Many-Objective Optimization Approach using Surrogate Models in Rotational Cattle Grazing

Abstract

In the context of many-objective problems, one of the most important problems is to be able to efficiently model and evaluate the different objectives of the problem. A strategy to speed up the computation, as well as to manage the uncertainties due to partial knowledge of the context, is to use surrogate models. The objective of this paper is to evaluate the hybrid use of surrogate models in the context of a many-objective optimization problem for livestock management. In particular, we propose to use hybrid objective functions, in which parts of the objective function originally constructed analytically are replaced by data-driven surrogate models. Specifically, we explore several hybrid schemes, where in each we combine different parts of the analytical objective function with other data-driven parts. Preliminary results show that the hybrid optimization models studied possess a competitive performance quality against the original purely analytical model, becoming an interesting proposal to manage the computational times and uncertainty of environments such as livestock farming. These results show the great potential of using models built with machine learning techniques to replace analytically constructed objective functions that are not always able to absorb the non-deterministic nature of livestock grazing, and represent an opportunity to further explore their usefulness in this context.