基于强化学习的多定位组件自动选择方法  

An automatic switching method for multiple location components based on reinforcement learning

作　　者：梁爽[1] 曹其新[1] 王雯珊[1] 邹风山[2] 

机构地区：[1]上海交通大学机器人研究所,上海200240 [2]新松机器人有限公司中央研究院,辽宁沈阳110000

出　　处：《智能系统学报》2016年第2期149-154,共6页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61273331)

摘　　要：在一个大规模的动态环境中,针对机器人各种定位传感器的局限性,提出了一种基于强化学习的定位组件自动选择方法。系统采用分布式架构,将机器人不同的定位传感器与定位方法封装为不同的组件。采用强化学习的方法,寻找最优策略,实现多定位组件的实时切换。仿真结果表明,该方法可以解决大型环境中,单一定位方法不能适用于整个环境的问题,能够依靠多定位组件提供可靠的机器人定位信息;环境发生改变时,通过学习的方法不需要重新配置组件,且与直接遍历组件后切换组件的方法相比,极大地减小了延时。To address the limitations of location sensors in large-scale dynamic environments, an automatic switching method for multiple robotic components based on reinforcement learning is proposed, This system uses distributed architecture and encapsulates different location sensors and methods into different middleware components. Reinforcement learning is employed to find the optimal strategy for deciding how to switch between components in real time. The simulation result shows that this method can solve problems that a single location method cannot in a large-scale environment and can provide reliable location information depending on multiple location components. This method can also effectively reduce the time delay compared with a method that first traverses all the compo- nents directly and then switches components.

关 键 词：移动机器人 定位 强化学习 中间件 MONTE CARLO方法 多传感器 模块化 分布式系统 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]