Algorithms with Greedy Heuristic Procedures for Mixture Probability Distribution Separation

Abstract

We propose new evolutionary algorithms with greedy agglomerative heuristic procedures for fuzzy clustering problem
based on model of mixture probability distribution separation. New algorithms use the EM algorithm
(Expectation Maximization) for local search and ideas of genetic evolutionary algorithms combined with greedy
agglomerative heuristic procedures as global search strategy. This combination allows to obtain better results
with higher values of the log likelihood objective function in comparison with the EM algorithm and its modification.
Results of testing of the electronic components shipped for the space industry are represented by arrays of data vectors
of very high dimensionality, up to hundreds of dimensions. One of the most important problems for increasing the quality
of the electronic units is detection of the homogeneous production batches of the electronic devices (components).
The EM algorithm and its modifications do not guarantee obtaining optimal or near optimal results with highest
log likelihood values for such problems. Computational experiments with electronic component testing data and classical
datasets for fuzzy clustering problems show that this new algorithm allows obtaining more precise results in comparison
known algorithms.