Polynomial neural networks for airborne applications

The authors have attempted to develop a polynomial neural network which will present a pilot with optimal maneuvers for missile evasion and to show the benefits of the network relative to a look-up table. They used an inductive, unsupervised learning model, which is well suited for dynamical systems and results in compact, computationally fast representations. An aircraft-missile computer simulation program is used to gather the data used in the neural network learning phase as well as in construction of the look-up table. The inputs to the neural model are variables such as aircraft velocity and initial range. Outputs are several maneuvers and associated maneuver times which provide the best chance for survival. Results show that a polynomial network can be generated that performs calculations approximating a simulator approximately a million times faster for the case of determining whether a aircraft needs to respond to a missile attack. Computer experiments suggest that for simple decision cases any of the three approaches are likely to be adequate: run a simulator in realtime, use a look-up table of precomputed runs, or use a polynomial network to approximate the simulator computations