CASIA OpenIR  > 毕业生  > 博士学位论文
平行学习理论及其在智能交通系统中的应用
林懿伦
2019-05-30
页数154
学位类型博士
中文摘要

随着机器学习技术的不断发展,越来越多的AI技术从理论变为了现实。计算机视觉、自然语言理解等研究方向实现了突破性进展,在智能安防、语音识别、图像识别、推荐算法等应用场景中,以深度学习为代表的 AI 技术逐步落地,并带来了显著的收益。

随着AI技术的应用不断增多,理论与现实的矛盾也逐渐突出,如何构建实用的机器学习系统逐渐成为研究者的关注热点。影响人们将理论实用化的障碍主要存在于三个方面。一是实验场景与现实场景的不一致性。二是数据效率与学习速度。三是控制行为在不同场景间的不一致性。

研究者提出了平行学习理论用于解决这一问题。与传统的机器学习理论不同,这一理论将智能体和环境视为一个整体的系统,而非对立的两个系统。这一理论指出,我们需要重新思考数据、知识和行为三者之间的关系。通过构建一个与真实系统平行存在的人工系统,我们可以在平行系统的异步执行和相互作用中实现三者的有机结合。这一方法首先使用先验知识和收集到的少量数据构建人工系统,该系统中同时包含了对环境的仿真和智能体的运行规则。然后通过人工系统生成大量的合成数据用以训练智能体。最后,通过智能体在平行系统中执行时产生数据的差异,我们可以修正平行系统间的差异并实现控制的目的。

为了更详细地阐释平行学习理论的概念和应用方法,我们首先对平行学习理论进行了较为完整的阐释,对其中三个主要的组成部分:描述学习、预测学习、引导学习进行了形式化的定义。然后,我们结合具体的机器学习算法,针对该理论在智能交通系统中的具体应用对其进行了较为详细的说明。

我们首先通过在无人驾驶中的应用,对使用描述学习建立人工系统的方法进行说明。通过采集智能汽车的行驶数据,结合简化的车辆动力学模型,我们以较高的数据效率构建了无人驾驶汽车的人工系统。通过在此人工系统中进行轨迹规划、车辆性能测试等任务,我们可以大大缩小需要搜索的行动空间,从而以更为高效的方式进行以往需要耗费大量金钱和时间的实车实验。

然后,我们通过交通流预测中的应用,说明如何使用预测学习产生所需要的人工数据,并通过实验验证这些合成数据带来的效果。我们提出了一种新的预测模型,称为模式敏感网络。该模型结合了描述学习和对抗生成网络,首先构建了一个简单的人工系统,再通过对人工系统中知识的迁移,实现对交通流数据的预测。实验表明,这一方法在不损失平均性能的同时,可以有效提高复杂模型在异常模式情况下的预测精度,并显著减少了训练所需要的时间和计算资源。

最后,我们通过城市交通控制中的应用,说明如何使用引导学习实现高效的执行,并对如何同步平行系统间的执行效果进行了初步的思考。我们使用了深度强化学习的最新研究成果,并结合交通系统的领域知识,实现了对多路口交通信号灯的实时控制。仿真实验的结果表明,这一城市交通控制系统可以在合理的时间内完成对交通系统控制的学习,效果较传统方法有显著的提升,且对各种交通情况都有较好的适应性。

平行学习理论的建立依赖众多机器学习技术与应用场景相关研究的发展,本文是对这一理论的早期探索。我们相信,本文的研究将有助于推动平行学习理论的发展,并为机器学习系统在实际中落地提供帮助。

英文摘要

With the continuous development of machine learning theory, more and more AI technique has come from theory to reality. In some application scenarios, such as intelligent security and recommendation system, machine learning technique represented by deep learning has gradually deployed and brought significant benefits.

 

With the increasing deployment of AI technique, the contradiction between theory and reality has become increasingly prominent. How to construct a practical machine learning system has gradually become an essential question in the AI field. The obstacles researchers faced can be divided into three types. One is the inconsistency between the experimental scene and the real scene, the second is the lack of data efficiency and slow learning speed, and the last is the inconsistency of control behavior between different scenarios.

 

Recently, a theoretical framework called parallel learning has been proposed to solve these problems.

Unlike traditional machine learning theory, this framework treats agents and environments as a holistic system rather than two separated systems.

This theory points out that we need to rethink the relationship between data, knowledge, and behavior. By constructing an artificial system that exists in parallel with the real system, we can achieve an organic combination of the three in the asynchronous execution and interaction of parallel systems.

Using a priori knowledge and a small amount of data collected, this method first builds an artificial system which contains both simulations of the environment and operating rules of the agent.

A large amount of synthetic data is then generated by the artificial system to train the agent.

Finally, we can improve the parallel systems, and also achieve the purpose of control by eliminating the difference generated by the parallel execution of the agent in the two systems.

 

In order to explain the concept and application methods of parallel learning theory,

we first gave a complete interpretation of the parallel learning theory and formalized its three main components: descriptive learning, predictive learning, and prescriptive learning.Then, we explained its applications for several typical problems in the intelligent transportation system (ITS) field.

 

We first illustrate how to use the descriptive learning method to build an artificial system through the application in autonomous driving. By collecting the driving data of the intelligent vehicle and combining the simplified vehicle dynamics model, we built the artificial system of the intelligent vehicle with high data efficiency. By performing trajectory planning, vehicle performance testing or other tasks in this artificial system, we can significantly reduce the action space need to be searched, and thus carry out real-vehicle experiments that require a lot of money and time more efficiently.

 

Then, through the application of traffic flow prediction, we explain how to use predictive learning to generate the required synthetic data. We propose a new predictive model called  Pattern Sensitive Network. This model first creates a simple artificial system by using collected data and then predict the traffic flow by transferring the knowledge acquired from the artificial system. Experiments show that this method can effectively improve the prediction accuracy under unusual cases without losing the average performance, and significantly reduce the computational resources required for training.

 

Finally, we demonstrate how to use prescriptive learning to achieve efficient execution by the application of urban traffic control. We use the deep reinforcement learning algorithm combined with the domain knowledge of the transportation system, to build a real-time multi-intersection controller. The simulation results show that this traffic controller can complete the learning procedure in a reasonable time, and its performance is significantly improved compared with the traditional methods while remaining good adaptability to various traffic situations. We also discuss how to synchronize the execution between parallel systems briefly.

 

This article is an early attempt of the establishment of parallel learning theory.

We believe our research will help promote the development of this theory and also inspire developers to build a better practical machine learning system.

关键词平行学习 强化学习 深度学习 智能交通系统
语种中文
七大方向——子方向分类人工智能基础理论
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/25763
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
林懿伦. 平行学习理论及其在智能交通系统中的应用[D]. 中国科学院自动化研究所. 中国科学院大学,2019.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
20190530 平行学习理论及其在智能(17378KB)学位论文 限制开放CC BY-NC-SA
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[林懿伦]的文章
百度学术
百度学术中相似的文章
[林懿伦]的文章
必应学术
必应学术中相似的文章
[林懿伦]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。