A comparison of neural network training algorithms

Six methods of training neural networks to represent aerodynamic data are compared. Prior research has shown that the extended Kalman filter provides speed and accuracy advantages over the back-propagation method, and further quickening of training is desirable. We investigate four filter-based approaches (standard filter without process "noise", setting lower bounds on estimate-error covariance, periodically re-initializing the estimate-error covariance, and adding fictitious "process noise") and compare them to approaches incorporating genetic algorithms. A genetic algorithm uses a global search of the network weight space to identify feasible weights. Here, the algorithm is used independently and as a starting step for an extended Kalman filter. While the initial global search is intuitively appealing and may have value in some applications, the randomly initialized additive-noise filter is found to be the fastest training method in the present study.