Dynamic Neural Network-Based Integrated Learning Algorithm for Driver Behavior

Vehicle driver's perception, judgment, decision and action towards traffic environment are usually uncertain and inconsistent during their driving processes. Thus, it is difficult to use the traditional driving decision model to accurately predict the driving behaviors under these circumstances. This paper proposes a DNNIA algorithm to describe driving behavior by dynamically integrating ANNs. Specifically, some ANNs are first trained to learn different kinds of driving behaviors based on sample data and small amounts of these ANNs with minimal generalization error E are then selected and integrated to predict the final driving behavior. The Lagrangian function method is used to resolve the coefficient ωi for optimal ensemble. Moreover, by introducing the idea of agent alliance, the study takes each individual ANN as an agent in the alliance and outputs the maximal value among all the weighted average outputs of the neuron in each individual ANN. The proposed method is evaluated on some benchmark datasets to show its effectiveness. In addition, the predicted driver's habitual behavior by DNNIA, such as braking pedal, consistently accord with that revealed by the sample data, which proves its practicality for real-world problems.

摘要针对传统驾驶决策模型难以体现驾驶员驾驶过程中对交通环境的感知、判断、决策、动作等环节存在不确定性和不一致性，提出了一种基于神经网络的驾驶行为动态集成学习算法——DNNIA。首先训练多个个体网络模拟驾驶行为，然后动态选择泛化误差E最小的个体网络进行集成，采用拉格朗日函数法求解最优集成权系数ωi，并引入agent联盟的思想，把联盟中的个体网络对应的神经元输出做加权平均后，取最大值作为输出。在标准数据集上验证了该算法的有效性，仿真实验中得到的驾驶员踩踏踏板的习惯行为仿真结果与实际采集的样本数据总体趋势基本吻合。