基于神经网络的连续状态空间Q学习已应用在机器人导航领域.针对神经网络易陷入局部极小,提出了将支持向量机与Q学习相结合的移动机器人导航方法.首先以研制的CASIA-I移动机器人和它的工作环境为实验平台,确定出Q学习的回报函数;然后利用支持向量机对Q学习的状态——动作对的Q值进行在线估计,同时,为了提高估计速度,引入滚动时间窗机制;最后对所提方法进行了实验,实验结果表明所提方法能够使机器人无碰撞的到达目的地.%Continuous Q-leaming algorithm based on neural has been used in robotic navigation domain for its simplicity and well-developed theory.Aiming at the neural easily falling into local minimum,a new mobile robot navigation method using Q-learning based on a Support Vector Machine (SVM) is proposed.According to the developed mobile robot CASIA-I and its working environment,an approach is proposed,used to determine the reward/penalty function of Q-learning.A SVM is used to estimate the Q-value of state-action pair on-line,at the same time,in order to decrease the on-time learning time of SVM, a sliding time-window is introduced.Experimental results are included to show that the action policy obtained through Q-learning based on SVM can make the mobile robot reach the destination without obstacle collision.
展开▼
