Optimisation of a complex simulation model

Abstract

In this paper we describe techniques utilised in the development of a scheme for identifying the regions in an 8-dimensional parameter space that gave optimal (or near-optimal) performance in a computational simulation of a real-world system. The system model, developed by Dexcel Ltd, attempts a detailed representation of pastoral dairying scenarios. It incorporates sub-models, themselves complex in many cases, of pasture growth, animal metabolism etc. Each evaluation of the objective function, a composite 'farm performance index', requires simulation of at least a one-year period of farm operation with a daily time-step and hence is computationally expensive. Since similar situations are likely to arise in other practical optimisation exercises, the results presented should have some quite general applicability.Two quite different methods of optimisation - Genetic Algorithm (GA) and Lipschitz Branch-and-Bound (LBB) algorithm are investigated and contrasted. Practical issues related to their efficient implementation in a Linux cluster parallel processing environment are discussed and their performance on the above problem is compared. The problem of visualisation of the objective function (response surface) in high-dimensional spaces is also considered in the context of the farm optimisation problem (where from a practical viewpoint knowledge of its behaviour in the region of optima is actually more important than the precise positions or values of the optima themselves). An adaption of the Parallel Coordinates visualization is described which helps visualise some important properties of the model’s output topography.