Learning building block structure from crossover failure

In the classical binary genetic algorithm, although crossover within a building block (BB) does not always cause a decrease in fitness, any decrease in fitness results from the destruction of some building blocks, in problems where such structures are well defined, such as those considered here. Those crossovers that cause both offspring to be worse, or one to be worse and one unchanged, are here designated as failed crossovers. Counting the failure frequency of singlepoint crossovers performed at each locus reveals something of the BB structure. Guided by the failure record, GA operators could choose appropriate points for crossover, in order to work more efficiently and effectively. Experiments on test functions Royal Road R1 and R2, Holland’s Royal Road Challenge function and H-IFF functions show that such a guided operator improves performance. While many methods exist to discover building blocks, this ”quick-and-dirty” method can sketch the linkage nearly ”for free�...

• Type: Conference
• Year: 2007
• Where: GECCO
• Authors: Zhenhua Li, Erik D. Goodman
• Keywords: Binary Genetic Algorithm, Building Blocks, GECCO 2007, Genetic Algorithms