基于强化学习的进化神经网络及其在机器人导航中的应用  

Evolutionary artificial neural networks based on reinforcement learning and its application in robot navigation

作　　者：李佳鹤[1] 姚明海[1] 

机构地区：[1]浙江工业大学信息工程学院,浙江杭州310032

出　　处：《浙江工业大学学报》2010年第6期655-660,共6页Journal of Zhejiang University of Technology

基　　金：浙江省自然科学基金资助项目(20080376)

摘　　要：神经网络的拓扑结构对网络的有效性起着十分重要的作用,网络建模中的主要困难就是如何有效地控制网络模型的结构进化趋势和复杂度.提出一种基于强化学习的进化神经网络(RL-EANN),采用强化学习方法对网络进化群体与外界环境交互的效果进行评价,使其无需任何先验知识即可进行学习进化,通过强化信号的约束来控制网络群体的拓扑结构进化趋势.并在移动机器人避障导航仿真实验中证明,采用RL-EANN能使仿真机器人在不确定环境中快速有效地学习避障和接近目标,取得较好的导航效果,实践证明该方法的合理性和有效性.The topology structure of neural network is very important for the effectiveness of neural network. The main difficulty in network modeling is how to control the topology evolution trend and complexity in network model efficiently. An evolutionary artificial neural network （RL-EANN） based on reinforcement learning is proposed in this paper. The performance of interaction between the network evolution group and its environment is evaluated by reinforcement learning. The reinforcement learning can be evolved without any priori knowledge, and the restriction of reinforcement signal is used to control the topology evolution trend of network group. It has been proved in the experiment of mobile robot obstacle avoidance navigation. The robot can efficiently learn obstacle avoidance and reaching target rapidly in uncertain environments with RL-EANN. The experiment achieves a good navigation performance. Experimental result proves that this method is rational and effective.

关 键 词：进化神经网络 强化学习 不确定环境 避障导航 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]