Sample size issues in the choice between the best classifier and fusion by trainable combiners