Abstract:
This thesis is primarily concerned with a description of four types of stochastic algorithms, namely the genetic algorithm, the continuous parameter genetic algorithm, the particle swarm algorithm and the differential evolution algorithm. Each of these techniques is presented in sufficient detail to allow the layman to develop her own program upon examining the text. All four algorithms are applied to the optimization of a certain set of unconstrained problems known as the extended Dixon-Szegö test set. An algorithm's performance at optimizing a set of problems such as these is often used as a benchmark for judging its efficacy. Although the same thing is done here, an argument is presented that shows that no such general benchmarking is possible. Indeed, it is asserted that drawing general comparisons between stochastic algorithms on the basis of any performance criterion is a meaningless pursuit unless the scope of such comparative statements is limited to specific sets of optimization problems. The idea is a result of the no free lunch theorems proposed by Wolpert and Macready. Two methods of presenting the results of an optimization run are discussed. They are used to show that judging an optimizer's performance is largely a subjective undertaking, despite the apparently objective performance measures which are commonly used when results are published. An important theme of this thesis is the observation that a simple paradigm shift can result in a different decision regarding which algorithm is best suited to a certain task. Hence, an effort is made to present the proper interpretation of the results of such tests (from the author's point of view). Additionally, the four abovementioned algorithms are used in a modelling environment designed to determine the structure of a Magnetic Cataclysmic Variable. This 'real world' modelling problem contrasts starkly with the well defined test set and highlights some of the issues that designers must face in the optimization of physical systems. The particle swarm optimizer will be shown to be the algorithm capable of achieving the best results for this modelling problem if an unbiased <font face="symbol">c</font>2 performance measure is used. However, the solution it generates is clearly not physically acceptable. Even though this drawback is not directly attributable to the optimizer, it is at least indicative of the fact that there are practical considerations which complicate the issue of algorithm selection.
