Discrete-time cellular neural network construction through evolution programs

The problem of the definition of the coefficients of a cellular neural network (CNN) system has been solved in many different ways. The first possibility is to choose them “by hand”, “mathematically” defining the function to be performed by the network. In some special case the applicability of “traditional” techniques such as back-propagation has been successfully demonstrated A more general approach to CNN training has been obtained using genetic algorithms. In this paper, a new kind of cellular neural network learning algorithm is presented, based on an evolution program, that is a “generalisation” of genetic algorithms. The evaluation of the fitness is run on a massively parallel system, the Connection Machine CM-2. This approach has been applied with promising results to the automatic design of CNN-based filter for image segmentation. A detailed description of the algorithm and an analysis of its computational cost are presented