基于候客点规划的空闲出租车路线推荐算法

为空闲出租车司机推荐有效的闲逛路线在提高出租车司机工作效率、减少乘客等待时间以及缓解交通压力方面具有重要作用。现有的研究工作主要集中于为空闲司机推荐完整的驾驶路线,没有考虑到真实路网环境下某些路段的可等待因素,使得推荐的路线因载客概率较低、行驶距离较长而花费成本较高。提出一种基于候客点规划的路线推荐算法,对出租车轨迹数据进行处理,并设计路径匹配算法将每个轨迹点与真实路段一一匹配。通过统计每个路段历史接载信息,并利用一种改进的多层感知机建立可预测时序接载概率的模型,结合路段的可等待因素设计一种最小花费成本的路线推荐算法。在真实数据集上的实验结果表明,与MNP、InExperence、Random算法相比,所提算法花费成本、巡航时间以及巡航路程均明显减少。     

基于时空上下文协同过滤的出租车载客点推荐算法

针对现有出租车载客点推荐算法忽略出租车所处上下文的情况,提出了一种基于时空上下文协同过滤的出租车载客点推荐算法。该算法将载客点信息映射到空间网格,通过在出租车司机驾驶行为相似度的计算中引入时间衰减因子,得到与目标出租车司机驾驶行为最相似的邻居集合,基于地点上下文过滤从相似邻居集合中选取感兴趣程度高的载客点推荐给目标出租车。在基于福州市出租车轨迹数据的实验中,时间衰减因子为0.7时,整体推荐效果最佳,同时该算法在邻居集合的不同大小时推荐准确率均优于传统协同过滤推荐算法。结果表明该算法与传统的协同过滤算法相比有更高的推荐准确度。     

基于出租车司机经验的约束深度强化学习算法路径挖掘

利用出租车司机经验,提出约束深度强化学习算法（CDRL）在线计算不同时间段内OD间最快路线。首先描述了路段经验数据库（ERSD）的提取; 然后介绍了CDRL方法,包括可选择约束路段生成和深度Q-lear-ning算法两个阶段,在第一阶段,生成OD（起终点）间可选择约束路段,在第二阶段,设计深度Q-learning算法学习出租车司机的经验,并根据他们的出发时间计算给定OD间的最快路线;最后,在广州CBD进行了应用实验。结果表明,CDRL方法计算在旅行时间上优于最短路径（SR）方法,且与最快路径（FR）方法计算路径差别不大;此外,CDRL方法在计算效率方面明显优于FR和SR方法,因此更适合OD间最快路径的在线计算。     

出租车载客点个性化推荐算法

不同的出租车司机在寻找乘客选取载客点时会有不同倾向,利用三种推荐算法对上海出租车司机载客点选取行为进行分析,根据司机对载客点的喜好程度进行个性化推荐.首先,利用基于用户和基于项目的协同过滤的算法来对出租车司机的载客点进行推荐,利用正确率指标来验证算法,实验证实了这两种算法的可行性;之后,考虑到出租车的载客行为受到时间的影响,在上述两种算法基础上增加了时间因子;最后,利用隐含因子模型（LFM）,将出租车与载客点的共现矩阵进行分解,根据分解所得矩阵进行兴趣度的分析.实验结果证明,三种方法可有效形成推荐,且LFM算法推荐准确率较高.     

基于MapReduce的出租车停泊点智能推荐算法

现有解决打车难问题的研究工作大部分是集中式地调度出租车,且大多方法在单一服务器上运行串行算法分析海量出租车GPS数据,计算量大,会遇到计算时间和计算资源的瓶颈。为此提出一种基于MapReduce的出租车停泊点智能推荐算法,为司机或乘客推荐更容易接到乘客或打到车的地点。算法通过挖掘大量出租车GPS轨迹数据,检测出停泊点,并生成停泊点知识库。再利用推荐模型为司机或乘客推荐最佳停泊点。实验分析了北京市真实出租车GPS轨迹数据,结果表明该算法能有效为司机和乘客推荐出停泊点,且在大数据量下具有较高的效率。     

基于海量出租车轨迹数据的智能推荐系统研究

本文旨在通过对出租车历史行驶轨迹进行机器学习,分析乘客的移动模式和出租车司机揽客行为模式,研究和设计一个智能推荐系统。该系统主要有离线数据挖掘部分和在线数据资源发布部分组成,离线数据挖掘采用Oracle结合Hadoop进行,以SQL存储过程开发为主,在线数据资源发布采用Java Web编写发布程序。实验测试表明,本文设计的出租车服务智能推荐系统可为乘客推荐更容易找到空驶出租车的地点,为出租车司机推荐快速招揽到乘客的地点,并实现自适应实时路况的优化路径推荐。与传统关系数据库的方法比较,本文提出的Oralce与Hadoop结合的混合模式性能更高。     

