Are ensemble-average learning curves reliable in evaluating the performance of adaptive filters?