Repairing normal EDAs with selective repopulation

The standard Estimation of Distribution Algorithm (EDA), usually, suffers from premature convergence due to an inherent inability to maintain an adequate variance and to preserve diverse candidate solutions. Normal multivariate EDAs have especially shown a lack of exploration even for convex objective functions. This article introduces several techniques which can be used to enhance the standard Normal multivariate EDA performance. The most important ones are based on (1) pre-selecting the candidate solutions to be evaluated, (2) replacing only a fraction of the population and (3) computing weighted estimators of the mean and covariance matrix. The resulting Normal EDA is competitive with similar approaches, as it is evidenced by statistical comparisons.