In this paper, we present a comprehensive framework for Neural Architecture Search (NAS), a cutting-edge technique in machine learning that automates the design of neural network architectures. This research aims to improve the computational efficiency of NAS by employing surrogate modeling, weight sharing, and early stopping. Surrogate models are used to predict performance without fully evaluating each candidate, while weight sharing reuses parameters across models to manage memory constraints. Early stopping is applied to terminate low-performing candidates, reducing the overall computational load. We also propose a novel RL-based strategy to identify a broader range of high-performance architectures. We validated our framework on benchmark datasets, including ImageNet, CIFAR-10, and LibriSpeech, showing faster NAS processes with high-quality results. Our findings offer actionable insights for real-world NAS applications, emphasizing efficiency gains and proposing pathways for future research.