交通场景下基于深度强化学习的感知型路径分配算法

Perceptual path allocation algorithm based on deep reinforcement learning in traffic scene

路径分配问题是交通数字孪生系统的重要研究方向之一,其重点是综合考量行驶需求的动态变化以及路网信息的实时改变,实现高效合理的路径规划。现阶段一些经典的分配算法如粒子群、Dijkstra等算法及其优化模型仅能达到全局静态最优,忽略了现实交通中的复杂变化。而逐渐推出的各种深度学习算法虽能进行全面的时空预测,但受限于海量历史数据的归纳分析以及较高的运算成本,难以大规模应用。鉴于此,提出了一种静态分配算法与深度强化学习算法结合的感知型路径分配算法,在行驶中依据实时路网信息和车辆当前状态,实现全局路径动态再分配及更新,相关算法的精度和效率在仿真实验中得到验证。

Path allocation problem is one of the important research directions of traffic digital twin system.Its focus is to comprehensively consider the dynamic changes of driving demand and the real-time changes of road network information,so as to realize efficient and reasonable path planning.At present,some classical allocation algorithms such as particle swarm optimization,Dijkstra and their optimization models can only achieve global static optimization,ignoring the complex changes in real traffic.Although various depth learning algorithms gradually introduced can carry out comprehensive spatio-temporal prediction,they are difficult to be applied on a large scale due to the inductive analysis of massive historical data and high operation cost.In view of this,this paper proposes a perceptual path allocation algorithm based on the combination of static allocation algorithm and deep reinforcement learning algorithm.During driving,the global path is dynamically redistributed and updated according to the real-time road network information and the current state of vehicles.The accuracy and efficiency of the proposed algorithm are verified in the simulation experiment.