动态FOCPA学习系统设计及在机器人运动平衡控制中的应用

Design of Dynamic FOCPA Learning System and Its Application to Robot Motion Balance Control

针对仿生自主学习系统的自组织和泛化能力问题,基于Skinner操作条件反射原理和模糊聚类算法设计了动态FOCPA(fuzzy operant conditioning probabilistic automaton)仿生自主学习系统.动态FOCPA学习系统不仅具有仿生的自学习和自组织能力,而且提高了学习的精度和速度.其在仅能获得环境微弱反馈信息的前提下,首先采用在线聚类的方法实现对输入空间的灵活划分,以确保映射规则的数目是最经济的;然后以取向值为评价信号,采用OC学习算法,在线自主学习输入状态到输出操作行为的最佳映射,并加入一个高斯噪声项对映射结果进行实时优化.此外,动态FOCPA学习系统还利用信息熵的评价能力,来验证自身的自学习和自组织能力.理论上分析了设计的OC学习算法的收敛性;通过对两轮柔性直立式机器人姿态平衡控制和速度控制的实验分析,验证了动态FOCPA学习系统的有效性.

Aiming at the ability of self-organization and generalization of bionic autonomous learning system,this paper constructs a dynamic fuzzy operant conditioning probabilistic automaton（FOCPA） bionic autonomous learning system based on Skinner operant conditioning（OC） theory and fuzzy clustering algorithm.The dynamic FOCPA learning system not only has bionic self-learning and self-organizing ability,but also can improve the learning speed and precision of learning system. Under the learning environment where only weak feedback information can be obtained,the FOCPA learning system firstly adopts online clustering algorithm to flexibly divide the input space to ensure that the number of mapping rules is the most economical.And then the learning system takes orientation value as evaluation signal and adopts the designed OC learning algorithm to autonomously learn the optimal mapping online from input states to output operant action,and a Gaussian noise term is added for optimizing the mapping result in real time.Moreover,by using the evaluating ability of information entropy, the self-learning and self-organizing ability is verified.The convergence of OC learning algorithm is proved from theory,and the further experiments on posture balancing control and velocity control of two-wheeled flexible upright robot prove the validity of dynamic FOCPA learning system.