Neural architecture search automates the design of deep neural network, and thus saves the costly manpower in applications like wildlife identification. However, the search may consume a lot of time to find the ideal architecture due to the bottleneck like estimating architecture performance. To reduce time cost, a novel method based on reinforcement learning is proposed in this paper. The proposed method economically predicts the architecture performance through regression trees rather than expensively estimating the performance. The trees are learnt from the performance and the attributes quantified from the architecture characteristics through the conversion functions developed specifically for quantization. The collection of attributes and the performances is called model pool. In experiments, a single search only consumes 6.5 hours and the resulting architecture contains 2.49 million parameters. Both the searching time and the parameter number are the best throughout all methods in comparison. The architecture is respectively tested on the datasets CIFAR-10, D3 built on Snapshot Serengeti dataset and NACTI-64 built on North American Camera Trap Images. Test error of CIFAR-10 is 4.61%, the accuracy of D3 is 95.43% which is better than that of human expert and hand-crafted networks, and the accuracy of NACTI-64 is 96.53% which is competitive throughout all networks in comparison.

神经体系结构搜索可自动执行深度神经网络的设计，从而节省了野生生物识别等应用中昂贵的人力。但是，由于存在诸如估计体系结构性能的瓶颈，因此搜索可能会花费大量时间来找到理想的体系结构。为了减少时间成本，本文提出了一种基于强化学习的新方法。所提出的方法通过回归树经济地预测了体系结构的性能，而不是昂贵地估计性能。通过专门针对量化开发的转换函数，从性能中学习树木，并从体系结构特征中量化属性。属性和性能的集合称为模型池。在实验中，一次搜索仅消耗6。5小时，生成的体系结构包含249万个参数。在所有比较方法中，搜索时间和参数编号都是最好的。该架构分别在数据集CIFAR-10，基于Snapshot Serengeti数据集构建的D3和基于北美相机陷阱图像构建的NACTI-64上进行了测试。CIFAR-10的测试误差为4.61％，D3的精度为95.43％，优于人类专家和手工制作的网络，NACTI-64的精度为96.53％，在所有网络中都具有竞争力。D3建立在Snapshot Serengeti数据集上，而NACTI-64建立在北美相机陷阱图像上。CIFAR-10的测试误差为4.61％，D3的精度为95.43％，优于人类专家和手工制作的网络，NACTI-64的精度为96.53％，在所有网络中都具有竞争力。D3建立在Snapshot Serengeti数据集上，而NACTI-64建立在北美相机陷阱图像上。CIFAR-10的测试误差为4.61％，D3的精度为95.43％，优于人类专家和手工制作的网络，NACTI-64的精度为96.53％，在所有网络中都具有竞争力。