewed English-language journal articles published between 2018 and 2025 were sourced from ScienceDirect, Web of Science (WoS), Scopus, IEEE Xplore, Springer, Google Scholar, Elsevier, and Taylor & Francis. After applying inclusion criteria and screening procedures, 46 relevant journal articles were included for analysis. Key AI applications identified include predictive maintenance, anomaly detection, real-time monitoring, autonomous process control, and smart supply chains. Adoption is facilitated by external enablers 5G infrastructure, regulatory support, and internal factors, organizational readiness, and workforce skills. Challenges include data quality issues, cybersecurity risks, legacy system integration, and limited model scalability. Technologies such as edge computing, cloud platforms, and federated learning are instrumental in mitigating these challenges. While adoption is growing, significant barriers remain. AI has the potential to drive operational efficiency and innovation in IIoT, provided these constraints are addressed. This paper offers a comprehensive taxonomy of AI applications and proposes a framework of adoption factors, offering valuable insights for researchers, practitioners, and policymakers involved in AI-driven industrial transformation.

Reinforcement learning (RL) has shown to be effective for simple automated cyber defence (ACD) type tasks. However, there are limitations to these approaches that prevent them from being deployed onto real-world hardware. Trained RL policies will often have limited transferability across even small changes to the environment setup. Instability during training can prevent optimal learning, a problem that only increases as the environment scales and grows in complexity. This work looks at addressing these limitations with a zero-shot transfer approach based on multi-agent RL. This is achieved by partitioning the task into smaller network machine subtasks, where agents learn the solution to the local problem. These local agents are independent of the network scale and can therefore be transferred to larger networks by mapping the agents to machines in the new network. Initial experiments show that this transfer method is effective for direct application to a number of ACD tasks. It is also shown that its performance is robust to changes in network activity, attack scenario and reduces the effects of network scale on performance.

Reinforcement learning (RL) solutions have shown considerable promise for automating the defense of networks to cyber attacks. However, a limitation to their real world deployment is the sample efficiency and generalizability of RL agents. This means that even small changes to attack types require a new agent to be trained from scratch. Meta-learning for RL aims to improve the sample efficiency of training agents by encoding pre-training information that assists fast adaptation. This work focuses on two key meta-learning approaches, MAML and ML3, representing differing approaches to encoding meta learning knowledge. Both approaches are limited to sets of environments that use the same action and observation space. To overcome this, we also present an extension to ML3, Gen ML3, that removes this requirement by training the learned loss on the reward information only. Experiments have been conducted on a distribution of network setups based on the PrimAITE environment. All approaches demonstrated improvements in sample efficiency against a PPO baseline for a range of automated cyber defense (ACD) tasks. We also show effective meta-learning across network topologies with Gen ML3.

This research aims to investigate the flying kinematics of bats, drawing inspiration from nature to benefit humans in achieving vision-independent flights. The project uses 50 cameras mounted inside a specially designed tunnel to analyze bat flying behavior. However, the simultaneous recording of all cameras generates numerous unnecessary image frames, prompting the need for a highly accurate filter. The study focuses on developing this filter using deep learning techniques to classify images based on the bat's location within each frame. We employed widely used one-stage object detection algorithms: YOLOv4, YOLOv4-tiny, and YOLOv5. Notably, YOLOv4, even though it is the older version we adapted, fulfills the intended objective of the study with better accuracy, for two different datasets, along with a higher learning process. Then, we closely looked at the tunable hyperparameters and augmentation techniques of the YOLOv4 model and adapted them in hyperparameter tuning on YOLOv5. Then we examined the impact of these hyperparameters and data augmentation techniques on the performance of YOLOv5L. Based on the result, we adapted the hyperparameters and data augmentation techniques, which have a positive impact on the YOLOv5 model. Improved YOLOv5L achieved better performance with a mean average precision (mAP) of 99.3% where the average precision (AP) of each classification scored more than 99% along with an auto anchor detection mechanism.

This study introduces a dynamically memory-adjusted whale optimization algorithm (DMA-WOA) for feature selection in polycystic ovary syndrome (PCOS) diagnosis. To overcome the standard WOA's limitations in balancing exploration and exploitation, DMA-WOA incorporated adaptive memory control to improve convergence stability and computational efficiency. In DMA-WOA adaptive control dynamics adjusted memory size and influence based on population diversity and fitness change, enabling consistent convergence in high-dimensional clinical data. The framework was evaluated on the only publicly available PCOS electronic health records dataset using diverse classifiers, including SVM, RF, LR, MLP, RNN, LSTM, GRU, TabTransformer, and TabNet. Results showed that DMA-WOA achieved superior accuracy, generalization, and runtime efficiency compared to baseline and standard WOA approaches, while comparative analysis with existing metaheuristics confirmed its enhanced optimization robustness and diagnostic reliability.

