Study of Specific Genetic Operators for Learning Kohonen Maps in Function of Initial Conditions

In this paper, we propose a new hybridization method of Self Organizing Maps (SOM) with genetic algorithms in order to design optimal quantified vectors of input speech signals. In particularly in the present work, we demonstrate that it is necessary to effectuate choice criteria of genetic operators concerning the way to evolve individuals. Based on the genetic algorithm, we determine which type of genetic operators is adequate to initial conditions and the objective to reach by guaranteeing the production of viable population converging towards similar performances. Experimental results show that the use of these operators depends on initial conditions and the objective to realize. The inter maps crossover allows decreasing the quantization error of a map. However, the intra map crossover allows decreasing the topological error of a map. Both recombination operators when applied successively aim at decreasing the distortion error of a map.