Skinner操作条件反射的一种仿生学习算法与机器人控制  

A Bionic Learning Algorithm Based on Skinner's Operant Conditioning and Control of Robot

作　　者：任红格[1] 阮晓钢[1] 

机构地区：[1]北京工业大学电子信息与控制工程学院,北京100124

出　　处：《机器人》2010年第1期132-137,共6页Robot

基　　金：国家863计划资助项目(2007AA04Z226);国家自然科学基金资助项目(60774077);北京市教委重点资助项目(KZ200810005002);北京市人才强教计划项目;高等学校博士学科点专项科研基金资助课题

摘　　要：针对两轮自平衡机器人的运动平衡控制问题,提出了基于Skinner操作条件反射理论的BP神经网络与资格迹相结合的仿生自主学习算法作为两轮机器人的学习机制.该算法利用资格迹能解决延迟影响、加快学习速度和提高可靠性的特点,将其与BP神经网络相结合构成复合学习算法,能够预测机器人将要获得的行为评价函数,并依据概率取向机制以一定的概率选择最大评价值对应的最优行为,从而使机器人能够在未知环境下通过与环境的交互、学习和训练,获得像人或动物一样的自主学习技能,实现对两轮机器人的运动平衡控制.最后,分别用基于Skinner操作条件反射理论的BP算法和BP资格迹复合算法对两轮机器人做了仿真实验并进行了比较.结果表明,基于Skinner操作条件反射理论的BP资格迹复合仿生自主学习算法的学习机制能够使机器人获得良好的动态性能和较快的学习速度,体现了机器人较强的自主学习技能和平衡控制能力.Aiming at the movement balance control problem of the two-wheeled self-balancing mobile robot, a bionic self-learning algorithm consisting of BP （backpropagation） neural network and eligibility traces based on Skinner＇s operant conditioning theory is put forward as a learning mechanism of the two-wheeled robot. The algorithm utilizes the characters of eligibility traces in resolving delay effect, increasing learning speed, and improving reliability and ability, so that the complex learning algorithm consisting of BP neural network and eligibility traces can predict the behavior evaluation function that the robot would obtain, and choose the optimum action corresponding to the biggest evaluation value according to the probability tendency mechanism by a certain probability. Thereby the two-wheeled robot can obtain the self-learning skills like a human or animal by interacting with, studying and training the unknown environment, and realize the movement balance control of the two-wheeled robot. Finally, two simulation experiments are done and compared using the BP algorithm and the complex learning algorithm consisting of BP neural network and eligibility traces based on Skinner＇s operant conditioning theory. The simulation results show that the learning mechanism of the complex learning algorithm consisting of BP neural network and eligibility traces based on Skinner＇s operant conditioning theory makes the robot obtain the better dynamic performance and the quicker learning speed, and reflect stronger self-learning skills and balance control abilities.

关 键 词：Skinner操作条件反射 资格迹 自主学习 平衡控制 两轮机器人 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]