Fitness uniform deletion: a simple way to preserve diversity
GECCO '05 Proceedings of the 7th annual conference on Genetic and evolutionary computation
Cooperative Multi-Agent Learning: The State of the Art
Autonomous Agents and Multi-Agent Systems
CIXL2: a crossover operator for evolutionary algorithms based on population features
Journal of Artificial Intelligence Research
Exploring the explorative advantage of the cooperative coevolutionary (1+1) EA
GECCO'03 Proceedings of the 2003 international conference on Genetic and evolutionary computation: PartI
| Hi-index | 0.00 |
A new family of "Distribution Replacement" operators for use in steady state genetic algorithms is presented. Distribution replacement enforces the members of the population to conform to an arbitrary statistical distribution, defined by its Cumulative Distribution Frequency, relative to the current best individual. As new superior individuals are discovered, the distribution "stretches" to accommodate the increased diversity, the exact opposite of convergence. Decoupling the maintenance of an optimal set of parents from the production of superior children allows the search to be freed from the traditional overhead of evolving a population of maximal fitness and, more significantly, avoids premature convergence. The population distribution has a significant effect on performance for a given problem, and in turn, the type of problem affects the performance of different distributions. Keeping mainly good individuals naturally does well on simple problems (as do distributions that exclude "median" individuals). With deceptive problems however, distributions which keep mainly bad individuals are shown to be superior to other replacement operators and also outperform classical generational genetic algorithms. In all cases, the uniform distribution proves suboptimal. This paper explains the details of distribution replacement, simulation experiments and discussions on the extension of this idea to a dynamic distribution.