Artificial evolution of neural networks and its application to feedback control

This paper develops novel neural networks suitable for direct embedment within a feedback loop. The structure of the embedded system is inspired by the proportional plus derivative control with moving average excitations. A parameter vector based uniform description of the problem of neural network design is presented. Difficulties associated with traditional mathematicallyguided design methods are discussed, which leads to the development of a genetic algorithm based evolutionary design method that overcomes these difficulties and makes direct neurocontrollers possible. Techniques are also developed to optimise the architecture in the same process of parameter training, leading to Darwin neural machines. The proposed methods are verified by examples of direct neurocontroller design for a delayed linear plant and a nonlinear plant.