The pointer network for reward maximisation in multi-target space mission sequence selection

Multi-target space mission scenarios such as asteroid rendezvous, debris removal or satellite servicing, require targeting several orbits in a single mission, often to be selected among a large set, and therefore choosing optimal sequences of these orbits to be visited. This paper demonstrates a reinforcement-learning-based framework for selecting the sequence of targets to be visited in large-scale multi-target mission optimisation problems. The sequence selection is a NP-hard combinatorial optimisation problem. The proposed method builds upon a neural network architecture for combinatorial optimisation originally developed for Euclidean problems, to produce estimates of the optimal sequence of targets in very short amounts of time. The neural network is trained using a policy-gradient reinforcement-learning approach. Once training is complete, the network can be evaluated in two ways: one of these (greedy decoding) produces solutions on average 15 % less optimal than Ant Colony Optimisation (ACO); the other (stochastic search) is on average 5 % less optimal than ACO, using an iterative process that is slower than greedy decoding but still orders of magnitude faster than ACO. The quality of the network’s solutions is shown both averaged over large amounts of problems, and demonstrated more closely on a few specific instances.