Incorporation of knowledge in genetic recombination
Harpal Maini · Syracuse University eBooks · 1994
Genetic algorithms (GA's) evolve sets of solutions to an optimization problem into better sets of solutions. GA's are robust and flexible since they use genetic recombination methods to help evolve new solutions. We have devised a general class of genetic recombination operators called Dynamic Knowledge-Based Nonuniform Crossover that enhances the performance of GA's. This class of recombination operators uses problem-specific information embodied in good solutions to bias the recombination of genetic material. These recombination techniques constitute a widely applicable problem specific technique that can be combined with other modifications to the traditional GA. To demonstrate our approach, we have drawn on optimization problems from a wide domain, including graph partitioning, soft-decision decoding of linear block codes, and the traveling-salesperson problem. Additionally, our methods adapt themselves to learn about the environment, progressively improving in performance. This means that our recombination techniques are capable of improving upon solutions to optimization problems obtained via fast heuristics.