基于最先策略增强学习的ART2神经网络

Foremost-Policy Reinforcement Learning Based ART2 Neural Network

提出一种基于最先策略增强学习的 ART2神经网络 FPRL-ART2(Foremost-Policy Reinforcement Learn-ing based ART2 neuraI network),并介绍其学习算法.为了达到在线学习的目的.在 FPRL-ART2中,从状态到行为值之间的映射中,选择第一个得到奖励的行为,而不是选择诸如1-step Q-Learning 中具有最优行为值的行为.ART2神经网络用于存储分类模式,其权重通过增强学习增强或减弱,达到学习的目的.并将 FPRL-ART2运用到移动机器人避碰撞问题的研究中.仿真实验表明,引入 FPRL-ART2后减少移动机器人与障碍物发生碰撞的次数,具有良好的避碰效果.

A foremost-policy reinforcement learning based ART2 neural network （FPRL-ART2） and its learning algorithm are proposed in this paper. To fit the requirement of real time learning, the first awarded behavior based on present states is selected in our Foremost-Policy Reinforcement Learning （FPRL） in stead of the optimal behavior in 1 step Q-Learning. The algorithm of FPRL is given and it is integrated with ART2 neural network. The stored weights of classified pattern in ART2 is increased or decreased by reinforcement learning. The FPRL-ART2 is successfully used in collision avoidance of mobile robot and the simulation experiment indicates that the times of collision between robot and obstacle is effectively decreased. The FPRL-ART2 makes favorable result of collision avoidance.