| Summary: | 碩士 === 國立臺灣大學 === 電機工程學研究所 === 100 === There is a common misunderstanding about building blocks. It is mistakenly believed that in genetic algorithms, if a fitness function can be separated into independent sub-problems, each sub-problem forms a building block. In this thesis we argue that even if a fitness function can be separated into independent sub-problems, each sub-problem does not necessarily form a building block. This thesis aims at examining if sub-problems in a fitness function form building blocks directly from the fitness function without performing genetic algorithms. To do so, this paper extends the convergence time model and the gambler’s ruin model so they can be applied to a larger variety of problems. With proposed models, the number of fitness evaluations can be estimated for both of these two cases: (1) some genes are transferred together in crossover (treated as a building block); (2) the genes are transferred separately. Therefore, we can compare the number of function evaluations and detect the existence of building blocks for a large family of fitness functions without actually performing a genetic algorithm.
|
|---|
