Inserting background knowledge in perceptrons through modification of the learning algorithm

Usually, knowledge to be learned by neural networks is represented implicitly in the training samples. The ability to insert knowledge apart from the implicit representations in training samples (“background knowledge”) gives rise to the hope that the learning and operation behavior of neural networks can be improved. In this paper, we develop a method to accomplish the insertion of expert knowledge into the error function during training. We modify the backpropagation learning algorithm such that the network is trained not only to minimize output error but also to consider further information provided by the expert users who train a multilayer perceptron with one hidden layer. The results are tested with an artificial example from design cost estimation using very small training set sizes of 10 samples. They show significant improvement compared to approaches which do not consider background knowledge