Hill Crunching Clustered Genetic Search and its Improvements
Henryk Telega, Igor T. Podolak · Repository of Wyższa Szkoła Biznesu – National-Louis University (Wyższa Szkoła Biznesu – National-Louis University) · 2005
Two modifications to the Hill Crunching Clustered Genetic Search (HC-CGS) algorithm are proposed in this paper. HC-CGS (see [Telega 1999], [Adamska et al. 2004] is a global optimization algorithm that was designed in order to solve such parameter inverse problems in which an approximation of certain level sets (central parts of basins of attractions of local minimizers) is required. The approximation of these sets can be useful when some additional criteria of optimization are considered after main results of parameter identification are obtained. The approximation is also helpful in stability analysis. In spite of some good properties of HC-CGS, tests have shown that its original version can be not effective for problems with more than 4 dimensions. Two modifications of HC-CGS are proposed in order to overcome the dimensionality limitation. In the first one clusters are remembered as ellipsoids. The first modification is based on the idea of cluster recognition with the use of Kohonen Self Organizing Maps (SOM) neural networks [Kaski 1997]. In the second one clusters are remembered as ellipsoids.