Research On Parallel Estimation Of Distribution Algorithms Using Bayesian Networks

Nature is the source from which human beings receive the inspiration. It has been proved a successful method to solve the practical problems by introducing the answers provided with nature for centuries. And it has formed a embranchment of bionics. We do not need to have a uefiniie whole description for the complicated problem, but just abide by the law of nature to generate new solutions. Based on this idea, Genetic Algorithms (GAs) becomes one of the most common methods to solve such questions.In GAs, many cross, variation factors can be used. However, some arithmetic operators are often chose improperly in practice. For example, choosing one-point cross operator is unfit when building blocks and encoding do not combine precisely, which is awful without knowing the prior knowledge. So it is significant to realize the distribution of useless building blocks and the joint information of the distribution to improve algorithm's efficiency. Estimaione distribution algorithm (EDAs) is one of the intelligent algorithms developed from the above idea. Compared with GAs, EDAs selects the best individuals grouped together form father solution sets instead of just incorporating father solution sets simply.A comparison between GAs and EDAs were given first, and the limitation of GAs was described in detail. Later the core of EDAs - probability map model was introduced and Bayesian network structure (Bayesian Beliefs Networks) was emphasized. It is a chart model that can show the joint probability distribution function about a set of variables. Next a question of 6 dimension OneMax function optimization problem was solved. And we proposed a method to construct a parallel Bayesian network which realized the parallelism of EDAs at last.In contrast to traditional estimate distribution algorithm, parallel EDAs has greatly improved the efficiency when optimizing continuous functions and real time questions.