Techniques for optimization of a dairy farm simulation model

Abstract

Dairy farms are complicated biological and managerial systems with many control variables. New Zealand's dairy farm research organization, Dexcel Ltd, has developed a computer simulation model, the Dexcel WFM (Whole Farm Model), to facilitate research into optimal farm configuration and management strategy. However, the complexity and high dimensionality of the Dexcel WFM makes its computational optimization a considerable technical challenge. This dissertation describes work that attempts to address some of the issues associated with that challenge.

The application of two types of "Black-box" optimization techniques, namely Lipschitz Branch-and-Bound Algorithm (LBBA) and Genetic Algorithm (GA) were studied in some detail. A novel development of the LBBA was proposed and shown to have superior performance in the high-dimensional problems of interest. A real coded implementation of the genome in the GA was demonstrated to be superior to its binary-coded counterpart, and the subsequent empirical comparison of the LBB and GA showed the GA to be more efficient on optimizing the Dexcel WFM as well as some other high-dimensional problems. Also described are studies on the effect of some implementation issues of the GA, including the identification of the best GA parameter configuration, the robust parallelization of the GA and the effect of a Tabu list on GA efficiency. Visualizations of the results of these investigations are presented using Sammon Mapping and Parallel Coordinate techniques to illustrate the topography of the Dexcel WFM response (objective) function.