Learning mixture models using a genetic version of the EM algorithm

The need to find new pattern recognition techniques that correctly classify complex structures has risen as an important held of research. A well-know n solution to this problem, which has proven to be very powerful, is the us e of mixture models, Mixture models are typically fitted using the expectat ion-maximization (EM) algorithm. Unfortunately, optimal results are not alw ays achieved because the EM algorithm, iterative in nature, is only guarant eed to produce a local maximum. In this paper, a solution to this problem i s proposed and tested in a complex structure where the classical EM algorit hm normally fails. This, we will do by means of a genetic algorithm (CA) wh ich will allow the system to combine different solutions in a stochastic se arch so as to produce better results. The reported results show the usefuln ess of this approach, and suggest how it can be successfully implemented. T wo new algorithms are proposed. The first one is useful when a priori infor mation of the observed data is not available. The second solution is useful for those cases where some knowledge of the structure of the data-set is k nown. This second solution has proven to converge faster than the first one , although the final results reached are very similar to each other. (C) 20 00 Published by Elsevier Science B.V. All rights reserved.