A modified screening estimation of distribution algorithm for large-scale continuous optimization

Continuous Estimation of Distribution Algorithms (EDAs) commonly use a Gaussian distribution to control the search process. For high-dimensional optimization problems, several practical issues arise when estimating a large covariance matrix from the selected population. Recent work in continuous EDAs has aimed to address these issues. The Screening Estimation of Distribution Algorithm (sEDA) is one such algorithm which, uniquely, utilizes the objective function values obtained during the search. A sensitivity analysis technique is then used to reduce the rank of the covariance matrix, according to the estimated sensitivity of the fitness function to individual variables in the search space.; In this paper we analyze sEDA and find that it does not scale well to very high-dimensional problems because it uses a large number of additional fitness function evaluations per generation. A modified version of the algorithm, named sEDA-lite is proposed which requires no additional fitness evaluations for sensitivity analysis. Experiments on a variety of artificial and real-world representative problems evaluate the performance of the algorithm compared with sEDA and EDA-MCC, a related, recently proposed algorithm. © Springer International Publishing Switzerland 2014.