The selection of informative markers has been studied massively as an alternative to reduce genotyping costs for the genomic selection (GS) application. Low-density marker panels are attractive for GS because they decrease computational time-consuming and multicollinearity beyond more individuals can be genotyped with the same cost. Nevertheless, these inferences are usually made empirically using "static" training sets and populations, which are adequate only to predict a breeding program's initial cycles but might not for long-term cycles. Moreover, to the best of our knowledge, none of these inferences considered the inclusion of dominance into the GS models, which is particularly important to predict cross-pollinated crops. Therefore, that reveals an important and unexplored topic for allogamous long-term breeding. To achieve this goal, we employed two approaches: the former used empirical maize datasets, and the latter simulations of long-term breeding cycles of phenotypic and genomic recurrent selection (intrapopulation and reciprocal). Then, we observed the reducing marker density effect on populations (mean, the best genotypes performance, accuracy, additive variance) over cycles and models (additive, additive-dominance, specific combining ability (SCA)). Our results indicate that the markers reduction based on different linkage disequilibrium (LD) levels is viable only within a cycle and brings a significant decrease in predictive ability over generations. Furthermore, in the long-term, regardless of the selection scheme adopted, the more makers, the better because they buffer LD losses caused by recombination over breeding cycles. Finally, regarding the accuracy, the additive-dominant models tend to outperform the additive ones and perform similar to the SCA.

Genomic prediction (GP) success is directly dependent on establishing a training population, where incorporating envirotyping data and correlated traits may increase the GP accuracy. Therefore, we aimed to design optimized training sets for multi-trait for multi-environment trials (MTMET). For that, we evaluated the predictive ability of five GP models using the genomic best linear unbiased predictor model (GBLUP) with additive + dominance effects (M1) as the baseline and then adding genotype by environment interaction (G × E) (M2), enviromic data (W) (M3), W+G × E (M4), and finally W+G × W (M5), where G × W denotes the genotype by enviromic interaction. Moreover, we considered single-trait multi-environment trials (STMET) and MTMET for three traits: grain yield (GY), plant height (PH), and ear height (EH), with two datasets and two cross-validation schemes. Afterward, we built two kernels for genotype by environment by trait interaction (GET) and genotype by enviromic by trait interaction (GWT) to apply genetic algorithms to select genotype:environment:trait combinations that represent 98% of the variation of the whole dataset and composed the optimized training set (OTS). Using OTS based on enviromic data, it was possible to increase the response to selection per amount invested by 142%. Consequently, our results suggested that genetic algorithms of optimization associated with genomic and enviromic data efficiently design optimized training sets for genomic prediction and improve the genetic gains per dollar invested.