ABSTRACT Neural Architecture Search (NAS) is an emerging subfield of automated machine learning (AutoML) that seeks to automate the design of deep neural networks. Traditional manual and heuristic-based architecture design is labor-intensive, time-consuming, and dependent on expert knowledge. NAS algorithms aim to minimize human involvement by automatically discovering architectures that achieve state-of-the-art performance across diverse tasks. This survey presents a comprehensive review of NAS methodologies, including reinforcement learning-based, evolutionary algorithm-based, gradient-based, and one-shot approaches. We examine the key components of NAS frameworks—search space formulation, search strategy design, and performance estimation methods—along with their computational costs, benchmark datasets, and evaluation protocols. Furthermore, we analyze scalability challenges, generalization capabilities, and deployment considerations, particularly for resource-constrained environments. Finally, we outline open research problems and promising future directions, with emphasis on hardware-aware, explainable, and zero-shot NAS paradigms. Keywords: Neural Architecture Search, AutoML, Deep Learning, Meta-Learning, Reinforcement Learning, Search Space, One-Shot NAS, NASBench
Atlas et al. (Mon,) studied this question.