Genetic drift in univariate marginal distribution algorithm

Like Darwinian-type genetic algorithms, there also exists genetic drift in Univariate Marginal Distribution Algorithm (UMDA). Since the universal analysis of genetic drift in UMDA is very difficult, in this paper, we just approach a certain kind of problem (WOneMax Problem). For WOneMax Problem, The individual space in UMDA can be denoted as a full binary tree, and the selecting process in UMDA can be considered as a process of cutting branch. We employ this binary tree to calculate the probability change of each variable between two adjacent generations. Comparing this change with our experimental data, we find that when the population size is limited, there exists genetic drift in UMDA. In order to avoid genetic drift, we model the probability of each variable as a signal with noise, and then use smoothing filter to eliminate genetic drift. Numerical results show this method is effective.