Machine Learning Techniques for Approximation of Objective Functions in Trajectory Optimisation

Objective function landscape approximation methods in evolutionary optimisation can be a beneficial technique, since they replace part of the calls to the original, often computationally expensive objective function, with calls to a faster and computationally cheaper function approximator, such as a polynomial or an artificial neural network. Moreover, in some cases, the approximate model smoothens rough fitness landscapes, facilitating the stochastic search. In this paper, we apply a neural network-based approximation technique to spacecraft interplanetary trajectory problems. These multimodal problems, recently introduced in the global optimisation community, can be very complex and characterised by the prevalence of many local optima and, in the worst cases, a heavy computation load involved with the calculation of the objective value of a given input vector. We perform the first steps to integrate an approximated model into a trajectory optimisation process, building a hybrid system where both the original trajectory model and the approximated model are carefully used in the optimisation process without degrading the quality of the final trajectory found.