基于Q学习的Agent在单路口交通控制中的应用  

Application of Agent-based Q-learning in the Traffic Flow Control of Single Intersection

作　　者：陈阳舟[1] 张辉[1] 杨玉珍[1] 胡全连[2] 

机构地区：[1]北京工业大学,北京100022 [2]江西师范大学,江西南昌330027

出　　处：《公路交通科技》2007年第5期117-120,共4页Journal of Highway and Transportation Research and Development

基　　金：北京市教委科技发展计划基金资助项目(TM2004100051);北京市自然科学基金资助项目(4042006);北京工业大学博士科研启动基金资助项目(52002011200402)

摘　　要：将Agent技术与Q学习算法相结合,应用到城市交通控制领域中,对单交叉口的交通流进行了控制研究,介绍了路口Agent的结构模型以及基于Q学习算法的学习机制的实现,提出了一种适用于交通控制的奖惩函数。即当红灯相位的饱和度大于绿灯相位的饱和度时,红灯相位的相对警界度在奖惩函数中占主导地位,此时大部分情况下会对Agent进行惩罚;在以后的决策过程中面对类似的交通状态Agent所选择的控制行为更倾向于将通行权切换给下一个相位,反之,Agent所选择的行为倾向于保持当前相位的通行权到下一决策时刻。并通过微观交通仿真软件Paramics对控制算法进行仿真研究,仿真结果表明该方法的控制效果优于定时控制,同时验证了奖惩函数的有效性。An approach of Agent technology combined with Q - learning is applied to urban traffic control, to study the single intersectio control.The model of intersection Agent an the implementation of the learning function based on Q-learning are introduced. A reward function which is fit to traffic control is put forward.The Agent will be punished when the red saturation is more than green＇s , when the relative security of red phase occupies dominant position in the reward function. In other words, in later decision-making process facing the similar traffic condition, the control behavior which Agent chooses would let the fight of way cut to the next phase. Otherwise, Agent would choose maintaining current phase right of way until next decision making. The experimental results indicate that the approach is better than the fixed control, and validate the effectiveness of the reward function.

关 键 词：交通工程 单交叉口 Q学习 奖惩函数 交通流 

分 类 号：U491[交通运输工程—交通运输规划与管理]