2000 年 36 巻 12 号 p. 1162-1171
UNDX (unimodal normal distribution crossover) proposed by Ono et al. shows an excellent performance on multiomodal or highly epistatic function optimization problems. But the performance of UNDX is not good on ill-scaled functions. Kita et al. proposed guidelines for designing crossover operators. UNDX-m and SPX are crossover operators based on Kita's guidelines. They show good performances on highly epistatic or ill-scaled functions. But UNDX cannot follow the guidelines. In this paper, we propose a new crossover operator named ENDX (extended normal distribution crossover). ENDX is an extension of UNDX so that ENDX can follow Kita's guidelines. Experimental result shows that ENDX shows a good performance on multimodal, highly epistatic and ill-scaled functions. Finally, we show whether these guidelines are proper for designing crossovers. For ENDX, the trade-off between the variety of children and the ratio of characteristics inherited from the parents to their children is necessary.