A Deep Learning Framework for Real-Time Intelligent Decision-Making Systems

Abstract

Real-time intelligent decision-making remains one of the most demanding challenges in modern artificial intelligence, particularly in dynamic,
resource-constrained environments where latency and accuracy must be simultaneously optimised. Existing approaches often sacrifice
interpretability or temporal responsiveness for predictive performance, leaving a critical gap in operational deployment. This study proposes a novel
hybrid deep learning framework, termed Temporal Attention-Guided Decision Network (TAG-DN), which integrates convolutional neural networks
(CNN), long short-term memory (LSTM) units, and a multi-head self-attention mechanism to enable precise, low-latency decision-making across
sequential data streams. The framework was evaluated on three benchmark datasets—KITTI autonomous driving sequences, UCF-Crime
surveillance streams, and a custom industrial IoT sensor dataset comprising 120,000 timestamped records. TAG-DN achieved a mean decision
accuracy of 94.7%, outperforming state-of-the-art baselines including standard LSTM (88.3%), Transformer-only models (91.2%), and CNN-BiLSTM
hybrids (92.6%). Average inference latency was measured at 11.4 milliseconds per decision cycle on an NVIDIA RTX 3080 GPU. The attention
mechanism improved temporal feature localisation by 18.3% over non-attention baselines. Ablation studies confirmed the complementary
contribution of each architectural component. The proposed framework demonstrates strong generalisation across diverse real-time application
domains and offers a scalable, deployable solution for intelligent systems requiring rapid, data-driven decisions.
Keywords: ; ; ; ; ; intelligent systems; temporal data