ROBOT NAVIGATION IN UNKNOWN ENVIRONMENT BASED ON DYNAMIC STRUCTURE SELF-ORGANIZING NEURAL NETWORK
Artificial neural network was an attractive approach applied in reinforcement learning due to its ability to approximate any complex functions and avoid dimension explosion. But usually it’s a complicated and insipid work to design the neural network structure and refine the network parameters. In this paper a neural network called dynamic structure self-organizing neural network (DSONN) is proposed. The structure of this network can be changed dynamically, according to the complexity between the input and target training patterns. The proposed network can not only insert and prune nodes in every hidden layer but also produce new hidden layers, so the structure of network can be changed dynamically and self-organized. In order to finish the mobile robot navigation mission, Q-learning is used to learn the control policy in “on-line fashion and the network approximate the value function in reinforcement learning. The paper illustrates the approach using a Pioneer3-DX robot started from a certain point and arrives at the goal point avoiding obstacles. Experiments have been done on function approximation and Q-learning of mobile robot navigation, the results showing this approach is effective.
mobile robot dynamic structure self-organizing neural network reinforcement learning golden divisional method
Junfei Qiao Ruiyuan Fan Honggui Han
College of Electronic Information and Control Engineering,Beijing University of Technology,Beijing 100124,China
国际会议
2008年拟人系统国际会议(2008 International Conference on Humanized Systems )(ICHS’08)
北京
英文
2008-10-18（万方平台首次上网日期，不代表论文的发表时间）