Optimisation techniques for a computer simulation of a pastoral dairy farm

This paper compares different methods of optimising the management variables in UDDER, a commercially-available computer simulation model of a pastoral dairy farm. The emphasis is on identifying the best optimisation strategy for this complex multi-dimensional system, taking the simulation model as a given constant. The optimisation methods studied are based on significantly different principles, with differing strengths and weaknesses: two hill-climbing algorithms (Nelder-Mead simplex and Powell's direction set), and a genetic algorithm (GA). Rather than examine all facets of dairy farm management, a single problem is optimised-that of maximising milkfat production while maintaining the health of the herd and pasture. The results show that while the GA can determine good regions within the search space quickly, it is considerably slower than either hill-climber at finding the optimal point within that region. The hillclimbers, in contrast, are fast but have a tendency to get trapped on local maxima and thus fail to find the true optimum. This led to the development of a hybrid algorithm which utilises the initial global search of the GA, followed by the more efficient local search of a hill-climber. This hybrid algorithm discovered near-optimal points much more quickly than the GA, and with more reliability than the hill-climber. (C) 1998 Elsevier Science B.V.