AutoML-Driven Architecture Optimization for Efficient and Scalable Neural Network Design
Abstract
AutoML-Driven Architecture Optimization is an emerging approach that automates the design of neural networks to improve efficiency, accuracy, and scalability across a wide range of applications. By reducing manual intervention, it enables rapid deployment of optimized models suited for different computational environments. However, traditional architecture design methods often involve extensive trial and error, are time-consuming, and fail to adapt well to varying resource constraints, leading to inefficient and inflexible models. To overcome these limitations, this paper introduces a framework based on Reinforcement Learning-based Neural Architecture Search (RL-NAS), which explores and identifies optimal architectures through a reward-driven search process. The proposed RL-NAS method is applied to optimize neural networks for image classification tasks across both edge and cloud platforms, striking a balance between model accuracy, latency, and resource usage. The experimental evaluation shows that RL-NAS generates architectures that significantly outperform manually designed baselines in terms of performance and computational efficiency. Furthermore, the method demonstrates robust adaptability to diverse deployment scenarios, confirming its effectiveness in scalable neural network design.