Robust evolution strategies

被引：18

作者：

Ohkura, K ^{[1
]}

Matsumura, Y

Ueda, K

机构：

[1] Kobe Univ, Fac Engn, Nada Ku, Kobe, Hyogo 6578501, Japan

[2] Univ Sussex, COGS, CCNR, Brighton BN1 9QH, E Sussex, England

来源：

APPLIED INTELLIGENCE | 2001年 / 15卷 / 03期

基金：

日本学术振兴会;

关键词：

evolution strategies; numerical optimization; strategy parameters; selectively neutral mutations; robustness;

D O I：

10.1023/A:1011234912985

中图分类号：

TP18 [人工智能理论];

学科分类号：

081104 ; 0812 ; 0835 ; 1405 ;

摘要：

Evolution Strategies (ES) are an approach to numerical optimization that shows good optimization performance. However, it is found through our computer simulations that the performance changes with the lower bound of strategy parameters, although it has been overlooked in the ES community. We demonstrate that a population cannot practically move to other better points, because strategy parameters attain minute values at an early stage, when too small a lower bound is adopted. This difficulty is called the lower bound problem in this paper. In order to improve the "self-adaptive" property of strategy parameters, a new extended ES called RES is proposed. RES has redundant neutral strategy parameters and adopts new mutation mechanisms in order to utilize selectively neutral mutations so as to improve the adaptability of strategy parameters. Computer simulations of the proposed approach are conducted using several test functions.

引用

页码：153 / 169

页数：17