基于出租车数据的载客热点与打车热点研究

面对城市出租车高空载率和乘客打车难问题,本文针对出租车司机端和乘客端分别进行载客热点和打车热点的分析研究,提出了一种基于DBSCAN算法的数据处理模型.利用这个模型对北京市182辆出租车的GPS轨迹数据进行处理,提高了数据精度;对于不同的受众,采用K-means算法对数据进行聚类分析,得到相关热点.实验表明,划分目标用户进行各热点的推荐不仅可以有效地为出租车司机提供高概率的载客热点,乘客打车难问题也有了一种可行的解决方法.     

大规模出租车起止点数据可视分析

在现代化城市中,出租车起止点数据是一类非常有用的交通大数据,其中蕴含着丰富的时空信息.为了挖掘潜在的出租车起止点时空模式,设计了一个出租车起止点数据可视分析系统.首先利用起止点分布的全局概览图从空间上确定需要进一步挖掘的区域;然后利用系统提供的套索或者矩形选择工具选择待分析区域,由所设计的环形像素图对该区域的起止点时空模式进行可视化编码;最后通过多可视化组件协同交互,从不同维度分析出租车起止点数据的潜在时空模式.将该系统用于杭州市出租车GPS真实数据,取得了良好的效果,既有助于交通管理部门按需调配车辆,也能帮助出租车司机获得更高收益.     

基于LightGBM的乘客候车路段推荐方法

目前城市出租车服务系统中出租车的营运效率普遍不高，一方面出租车有较高的空载率，另一方面由于乘客盲目选择打车地点而造成打车难的问题。针对上述问题，提出了一种基于LightGBM的乘客候车路段推荐方法。该方法从城市出租车的历史轨迹信息中提取影响空载的时空特征，利用LightGBM框架预测各个路段出租车的空载数量，并以此计算各个路段在未来一个时隙内成功打到空车的概率，最终根据乘客所处位置将其附近成功打车概率最高的路段推荐给乘客。仿真实验利用了西安市出租车数据对模型进行训练和验证，并将实验效果与SVR、GBDT等方法进行了对比。实验结果表明，所提方法的RMSE、MAE、MAPE等指标均优于其他方法，该模型对解决出租车与乘客人车矛盾有一定实用价值。     

基于网格化的出租车空载寻客路径推荐

出租车在城市交通中扮演着十分重要的角色。通过研究出租车空载寻客路径推荐来提高出租车载客效率,具有较大的现实意义。许多城市以"环"进行区域划分,使出租车订单OD(Origin-Destination)数据呈现出环内相似、环间不同的分布特点。基于此,对订单数据进行环形切分,结合区域面积和订单数量建模,计算出租车载客核心点。提出网格化的出租车空载寻客曼哈顿路径算法,将出租车与载客核心点之间的区域进行网格化处理,找出载客概率最大的一条曼哈顿路径推荐给空载出租车司机。实验表明,较直接聚类方法,先进行环形数据切分计算出的载客核心点分布更加均匀、合理。基于网格化方法推荐的最优曼哈顿路径载客概率不低于经典的基于最短距离的路径规划算法获得的路径。     

A method for real-time trajectory monitoring to improve taxi service using GPS big data

As taxi service is supervised by certain electronic equipment (e.g., global positioning system (GPS) equipment) and network technique (e.g., cab reservation through Uber in USA or DIDI in China), taxi business is a typical electronic commerce mode. For a long time, taxi service is facing a typical challenge, that is, passengers may be detoured and overcharged by some unethical taxi drivers, especially when traveling in unfamiliar cities. As a result, it is important to detect taxi drivers’ misbehavior through taxi’s GPS big data analysis in a real-time manner for enhancing the quality of taxi services. In view of this challenge, an online anomalous trajectory detection method, named OnATrade (pronounced “on a trade,” which means activities in a taxi trade on the fly), is investigated in this paper for improving taxi service using GPS big data. The method mainly consists of two steps: route recommendation and online detection. In the first step, route candidates are generated by using a route recommendation algorithm. In the second step, an online anomalous trajectory detection approach is presented to find taxis that have driving anomalies. Experiments evaluate the validity of our method on large-scale, real-world taxi GPS trajectories. Finally, several value-added applications benefiting from big data analysis over taxi’s GPS data sets are discussed for potential commercial applications.     

