交互门控循环单元及其在到达时间估计中的应用  

Interactive gated recurrent unit and its application for estimated time of arrival

作　　者：孙翊文 王宇璐 傅昆 王征 张长水[1] 周东华[3,1] 叶杰平 Yiwen SUN;Yulu WANG;Kun FU;Zheng WANG;Changshui ZHANG;Donghua ZHOU;Jieping YE(Department of Automation,Tsinghua University,Beijing 100084,China;DiDi AI Labs,Beijing 100094,China;College of Electrical Engineering and Automation,Shandong University of Science and Technology,Qingdao 266590,China)

机构地区：[1]清华大学自动化系,北京100084 [2]滴滴出行人工智能实验室,北京100094 [3]山东科技大学电气工程与自动化学院,青岛266590

出　　处：《中国科学：信息科学》2021年第5期822-833,共12页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:61751307,61876095)资助项目。

摘　　要：门控循环单元(gated recurrent unit, GRU)是一种有代表性的深度神经网络,它在众多序列学习任务中达到了国际领先的水平.然而,在门控循环单元的每个时间步之间,输入信息与隐含状态信息缺乏交互,这对更好地挖掘上下文语义信息带来了挑战.针对这个问题,本文提出了一个新颖的序列学习通用的语义特征提取模型:交互门控循环单元(interactive gated recurrent unit, InterGRU),可以让输入与隐含状态向量在各时间步间进行多轮充分的交互.并且,在到达时间估计(estimated time of arrival, ETA)这个有代表性、有挑战的时空序列预测任务上,本文提出了一套基于交互门控循环单元的深度学习框架(InterGRU-ETA).本文在来自滴滴出行平台真实场景下的海量数据集上充分地实验验证了InterGRU-ETA.结果表明,我们的框架在预测准确率上优于目前国际上最先进的方法.这反映了交互门控循环单元在捕获序列语义信息上的性能优势和广阔前景.A gated recurrent unit(GRU) is a representative deep neural network that has achieved promising results in many sequence learning tasks. However, there is a lack of interaction between the input and the hiddenstate among each time step of GRU, resulting in the challenges to mine contextual semantic information effectively.In this paper, we propose a novel deep learning method called interactive gated recurrent unit(InterGRU) to solve this problem, which allows full interaction between the input and the hidden state at various time steps.Furthermore, we propose a deep learning framework, InterGRU-ETA, based on InterGRU for the estimated time of arrival(ETA) which is a representative and challenging time series forecasting task. Our framework has been fully experimentally verified on the large-scale real-world datasets from the Didi Chuxing platform. The results on massive historical vehicle travel data show that InterGRU-ETA is superior to other state-of-the-art algorithms.This can reflect the advantages of InterGRU in capturing sequential semantic information.

关 键 词：门控循环单元 到达时间估计 深度学习 时空序列预测 智能交通系统 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]