An interval computation approach to backpropagation

A new learning algorithm is introduced to allow interval valued inputs. This procedure is based on gradient descent and error backpropagation. It has many advantages over the classical backpropagation including the possibility of dealing with missing values attributes. We establish a framework especially useful for applications with incertitude in the input. We propose a cost function reflecting a trade-off problem between the goal of including the target value into the interval valued output, and minimizing this interval size. Simulated numerical results are presented