The aerodynamic stability and control derivative database for the F-15 ACTIVE aircraft's six degree-of-freedom simulation is currently being modeled using neural networks. The objective is to develop pre-trained neural networks using this database, and upon achieving acceptable levels of size and accuracy, to install the neural networks on the F-15 ACTIVE aircraft for in-flight experimentation in on-line learning and reconfigurable flight controls. The material presented in this paper examines a representative subset of the entire aerodynamic stability and control derivative database in order to: 1) develop accuracy criteria that neural networks must achieve in order to accurately model the database, and 2) develop guidelines for pre-training that will help achieve the accuracies while minimizing network size. The results show that neural networks must be within ±3.77%, ±15%, or ±50%, depending on individual derivative sensitivities and relative importance rankings. Results also indicate that overall network size requirements can be reduced by 70% without significantly impacting accuracy by modeling several derivatives at once, rather than individually.