Timely and accurate diagnosis of lung diseases is critical for reducing related morbidity and mortality. Lung ultrasound (LUS) has emerged as a useful point-of-care tool for evaluating various lung conditions. However, interpreting LUS images remains challenging due to operator-dependent variability, low image quality, and limited availability of experts in many regions. In this study, we present a lightweight and efficient deep learning model, ParSE-CNN, alongside fine-tuned versions of VGG-16, InceptionV3, Xception, and Vision Transformer architectures, to classify LUS images into three categories: COVID-19, other lung pathology, and healthy lung. Models were trained using data from public sources and Ugandan healthcare facilities, and evaluated on a held-out Ugandan dataset. Fine-tuned VGG-16 achieved the highest classification performance with 98% accuracy, 97% precision, 98% recall, and a 97% F1-score. ParSE-CNN yielded a competitive accuracy of 95%, precision of 94%, recall of 95%, and F1-score of 97% while offering a 58.3% faster inference time (0.006 s vs. 0.014 s) and a lower parameter count (5.18 M vs. 10.30 M) than VGG-16. To enhance input quality, we developed a preprocessing pipeline, and to improve interpretability, we employed Grad-CAM heatmaps, which showed high alignment with radiologically relevant features. Finally, ParSE-CNN was integrated into a mobile LUS workflow with a PC backend, enabling real-time AI-assisted diagnosis at the point of care in low-resource settings.

Remote sensing has become a vital tool across sectors such as urban planning, environmental monitoring, and disaster response. Although the volume of data generated has increased significantly, traditional vision models are often constrained by the requirement for extensive domain-specific labelled data and their limited ability to understand the context within complex environments. Vision Language Models offer a complementary approach by integrating visual and textual data; however, their application to remote sensing remains underexplored, particularly given their generalist nature. This work investigates the combination of vision models and VLMs to enhance image analysis in remote sensing, with a focus on aircraft detection and scene understanding. The integration of YOLO with VLMs such as LLaVA, ChatGPT, and Gemini aims to achieve more accurate and contextually aware image interpretation. Performance is evaluated on both labelled and unlabelled remote sensing data, as well as degraded image scenarios that are crucial for remote sensing. The findings show an average MAE improvement of 48.46% across models in the accuracy of aircraft detection and counting, especially in challenging conditions, in both raw and degraded scenarios. A 6.17% improvement in CLIPScore for comprehensive understanding of remote sensing images is obtained. The proposed approach combining traditional vision models and VLMs paves the way for more advanced and efficient remote sensing image analysis, especially in few-shot learning scenarios.

For identifying foliar diseases in crops at an early stage, accurate detection is necessary in maintaining food security, minimizing economic losses, and cultivating sustainable agriculture. In staple crops, potato is highly vulnerable to lethal diseases like Early Blight and Late Blight that can drastically affect both the quality and the quantity of the yield. Conventional diagnostic procedures using visual observation and/or laboratory examinations are frequently tedious, time-consuming, and susceptible to error. To address these problems, in this research, we propose a novel deep learning architecture using a customized convolutional neural network (CNN) for classifying potato leaf images into three distinct classes, namely Early Blight, Late Blight and Healthy. The model is trained on a selective and heavily augmented subset of the PlantVillage dataset containing 11,593 images and further optimized using regularization techniques like dropout and batch normalization. The system architecture is intended to keep the tradeoff between performance and computational efficiency, so as to fit real-world agricultural scenarios. To increase interpretability and improve trust, we use the Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the regions in space of the leaves that most contribute to the prediction of the model. The experimental results show superior performance and the proposed model reaches 99.14% accuracy and close-to-perfect precision, recall and F1-scores in all of the classes. Grad-CAM visualizations validate that the model is robust in attending to biologically meaningful regions for the disease symptoms. In addition, we perform comparative analyses against recent state-of-the-art models, and demonstrate that the proposed approach outperforms the others in accuracy and interpretability.

Cyber-attacks pose a security threat to military command and control networks, Intelligence, Surveillance, and Reconnaissance (ISR) systems, and civilian critical national infrastructure. The use of artificial intelligence and autonomous agents in these attacks increases the scale, range, and complexity of this threat and the subsequent disruption they cause. Autonomous Cyber Defence (ACD) agents aimto mitigate this threat by responding at machine speed and at the scale required to address the problem. Additionally, they reduce the burden on the limited number of human cyber experts available to respond to an attack. Sequential decision-making algorithms such as Deep Reinforcement Learning (RL) provide a promising route to create ACD agents. These algorithms focus on a single objectivesuch as minimising the intrusion of red agents on the network, by using a handcrafted weighted sum of rewards. This approach removes the ability to adapt the model during inference, and fails to address the many competing objectivespresent when operating and protecting these networks. Conflicting objectives, such as restoring a machine from a back-up image, must be carefully balanced with the cost of associated down-time or the disruption to network traffic or services that might result. Instead of pursuing a Single-Objective RL (SORL) approach, here we present a simple example of a multi-objective network defense game that requires consideration of both defending the network against red-agents and maintaining the critical functionality of green-agents. Two Multi-Objective Reinforcement Learning (MORL) algorithms, namely Multi-Objective Proximal Policy Optimization (MOPPO) and Pareto-Conditioned Networks (PCN), are used to create two trained ACD agents whose performance is compared on our Multi-Objective Cyber Defense game. The benefits and limitations of MORL ACD agents in comparison to SORL ACD agents are discussed based on the investigations of this game.