Neural Architecture Search (NAS) has automated the design of deep neural networks and achieved state-of-the-art performance across computer vision and natural language processing tasks. However, most NAS methods focus primarily on predictive accuracy while ignoring practical deployment constraints such as inference latency, memory footprint, and energy consumption. These limitations make conventional NAS unsuitable for edge and resource-constrained environments. This paper proposes a novel Energy-Aware Multi-Objective Neural Architecture Search (EA-MO-NAS) framework that jointly optimizes model accuracy, parameter size, inference latency, and energy consumption. The proposed method integrates reinforcement learning for guided architecture sampling with evolutionary mutation and adaptive search-space pruning to reduce computational overhead. A dynamic multi-objective fitness function balances competing optimization goals. Experimental evaluations on CIFAR-10, CIFAR-100, and Tiny-ImageNet demonstrate that the proposed method achieves competitive accuracy while reducing energy consumption by up to 45% compared to conventional NAS approaches. The framework enables sustainable and deployment-friendly automated neural network design.
Devi et al. (Sun,) studied this question.