Neurocontrollers designed by a genetic algorithm

The paper discusses problems existing in neural network design using mathematically guided training methods. It presents a genetic algorithm based design technique to train the network, which overcomes all these problems. The paper also presents suitability conditions for using the genetic algorithm based design methods and develops, under these conditions, direct neurocontrollers with a novel structure inspired by proportional plus derivative control. Techniques are also developed to select the architectures in the same process of parameter training. The proposed methods are validated by several examples, including one with plant transport delay