基于进化优化的神经架构搜索方法综述

A Survey of Evolutionary Optimization-based Neural Ar-chitecture Search

探讨了基于进化优化的神经架构搜索(Neural Architecture Search,NAS)算法,该类算法通过模拟生物进化的过程,能自动发现适应于特定任务的神经网络结构.本文回顾了进化优化算法在全局搜索、适应性和对离散连续问题的灵活性等方面的优势,详细介绍了基于进化计算的NAS方法在神经网络设计中的应用,强调了其在非连续、高维搜索空间中的潜力,总结了该领域的主要进展和挑战.最后,基于进化优化的NAS算法在推动神经网络架构搜索领域取得更多突破进行展望.未来的研究方向包括提高算法的计算效率、引入智能的搜索策略、克服高维搜索空间的问题以及该算法与深度学习的结合.

This paper explores the evolutionary computation(EC)-based neural architecture search(NAS)algorithms,which can automatically discover neural network structures suitable for specific tasks by simulating the process of biological evolution.This paper reviews the advantages of EC-based NAS algorithms in global search,adaptability and flexibility for both continuous and discrete optimization problems.The application of EC in NAS,especially in consideration of its potential in searching in discrete and high-dimensional space,is introduced in detail.Furthermore,this paper summarizes the main advances and challenges in EC-based NAS algorithms.Finally,this paper looks forward to the prospect of EC-based NAS algorithms in driving more breakthroughs in the field of neural network architecture search.Future research directions include improvement of the computa-tional efficiency of the algorithms,introduction of intelligent search strategies,solution of the problem of high-dimensional search space,and the integration with deep learning.