In recent years, neural architecture search (NAS) has achieved unprecedented development because of its ability to automatically achieve high-performance neural networks in various tasks. Among these, the evolutionary neural architecture search (ENAS) has impressed the researchers due to the excellent heuristic exploration capability. However, the evolutionary algorithm-based NAS are prone to the loss of population diversity in the search process, causing that the structure of the surviving individuals is exceedingly similar, which will lead to premature convergence and fail to explore the search space comprehensively and effectively. To address this issue, we propose a novel indicator, named architecture entropy, which is used to measure the architecture diversity of population. Based on this indicator, an effective sampling strategy is proposed to select the candidate individuals with the potential to maintain the population diversity for environmental selection. In addition, an unified encoding scheme of topological structure and computing operation is designed to efficiently express the search space, and the corresponding population update strategies are suggested to promote the convergence. The experimental results on several image classification benchmark datasets CIFAR-10 and CIFAR-100 demonstrate the superiority of our proposed method over the state-of-the-art comparison ones. To further validate the effectiveness of our method in real applications, our proposed NAS method is applied in the identification of lumbar spine X-ray images for osteoporosis diagnosis, and can achieve a better performance than the commonly used methods. Our source codes are available at https://github.com/LabyrinthineLeo/AEMONAS.

近年来，神经架构搜索（NAS）因其在各种任务中自动实现高性能神经网络的能力而获得了前所未有的发展。其中，进化神经架构搜索（ENAS）以其出色的启发式探索能力给研究人员留下了深刻的印象。然而，基于进化算法的 NAS 在搜索过程中容易丢失种群多样性，导致幸存个体的结构极其相似，从而导致过早收敛，无法全面有效地探索搜索空间。为了解决这个问题，我们提出了一种新的指标，称为架构熵，用于衡量人口的架构多样性。基于该指标，提出了一种有效的抽样策略，以选择具有保持种群多样性潜力的候选个体进行环境选择。此外，设计了拓扑结构和计算操作的统一编码方案，以有效表达搜索空间，并提出相应的种群更新策略以促进收敛。在几个图像分类基准数据集 CIFAR-10 和 CIFAR-100 上的实验结果证明了我们提出的方法优于最先进的比较方法。为了进一步验证我们的方法在实际应用中的有效性，我们提出的 NAS 方法应用于腰椎 X 线图像的识别，用于骨质疏松症的诊断，并且可以达到比常用方法更好的性能。