摘要
提出一种多维函数的通用进化逼近方法 .通过构造一类结合采样函数和样条函数优点的基本函数族 ,提出一种单调函数逼近方法 ,并借助采样函数的有界变差特点 ,将该方法推广到一般函数情形 ,这两种函数的逼近都可通过遗传算法完成 .该方法的优点在于可以简单一致地推广到更高维函数的逼近 ,并使逼近复杂度与维数成线性关系 ,降低学习算法难度 .试验表明 ,该方法是有效的 .基于文中单调函数逼近技术提出的一种新的决策策略学习方法已成功地应用于某移动机器人控制器设计中 .
An evolutionary approximation method for multi dimensional functions is proposed in this paper. Firstly, combining sampling functions with spline functions, a class of base functions are constructed. Based on these base functions, an approximation model for monotone functions is given. Then, according to the limited variation property of sampling functions, the new model is extended to general function approximation. These approximation tasks are implemented by Genetic Algorithms. The newly proposed method has the advantages of simplicity, uniformity for multi dimensional function approximation, and the approximation complexity is linear with the dimensionality, which will reduce the computational cost considerably. Based on the approximation method, a novel decision making strategy learning model (Evolutionary Decision Making) has been proposed and successfully applied to the controller design task of mobile robots.
出处
《计算机学报》
EI
CSCD
北大核心
2000年第6期593-601,共9页
Chinese Journal of Computers
基金
国家自然科学基金!( 69783 0 0 7
6990 3 0 10 )
湖南省自然科学基金!( 96-0 6)
武汉大学软件工程国家重点实验室开放课题基金!( SEL
关键词
函数逼近
多维函数
进化逼近
移动机器人
function approximation, genetic algorithms, decision making strategy learning, evolutionary robotics
