A comparison of design algorithms for choosing the training population in genomic models

In contemporary breeding programs, typically genomic best linear unbiased prediction (gBLUP) models are employed to drive decisions on artificial selection. Experiments are performed to obtain responses on the units in the breeding program. Due to restrictions on the size of the experiment, an efficient experimental design must usually be found in order to optimize the training population. Classical exchange-type algorithms from optimal design theory can be employed for this purpose. This article suggests several variants for the gBLUP model and compares them to brute-force approaches from the genomics literature for various design criteria. Particular emphasis is placed on evaluating the computational runtime of algorithms along with their respective efficiencies over different sample sizes. We find that adapting classical algorithms from optimal design of experiments can help to decrease runtime, while maintaining efficiency.