Evolutionary search with soft selection (Q1112757)

A new method of global random search in \(R^ n\) devised from a model of evolution is proposed. The model can be regarded as a randomization of Eigen's model of macromolecular evolution. The novelty of the method is in a random choice of modification basepoints with probabilities depending on results of the last group of trials - soft selection. In other aspects, the proposed search and the widely used evolutionary search are identical, and the performance of the former can be conveniently assessed against the latter. The method appears especially efficient in crossing gaps between regions of attraction of local optima. This effect, crucial for global optimization, comes from an ability of soft selection to shift distributions of trials toward gaps. Numerical estimates show significant advantages of soft selection which grow with the gap size and the problem dimension.