摘要
传统的Q学习已被有效地应用于处理RoboCup中传球策略问题,但是它仅能简单地离散化连续的状态、动作空间。文章提出一种改进的Q学习算法,提出将神经网络应用于Q学习,系统只需学习部分状态—动作的Q值,即可进行Q学习,有效的提高收敛的速度。最后在RoboCup环境中验证这个算法,对传球成功率有所提高。
Q-learning has traditionally been used effectively in dealing with RoboCup ball tactics,but it is only a simple discretization of continuous state and action space.Proposed a modified Q learning algorithm,neural network applied to Q learning,the system only need to learn some of the state-action Q value,you can get a continuous approximation of Q value,and can effectively improve generalization ability.Finally,in the RoboCup environment,the algorithm is proved to achieve optimal playing strategy,and effectively improves the success rate of passing ball.
出处
《四川理工学院学报（自然科学版）》
CAS
2011年第4期417-421,共5页
Journal of Sichuan University of Science & Engineering(Natural Science Edition)
