一种基于分类回归树的无人车汇流决策方法

决策规划是无人驾驶技术中的重要环节.由于道路结构变化或障碍物引起的车辆被动换道多采用基于逻辑规则或优化算法的决策方式.本文以通行量为优化目标，提出一种基于分类回归树（Classification and regression tree，CART）的汇流决策方法.依据交通流参数，选择大量具有代表性的车辆汇流场景.对场景中车辆的汇流决策序列进行编码，采用遗传算法搜索使得通行量最大的决策方案.将寻优获得的大量汇流决策序列作为样本，训练分类回归树.选取车辆自身信息及与周围车辆的关系等以描述环境特征，运用分类回归树描述环境特征与决策结果的映射关系，获得一种通行量最优的汇流决策方法.在软件中进行仿真实验，对比既有方法，基于分类回归树的汇流方法能够有效减少汇流行为对车流的扰动，在大流量情形下依旧能保持较高的通行效率.此外，该方法对实际实施中可能存在的环境感知误差，如定位误差，有一定的鲁棒性.

Classification and Regression Tree Based Traffic Merging for Method Self-driving Vehicles

Decision-making and planning are important technologies of unmanned vehicle. Logical rule and optimization algorithm are commonly applied to passive merging strategy for road structure change or obstacles. A traffic merging strategy aiming to improve throughput is proposed in this paper. According to different traffic parameters, a large number of typical traffic merging scenarios are selected. For vehicles in different scenarios, decision sequences are encoded and optimal merging decision is obtained by genetic algorithm based on remainder stochastic sampling with replacement (RSSR). Those optimal decisions are used to train classification and regression tree (CART). Specifically, the environmental feature is described by vehicle state and relationship between other vehicles around. Then the relationship between environmental features and decision is modeled by classification and regression tree. Compared with the previous merging strategy it is shown by simulation that the merging strategy based on CART can effectively mitigate disturbance on traffic flow, brought by merging maneuver, and maintain a high through efficiency even in large flow circumstances. Moreover, this method is also rather robust to environmental perception errors, such as positioning error which may exist in implementation.