Pattern-based preservation of building blocks in genetic algorithms

As stated in the building block hypothesis, we expect genetic algorithms (GAs) to create building blocks (BBs) and combine them appropriately in the evolutionary process. However, such BBs are often destroyed by unwanted crossovers, soon after they are created. Also, we may suffer from a "loose" encoding of chromosomes since BBs are in general unknown. In this paper, we propose a framework named GAP (GA with patterns), in which key patterns are extracted from significantly "good" chromosomes and protect such key patterns against unwanted crossover. GAP is applicable to optimization problems with fixed-point encoding and permutation encoding in a uniform fashion, and unlike perturbation-based linkage learning methods, GAP does not require extra fitness evaluations. Experimental results with the royal road problems and traveling salesman problems show the performance improvement of GAP over standard GAs.