Fig. 3

The bar plot shows the prediction ability (average of environments and traits) according to different training sets (TRS) used for genomic prediction for the HEL (Helix) dataset. The plot represents the training sets by OTS1, OTS2, OTS3, Random1, Random2, Random3 and CV2. OTSs are our optimized training sets, that increase in size (from 4.6 to 19.2% of TRS) by combining one, two, or three replications of the populations selected by the LA-GA-T algorithm. Random’s are our randomly sampled training sets with the same size of OTSs. CV2 is our benchmark scenario, where we used 70% of TRS, and a cross-validation scheme (CV2), under a multi-trait-multi environment model (MTMET). GET (genotype x environment x trait) and GWT (genotype x environmental covariates x trait) are the kernels used as the basis for the selection of information by the LA-GA-T algorithm. GET-R and GWT-R represents the equivalent of OTS, in terms of sample size, but for the random samples. Error bars represents the standard error