Why to merge neural networks is more favorable than to compete