Rule extraction from artificial neural network with optimized activation functions

A novel method of rule extraction from artificial neural network with optimized activation function is proposed. Weight-decay approach is used in training and the unnecessary connections in the neural network are pruned at the cost of an increase in the error function within a predetermined limit. A penalty term is added in the activation function to facilitate the values of hidden and output nodes to have better approximation to 0 or 1, which is of great help in symbolic rule extraction in neural network. With the optimized activation function, the rule extraction becomes much easier and simpler. Rule extraction has been experimented on two public datasets of iris and breast-cancer, which results showed that the proposed method has a better rule overcast accuracy than the commonly used methods, such as decision tree algorithm C4.5 and RX algorithm.