A Parallel Genetic Algorithm for Pattern Recognition in Mixed Databases

Structured data bases may include both numerical and non-numerical attributes (categorical or CA). Databases which include CAs are called “mixed” databases (MD). Metric clustering algorithms are ineffectual when presented with MDs because, in such algorithms, the similarity between the objects is determined by measuring the differences between them, in accordance with some predefined metric. Nevertheless, the information contained in the CAs of MDs is fundamental to understand and identify the patterns therein. A practical alternative is to encode the instances of the CAs numerically. To do this we must consider the fact that there is a limited subset of codes which will preserve the patterns in the MD. To identify such pattern-preserving codes (PPC) we appeal to neural networks (NN) and genetic algorithms (GA). It is possible to identify a set of PPCs by trying out a bounded number of codes (the individuals of a GA’s population) and demanding the GA to identify the best individual. Such individual is the best practical PPC for the MD. The computational complexity of this task is considerable. To decrease processing time we appeal to multi-core architectures and the implementation of multiple threads in an algorithm called ParCENG. In this paper we discuss the method and establish experimental bounds on its parameters. This will allow us to tackle larger databases in much shorter execution times.