摘要
再励学习 ,作为一种新兴的智能学习模式 ,由于学习机制简单 ,不需要任何先验知识 ,也不需要样本数据 ,被越来越多地用于未知环境模型系统的学习。而目前再励学习存在的问题之一是学习速度不高 ,难以保证系统的实时性。在已有的再励学习系统中 ,再励函数多采用无模型表示结构 ,这种结构过于简单粗糙 ,也是再励学习学习效率低下的主要原因之一。因此 ,本文结合多机器人协调避障路径规划问题 ,提出一种新的基于模糊模型的再励函数结构 ,这种结构将反映机器人基本行为如躲避障碍物、其它机器人和趋向目标等的再励函数子函数进行分层建模 ,并取模糊加权和来表示总的再励函数。仿真试验表明 ,使用基于模糊模型的再励函数结构使再励学习的收敛速度要高于无模型结构。
As a newly rising intelligent learning mode, reinforcement learning is being applied more and more in a learning system with unknown environment model because of its simple learning mechanism and no need of knowledge of the system or sample data in advance. However, one of the problems of the reinforcement learning method is that its learning speed is too low to ensure the real_time system. Researchers have studied to speed up learning by improving learning algorithm and adopting intelligent exploration policy or applying the hierarchical reinforcement learning method, etc. However, how to describe the reinforcement function and how the reinforcement function affects the learning speed are seldom studied. In the existing reinforcement learning system, the model_free reinforcement function artificially defined is usually used. Its simple and rough expression is one of the causes of the low efficiency of learning. In this article, a new fuzzy model_based reinforcement function structure is presented. It is described according to the actual application in the conflict-free path planning problem of a cooperative multiple mobile robot system. In this system, the robot behaviors are divided into three basic kinds moving to the goal, avoiding obstacles and other robots. Then, the subfunctions reflecting these basic behaviors of robots are hierarchically and fuzzily modeled, and the final reinforcement function is expressed by the sum of fuzzy weighted sub-functions. The fuzzy model based reinforcement function has more accurate expression of the influence of each robot's action on the environment. The simulation shows that using the fuzzy model based reinforcement functions in reinforcement learning algorithm can further speed up the convergence than using model-free reinforcement functions.
出处
《光学精密工程》
EI
CAS
CSCD
2002年第2期148-153,共6页
Optics and Precision Engineering
关键词
机器人
再励学习
再励函数
模糊模型
避障路径规划
robots reinforcement learning
reinforcement function
fuzzy model
conflict-free path planning