Three-valued neural networks for test generation
Hideo Fujiwara · 2002
A three-valued (0, 1, and 1/2) neural network, which is an extension of the binary Hopfield model, is proposed, and it is shown that the test generation problem can be solved by the three-valued model more effectively than by the binary model. In the three-valued model, the energy function of networks, hyperplanes of neurons, and update rules of neuron states are extended so that the third value, 1/2, can be treated satisfactorily. It is proved that the proposed three-valued model always converges. To escape from local minima, an extension of Boltzmann machines, in which the update rules are modified by introducing probabilities of neuron states, is presented.>