摘要
神经网络架构搜索（NAS）作为自动化机器学习的重要分支，旨在通过算法自动设计最优神经网络结构，以克服传统手工设计的局限性。本文针对现有NAS算法存在的计算成本高、搜索空间庞大及搜索效率低等问题，提出一种基于强化学习与元学习相结合的改进型NAS算法。该方法首先构建一个高效的搜索空间，利用元学习技术对初始架构进行预训练，从而加速搜索过程；其次引入多任务学习机制，在搜索过程中同时优化多个相关任务，提升模型泛化能力；再次采用渐进式搜索策略，从简单到复杂逐步构建网络架构，有效降低搜索难度。实验结果表明，所提算法在CIFAR - 10和ImageNet数据集上均取得了优于基准NAS算法的性能，不仅大幅减少了搜索时间，而且生成的网络架构在准确率方面也有显著提高。此外，该算法还具有良好的可扩展性，能够适应不同规模的数据集和应用场景，为神经网络架构的自动化设计提供了新的思路与方法，推动了NAS技术向更高效、实用的方向发展。
关键词：神经网络架构搜索；强化学习与元学习；多任务学习机制；渐进式搜索策略；高效搜索空间构建
Abstract
Neural Architecture Search (NAS), as a crucial branch of automated machine learning, aims to automatically design optimal neural network structures through algorithms, thereby overcoming the limitations of traditional manual design. This paper addresses the issues of high computational cost, vast search space, and low search efficiency in existing NAS algorithms by proposing an improved NAS algorithm that combines reinforcement learning with me ta-learning. The method first constructs an efficient search space and employs me ta-learning techniques to pre-train the initial architecture, accelerating the search process. It then introduces a multi-task learning mechanism to optimize multiple related tasks simultaneously during the search, enhancing model generalization. Furthermore, a progressive search strategy is adopted, gradually building network architectures from simple to complex, effectively reducing search difficulty. Experimental results demonstrate that the proposed algorithm outperforms benchmark NAS algorithms on both CIFAR-10 and ImageNet datasets, not only significantly reducing search time but also achieving notable improvements in accuracy. Additionally, the algorithm exhibits excellent scalability, adapting to datasets and application scenarios of varying scales, providing new approaches and methods for the automated design of neural network architectures and promoting the development of NAS technology towards more efficient and practical directions.
Keywords：Neural Network Architecture Search; Reinforcement Learning And me ta-Learning; Multi-Task Learning Mechanisms; Progressive Search Strategies; Efficient Search Space Construction
目 录
摘要 I
Abstract II
一、绪论 1
(一) 研究背景与意义 1
(二) 国内外研究现状 1
(三) 本文研究方法 2
二、NAS算法基础理论 2
(一) 神经网络架构搜索概述 2
（二） NAS算法的分类 3
（三） NAS算法的关键技术 4
三、NAS算法性能优化研究 4
(一) 搜索空间设计优化 4
(二) 搜索策略改进方法 5
(三) 评估机制的优化 6
四、NAS算法应用与展望 6
结 论 10
参考文献 11
