基于长短期记忆-异步优势动作评判的智能车汇入模型  

Intelligent Vehicle Merging Model Based on Long Short-Term Memory and Asynchronous Advantage Actor Critic Algorithm

作　　者：吴思凡 杜煜[1] 徐世杰 杨硕[1] 杜晨 Wu Sifan;Du Yu;Xu Shijie;Yang Shuo;Du Chen(Beijing Union University,Beijing 100101;Beijing Key Laboratory of Information Service Engineering,Beijing 100101)

机构地区：[1]北京联合大学,北京100101 [2]北京市信息服务工程重点实验室,北京100101

出　　处：《汽车技术》2019年第10期42-47,共6页Automobile Technology

基　　金：国家自然科学基金项目(91420202);北京联合大学研究生资助项目

摘　　要：针对以往强化学习中智能车汇入车流算法训练时间复杂度高、收敛速度慢的问题,提出基于长短期记忆-异步优势动作评判算法的智能车汇入模型。在异步优势动作评判算法的基础上,结合长短期记忆神经网络,有效地解决训练模型时间和模型收敛的问题。试验结果表明,该算法提高了模型收敛速度与汇入成功率,同时降低了时间复杂度,适合汇入车流场景。Due to the original intensive learning of intelligent vehicle ramp merging features high time complexity and slow converging speed for training,this paper proposes an intelligent vehicle merging model based on Long Short-Term Memory(LSTM)-Asynchronous Advantage Actor-Critic(A3C)algorithm.This algorithm combines the LSTM neural network with the original A3C algorithm,which can effectively solve the problem of high time complexity and non-convergence of the model.Experiments show that the LSTM-A3C algorithm improves the convergence speed and the convergence success rate,and reduces the time complexity,which is more suitable for the application scenario of intelligent vehicle ramp merging.

关 键 词：智能车 汇入车流 异步优势动作评判算法 长短期记忆神经网络 时间复杂度 收敛速度 

分 类 号：U461.99[机械工程—车辆工程]