Near-Optimal Flight Trajectories Generated by Neural Networks

Abstract

A particular rendezvous mission is modeled, which is difficult to perform in practice by pilots. It serves as a demonstration of using neural networks to generate control laws. The problem emphasizes the need for a control law since it is non-linear and has a sensitive element included. A particular formulation of the optimal criteria is made. The optimization method used must be efficient and cheap to use as a large set of optimal solutions is needed as learning material for the neural networks. The worst problem with neural network is to show it all information needed for control within the whole state space. The application in this paper shows some of the difficulties found during our experimental sessions.