Genetic Algorithms Training of Neural Nets for Aircraft Guidance and Control

Glenn Gebert and Murray Anderson, Sverdrup Technology, Inc., USA; Johnny Evers, Air Force Research Laboratory, USA

As airframe become more and more complex, and are called upon to perform increasingly stressful maneuvers, autopilots must be robust enough to adequately stabilize the airframe in the highly non-linear, strongly cross-coupled environments. Classic autopilot design can achieve stability throughouthe flight envelope, but generally lack robustness for design and environmental ehanges. Guidance and control routines composed of a neural net architecture offer a promising ability to process multiple inputs, generate the appropriate outputs, and provide greater robustness. However, difficulty can arise in the training process of the neural nets. In the present study, a feedforward neural net was used for the guidance and control routines on typical airframes. The neural nets were trained through genetic algorithms. The work attempts to model the biological process of the "thinking" aspect of the airframes by the us of a neural nets trained through natural slection as put forth in the Theory of Evolution. The present study produced an autopilot that learned to control itrates and maneuver (with a full six degrees-of-freedom) across an arena to a target.

This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.