In this paper, the concept of constrained neural networks" is introduced, and the It's usefulness is validated by several examples. In conventional approach, the neural network structure is developed without considering the underlying structure of the problems. Hence, conventional networks contain a large number of weights and require considerable training patterns and time. A novel neural network structure is proposed for constrained problems. Symmetric neural networks are also proposed and can be considered as an example of constrained neural networks. Because these neural networks can not be trained by the conventional training algorithm such as back propagation, which destroy the structures of the weights,an appropriate training algorithm is developed. It's effectiveness is demonstrated by applications."
