Interval neural networks

Abstract

In the paper we consider an architecture and properties of neural networks that have interval weights and interval biases. This model of a neural network takes into consideration inaccuracies in technical realisation of neuron in-out characteristics. A neural network with such architecture maps an input vector into interval response. We consider an architecture of four-layer feedforward network. A learning algorithm is derived from the cost function in a similar manner to the backpropagation algorithm. We examined properties of these nets using computer simulation.