神经网络架构搜索(Neural Architecture Search, NAS)通过设计高效的搜索方法，自动获取性能优异的网络结构，避免算法开发过程中大量的人工调优工作，降低算法开发难度，因而NAS的研究备受关注。然而，面向实际应用，适配硬件和任务的快速NAS方法的研究仍存在亟待解决的关键科学问题：1、NAS常被建模为奖励延迟马尔可夫决策过程，其搜索算法存在收敛速度慢的问题，研究奖励重分配的强化学习搜索算法，以加快搜索；2、目前NAS方法通常是针对单一任务定义随机初始的搜索空间，存在先验知识不足、难以快速适配不同任务的问题，针对多种任务构建具备数据和专家知识的搜索空间，实现快速适配任务的NAS方法；3、目前NAS方法通常没有考虑硬件适配问题，搜索的网络结构复杂，难以应用在资源受限的设备上，通过引入硬件约束代价，研究适配硬件的NAS方法。本课题将为NAS的落地应用奠定扎实的方法和理论基础。

Neural Archiitecture Search (NAS) automaticly searchs more promising architectures, which improves the study of neural network algorithm. NAS has attracted a lot of attention from scholars all over the world. However, NAS methods for edge hardwears and multiple tasks still have several problems needed to be solved: 1) NAS is usually modeled as a Markov decision process with fully delayed reward, whose search algorithm is with unstable and slow convergence. To solve this, we study reward decomposition and reward redistribution-based Reinforcement Learning (RL) search algorithm for speeding up the convergence process. 2) The search space of NAS uaually is designed only for one single task and initialized with random weights, which can't quickly adapt to different tasks. Therefore, we build a search space with data and expert knowledge for multiple tasks. 3) The current NAS methods usually do not consider the computing constraints of edge hardwares, and obtain neural networks with complex architectures and large parameters which can't apply in edge hardwares. This project aims to search neural networks for edge hardwares by introducing the computing constraints into the evaluation stratege. This project provides the theoretical foundation for the practical applicatiion of NAS methods.