Toward a general framework for AI-enabled prediction in crop improvement

A theoretical framework for AI and ensembled prediction for crop improvement is introduced and demonstrated using the logistic map. Symbolic/sub-symbolic AI-based prediction can increase predictive skill with increase in system complexity. The curse of dimensionality in genomic prediction has been established and hampers genetic gain for complex traits. Artificial intelligence (AI) that fuses symbolic and sub-symbolic approaches to prediction is emerging as an approach that can deal effectively with this problem. By leveraging information across physiological and genetic networks, it is plausible to increase prediction accuracy by harnessing prior knowledge and computation approaches. Ensembles of models de facto implement the diversity prediction theorem, and thus enable breeders identify subnetworks of genetic and physiological networks underpinning crop response to management (M) and environment (E). Here, we introduce a theoretical framework for AI-enabled prediction in crop improvement. This framework brings together elements of dynamical systems modeling, ensembles, Bayesian statistics and optimization. We demonstrate properties of this framework and limits to predictability using a simple logistic map. We show that heritability and level of predictability decrease with increase in system complexity that conforms well with prior empirical evidence. We show that predicting systems states is an inferior strategy to predicting system process rates for complex systems. This holds for both the level of predictability and for the ability to use the data generating functions to produce a view of the system state space that can help breeders develop an intuition for how biological interventions can affect the performance of the crops. By integrating biological knowledge and computational approaches to prediction, it is feasible to increase predictive accuracy in breeding systems and therefore hasten the rate of genetic gain.