In commercial breeding, new genotypes are constantly being created and need to be tested to understand how a specific seed will perform in its target locations. A major constraint is that a genotype needs to go through multiple years of testing before it can be commercialized. With the volume of new genotypes that are constantly being enhanced, it is unrealistic to test every genotype in every target environment. Here, a methodology has been created that considers the fact that there are limited resources, whether it be limited space or a limited number of each genotype available in a single planting season. This new approach works by using the observations of genotypes that were planted and then inferring the performance of specific genotypes in certain environments. For agricultural crops, not all genotypes respond in the same way when planted in a certain environment. This phenomenon is describing genotype by environment (GxE) interaction. Numerous methods exist that aim to predict plant performance and specifically quantify and understand the GxE interaction. Here, five models are first evaluated on four different crop datasets. The Biclustering model is one model considered and it is effective at determining which genotypes have no GxE interaction in a subset of environments. This model works well with sparse data which is what exists in practice. Therefore, the Biclustering model is used to find subsets of genotypes and environments that have little to no GxE interaction.

In a subset of genotypes and environments with no interaction, genotypes can be planted in a strategic, methodical pattern so that the phenotype of unplanted genotypes can be inferred. Depending on the amount of physical resources available, two approaches can be utilized to gain information about unplanted genotypes. Given a set number of genotypes that can be planted, the first approach aims to maximize the number of known genotype/ environment pairs. The term genotype/environment pair refers to the phenotype that exists for a single genotype in a single environment. The second approach determines how many observations are required to infer every genotype/environment pair within a dataset. Additional constraints can be introduced to create a more realistic model.

The effectiveness of these two approaches can be illustrated using small-scale experimental designs that can be translated to full-scale commercial cases. In order to evaluate the effectiveness of the experimental designs created, both optimized and random models are compared to the original phenotypic responses. Validation indicates that optimizing the location of genotypes allows more inferences to be made, implying that creating an optimized planting plan can improve the understanding of genotypes. If this approach is applied in practice, it can facilitate further research as additional information can be gained from existing resources.