Interval-valued differential evolution for evolving neural networks with interval weights and biases

The ordinary differential evolution (DE) algorithm employs real-valued vectors as genotypes. The author previously proposed an extension of DE which can handle interval-valued genotypes. In this paper, the proposed method is applied to evolution of neural networks with interval connection weights and biases. Experimental results show that the interval DE can evolve neural networks which model interval functions well despite that no training data is explicitly provided.