An iterated local search algorithm for the construction of large scale D-optimal experimental designs

We focus on the D-optimal design of screening experiments involving main-effects regression models, especially with large numbers of factors and observations. We propose a new selection strategy for the coordinate-exchange algorithm based on an orthogonality measure of the design. Computational experiments show that this strategy finds better designs within an execution time that is 30% shorter than other strategies. We also provide strong evidence that the use of the prediction variance as a selection strategy does not provide any added value in comparison to simpler selection strategies. Additionally, we propose a new iterated local search algorithm for the construction of D-optimal experimental designs. This new algorithm clearly outperforms the original coordinate-exchange algorithm.