Traditional and evolved dynamic neural networks for aircraft simulation

This paper presents results in applying gradient and evolutionary programming (EP) techniques to training dynamic neural network models of aircraft response. The gradient methods modify the weights of predefined neural network structures to learn the desired mapping. We show that this approach is quite effective as long as the predefined network topology is capable of modeling the dynamic system. We examine several dynamic neural network structures: two recurrent architectures and the memory neuron network. The evolutionary programming algorithm determines not only the weights of a dynamic neural network, but also the topology. The EP algorithm is applicable to a much broader class of problems, since a predefined topology does not need to be in place beforehand and the dynamics of the system do not need to be known