WHY DGAS WORK WELL ON GA-HARD FUNCTIONS

What makes a problem easy or hard for a genetic algorithm (GA)? Much p revious work has studied this question by applying Walsh analysis.(4)) In this paper, we demonstrate a function that is GA-hard by analyzing the Walsh coefficients of this function's Walsh decomposition. Then, we construct five functions with differing degrees of difficulty for g enetic algorithms. Some are GA-easy and some are GA-hard. In a previou s paper,(29)) wh have proposed a novel selection method, disruptive se lection. This method devotes more trials to both better solutions and worse solutions than it does to moderate solutions, whereas the conven tional method allocates its attention according to the performance of each solution. Experimental results show that DGAs (GAs using disrupti ve selection) perform very well on both GA-easy and GA-hard functions. Finally, we discuss why DGAs outperform conventional GAs.