Adaptive Neural Architectures for Real-Time Intelligent Decision Systems

Abstract

Real-time intelligent decision-making has become increasingly critical in dynamic and data-intensive environments such as finance, healthcare, and smart infrastructure. However, conventional deep learning (DL) models often struggle to adapt to rapidly changing data distributions and often fail to deliver the low-latency responses required for time-sensitive applications. To address these limitations, this research proposes an adaptive neural architecture framework designed for real-time intelligent decision systems. The proposed model integrates dynamic structural adaptation with a Quokka Swarm Optimized Adaptive Recurrent Neural Network (QSO-ARNN) to effectively capture temporal dependencies while continuously adjusting model parameters based on streaming data inputs, which contain 9,500 records from dynamic environments. Z-score normalization is utilized to normalize the dataset, while feature extraction using Linear Discriminant Analysis (LDA) improves the model's capacity to differentiate between various classes. The proposed approach is evaluated on large-scale, high-velocity data streams, simulating real-world decision-making scenarios. It enhances accuracy and responsiveness through data normalization and feature extraction for high-throughput, low-latency predictions. With a throughput of 0.1565 Gbps, a latency of 0.030 seconds, a root mean square error (RMSE) of 0.0154, a mean absolute error (MAE) of 0.0125, a mean absolute percentage error (MAPE) of 6.10%, and a high coefficient of determination that was stimulated in Python, the experimental results show excellent performance. Overall, the proposed adaptive neural architecture provides a scalable, efficient, and high-performing solution for real-time intelligent decision-making, significantly outperforming traditional static models in both accuracy and responsiveness.