|
M. E. H. Pedersen
Ceci n'est pas une Homepage
ThesisSimplifying Swarm Optimization
AbstractThis thesis is about the simplification of black-box / direct-search optimization methods, which by definition do not use gradient information to guide their search for an optimum, but merely need a fitness measure for each candidate solution to the optimization problem. To achieve such simplification, a technique for tuning the behavioural parameters of an optimization method is first developed. This technique is known as Meta-Optimization, in that it employs an additional layer of optimization on top of the method whose behavioural parameters are to be tuned. Meta-Optimization is then used ad hoc to identify which behavioural parameters can be eliminated from an optimization method. It is found that simplifications of two popular swarm-based / multi-agent optimization methods are indeed possible. When the methods are used for optimizing simple benchmark problems in very long optimization runs, the optimization methods cannot only be made a great deal simpler, but their performance on those benchmark problems are also greatly improved through this tuning and simplification. This performance increase however, does not translate to a real-world case study, for which the optimization methods must be specifically tuned and simplified, if not to impair their performance. The cause of this might be the different search-space topologies for the benchmark and real-world problems, but it may just as well be caused by the significantly fewer optimization iterations used for the real-world problems than used for the benchmark problems. The results of these studies therefore suggest that the longer optimization runs are allowed, the simpler the optimization method can be made. Download
(C) Copyright 2006-2008 Hvass Labs. All rights reserved. |