摘要
针对在基于行为的移动机器人沿墙导航控制器的设计中缺乏足够的先验知识的问题,采用Q-学习方法让机器人通过学习来自动构建导航控制器。将模糊神经网络和Q-学习相结合,用模糊神经网络直接逼近连续状态和动作空间中的Q值函数。利用对Q值函数的优化获得控制输出。模糊神经网络中的节点根据状态动作对的各个分量和时间差分的新颖性进行自适应地添加和构造,这样不仅能克服节点选择的困难还能使网络保持适度的规模。网络中的参数采用扩展卡尔曼滤波方法进行自适应调整。基于Khepera 2机器人的沿墙导航实验验证了该方法的有效性和优越性。
The Q-learning was introduced into navigation control of the wall-following task of mobile robots where there was no enough priori knowledge available.The Q-value function was approached directly u-sing Fuzzy Neural Network（FNN）.The optimization method was used to search the greedy action with maximum Q-value.The nodes of FNN were created incrementally and adaptively according to every ele-ment of the current pair of state-action and Temporal Difference（TD）,which overcame the difficulties of the choice of nodes and ensured an economic size of the network.Moreover the parameters of the FNN were updated using Extended Kalman Filter（EKF）.The results of the wall-following task of Khepera 2 mobile robot demonstrate the superiority and validity of the proposed method.
出处
《电机与控制学报》
EI
CSCD
北大核心
2010年第6期83-88,97,共7页
Electric Machines and Control
基金
国家自然科学基金(60703106)
关键词
Q-学习
模糊神经网络
沿墙导航
移动机器人
Q-learning
fuzzy neural network
wall-following navigation
mobile robots
