H∞ Optimal Training Algorithms and their Relation to Backpropagation