基于梯度提升决策树的汇合交互作用研究

Study of merging interactions based on gradient boosting decision tree

为了研究高速公路目标车道领跟车在交织区与汇合车辆的交互作用,基于梯度提升决策树(GBDT)建立交织区汇合交互作用模型.引入目标车道领跟车与前车、后车和汇合车辆的速度差、时间间隙、冲突评价指标及横向位置,分析汇合车辆与领跟车之间的交互行为.利用美国NGSIM数据集中目标车道领跟车与汇合车辆的轨迹数据对模型进行训练和测试,比较不同损失函数对模型的拟合效果,对汇合加速度进行偏效应分析.研究结果表明,基于平方损失函数(LS)的GBDT模型精度高于基于最小绝对偏差(LAD)和胡贝尔(Huber-M)损失函数的模型.在汇合行为的各研究对象中,汇合车辆的预测精度高于领跟车,汇合车辆的横向位置在汇合交互作用中的影响程度最高.GBDT模型用于汇合交互行为不仅可以准确预测目标车道领跟车与汇合车辆之间的交互作用,也能够获取影响变量与加速度之间隐藏的非线性关系.

The gradient boosting decision tree(GBDT)was applied to establish the interaction model of the merging behavior in the weaving area in order to analyze the interactions between merging vehicles and their putative leading and following vehicles in the target lane during merging process.Influencing variables including the speed difference,time gaps,the traffic conflicts and horizontal position among the putative leading and following vehicles in the target lane,leading and following vehicles in the original lane and the merging vehicles were extracted to analyze the merging interactions between merging vehicles and putative leading and following vehicles in the target lane.Vehicle trajectory data of NGSIM dataset were used to train and test the proposed model.Different loss functions were compared based on the fitting accuracy,and the partial effects of influencing variables on the merging acceleration were analyzed.Results show that the least squares(LS)loss function can produce the higher accuracy compared with least absolute deviation(LAD)and Huber-M loss functions.The prediction accuracy of merging vehicle is the highest among all the vehicles.The lateral position of the merging vehicle plays the most important role in the merging interaction.The GBDT model can accurately predict the interactions during merging process and deeply mine the hidden nonlinear relationships between the influencing variables and vehicle accelerations.