Feature selection provides a great enhancement in the process of building a classifier model. A recent approach to feature selection is the use of Estimation of Distribution Algorithms (EDAs). Those algorithms?s performance is greatly affected by the initial population, so prior knowledge about the problem is very important. The most important prior knowledge about the features is the relative order of importance observed among them, which can be obtained by some statistical measure. Based on the use of that kind of knowledge, some improvements are proposed and theoretically discussed. An experiment is presented, which evaluates potential benefits of those alternatives.