Application of cellular automata and type-2 fuzzy logic to dynamic vehicle path planning

Nowadays, most road navigation systems’ planning of optimal routes is conducted by the On Board Unit (OBU). If drivers want to obtain information about the real-time road conditions, a Traffic Message Channel (TMC) module is also needed. However, this module can only provide the current road conditions, as opposed to actually planning appropriate routes for users. In this work, the concept of cellular automata is used to collect real-time road conditions and derive the appropriate paths for users. Notably, type-2 fuzzy logic is adopted for path analysis for each cell established in the cellular automata algorithm. Besides establishing the optimal routes, our model is expected to be able to automatically meet the personal demands of all drivers, achieve load balancing between all road sections to avoid the problem of traffic jams, and allow drivers to enjoy better driving experiences. A series of simulations were conducted to compare the proposed approach with the well-known A* Search algorithm and the latest state-of-the-art path planning algorithm found in the literature. The experimental results demonstrate that the proposed approach is scalable in terms of the turnaround times for individual users. The practicality and feasibility of applying the proposed approach in the real-time environment is thus justified.     

The design of an IoT-based route optimization system: A smart product-service system (SPSS) approach

The aim of route optimization system (ROS) is to design a set of vehicle routes to fulfill transportation demands, in an attempt to minimize cost and/or other negative social and environmental impacts. ROS, established based on the fruitful studies of vehicle routing problem (VRP), has been applied in various industries and forms. During daily operations, dynamic traffic conditions, varying restriction policies, road constructions, drivers’ progressing familiarity with the routes and destinations are all common factors affecting the performance of ROS. However, most current systems are designed in a one-way and open-loop manner, i.e. these systems do not track how the planned vehicle routes are performed, which hinders the continuous improvement of the system and would lead to the failure of the system. This study proposes a smart product-service system (SPSS) approach to design an IoT-based ROS, arguing that the product (i.e. the ROS) and services (updating base data and learning users’ behaviors automatically to optimize the system) should be designed as a bundle. For this end, IoT devices are employed to acquire real-time information and feedbacks of vehicles and drivers, which are used to assess the execution of planned routes and dynamically modify the base data. Moreover, the driving records from IoT devices reveal drivers’ improving familiarity with routes and destinations, which will be considered to optimize the assignment of routes to drivers. Finally, we use a case of retailing industry to show the advantages of the proposed SPSS approach.     

面向出租车出行规律的预测式可视分析方法

针对传统交通数据可视分析方法缺乏预测分析能力的问题,提出了基于出租车出行数据的预测式可视分析方法,支持用户更有效地探索未来的交通状况.在可视分析模型中,提出了结合天气、星期几等多种非交通因素的预测模型,提高了预测的准确度;提出了基于预测数据和广义地点类型约束的路径规划方法,获得了更优的路径规划结果;以多种可视化手段分析和预测了出租车司机的运营状况,帮助司机进行运营决策.以温州市出租车数据进行的实验结果表明,与传统方法相比,文中方法能更准确地预测交通状况和运营状况,并获得更合理的路径规划结果.     

一种权衡风险收益的推荐方法

日常生活中,人们面临众多需要在可选对象中进行抉择的问题。其中一些往往需要衡量风险及收益进行决策,对此已有的推荐方法依赖于用户或相似用户的历史数据,因此在类似打车地点推荐等缺乏这些数据或类似数据可重复利用度低的情况下,需要一种不依赖用户方面数据,同时能够权衡可选对象的风险及收益进行推荐的推荐方法。以经济学领域的现代投资组合理论为基础,提出一种可应用于上述场景的推荐方法。并以打车位置推荐为例说明如何使用该方法,以及同以往对于该理论的应用相比,应该如何更为适当地选择风险及收益的计算策略。在真实的数据集上进行实验,验证方法中权衡推荐策略的有效性。     

Cognition-based hierarchical en route planning for multi-agent traffic simulation

The goal of this study is to model drivers’ cognition-based en route planning behaviors in a large-scale road network via the Extended Belief-Desire-Intention (E-BDI) framework. E-BDI is a probabilistic behavior modeling framework based on agents’ own preferences of multiple attributes (e.g., travel time and its variance) and daily driving experiences. However, it is challenging to use the E-BDI framework for the demonstration of drivers’ en route planning behavior in a large-scale road network due to its high computational demand. To handle the computation issue, a hierarchical en route planning approach is proposed in this study. The proposed E-BDI-based en route planning approach consists of three major procedures: (1) network partitioning, (2) network aggregation, and (3) E-BDI-based en route planning. The Java-based E-BDI module integrated with DynusT^® traffic simulation software is developed to demonstrate the proposed en route planning approach in Phoenix, Arizona road network involving 11,546 nodes and 24,866 links. The demonstration results reveal that the proposed approach is computationally efficient and effective in representing various en route planning behaviors of drivers in a large-scale road network.