Enhancing the Performance of Maximum-Likelihood Gaussian EDAs Using Anticipated Mean Shift

Estimation-of-Distribution Algorithms (EDAs) are a specific type of Evolutionary Algorithm (EA). EDAs are characterized by the way in which new solutions are generated. The information in all selected solutions is combined at once. To this end, an interim representation that compresses and summarizes this information is used: a probability distribution over the solution space. New solutions are generated by sampling. Efficient optimization is guaranteed under suitable conditions. In practice it is however impossible to meet these conditions in general because arbitrarily complex distributions are required. Hence, practical techniques are required. In this paper, we focus on optimization of numerical functions using continuous distributions. The use of the normal distribution or combinations thereof is the most commonly adopted choice. It has already been so since the first EDAs in continuous spaces were introduced. An important question is how efficient EDAs are in the continuous domain using such practical distributions.