Abstract:

Evolutionary Algorithms （EAs） are widely adopted in Neural Architecture Search （NAS） due to their strong global exploration capability. However， traditional mutation ways often lack performance information guidance and have limited local exploitation capability， thereby constraining search efficiency and architecture quality. Therefore， this paper proposed a Diffusion-Guided Evolutionary NAS （DGEA） algorithm to enhance mutation quality and search capability jointly via distribution-aware and bias control mechanisms. The core of DGEA is to use a discrete diffusion model to generate structural perturbations， aiding evolutionary search process. By learning structural priors from the distribution of high-performance architectures， DGEA performed controlled noise addition and conditional denoising to create new high-performance architecture. At the same time， DGEA introduced an adaptive mutation strategy to balance exploration and exploitation dynamically. Besides， DGEA integrated a lightweight Harmony Search （HS） operator as local enhancement module for fine-tuning neighborhoods， thereby suppressing premature convergence. Experimental results on NAS-Bench-101， NAS-Bench-201， and ImageNet100 transfer tasks show that DGEA boosts search efficiency and effectiveness significantly. Under similar computational budgets， compared to EG-NAS （Neural Architecture Search with Fast Evolutionary Exploration）， GEA （Guided Evolutionary Neural Architecture Search with Efficient Performance Estimation） and REA （Regularized Evolution for Image Classifier Architecture Search）， DGEA reduces the architecture evaluations needed to reach the same accuracy level by 10%-84% with the overall accuracy gains of 0.28 to 0.42 percentage points. Ablation studies verify the contributions and synergistic effect of Diffusion-Guided Mutation （DGM） mechanism and HS module. It can be seen that by integrating distribution-aware generative mutation with adaptive HS local optimization， DGEA improves evolutionary search’s dynamic performance and convergence stability.

Key words: Neural Architecture Search (NAS), Evolutionary Algorithm (EA), Diffusion-Guided Mutation (DGM), distribution modeling, Harmony Search (HS), local optimization

摘要：

进化算法（EA）在神经架构搜索（NAS）中因具备较强全局探索能力而被广泛采用。然而，传统变异方式常缺乏性能信息引导，且局部开发能力有限，导致搜索效率与架构质量提升受限。因此，提出一种扩散引导的进化NAS （DGEA）算法，利用分布感知与偏置控制机制共同提升变异质量和搜索能力。该算法的核心是利用离散扩散模型生成结构化扰动，通过从高性能架构分布中学习到的结构先验，执行可控的噪声添加与条件去噪以生成新的高性能架构。同时，引入自适应变异策略动态平衡探索与利用；此外，集成一个轻量级的和声搜索（HS）算子作为局部增强模块微调架构邻域，以抑制早熟收敛。在NAS-Bench-101、NAS-Bench-201以及ImageNet100迁移任务上的实验结果表明，DGEA显著提升了搜索效率与效果。在相近的计算预算下，相较于EG-NAS （Neural Architecture Search with Fast Evolutionary Exploration）、GEA （Guided Evolutionary NAS with Efficient Performance Estimation）和REA （Regularized Evolution for Image Classifier Architecture Search），DGEA达到相同准确率水平所需架构评估数量减少10%~84%，准确率整体提升0.28~0.42个百分点。消融实验证实了扩散引导变异（DGM）机制和HS模块的贡献以及协同作用。通过整合分布感知的生成式变异与自适应HS局部优化，DGEA有效提升了进化搜索的动态性能与收敛稳定性。

关键词: 神经架构搜索, 进化算法, 扩散引导变异, 分布建模, 和声搜索, 局部优化