Multi-appointment missions: a transformer estimates the cost of a satellite round trip 100,000 times faster than an optimisation algorithm

Trained on 850,000 sequences, the model evaluates a route in less than a millisecond with a 5 per cent error rate, and outperforms the best-known solutions to the GTOC9 problem.

Designing a mission where several spacecraft must join a series of targets — debris to be removed, satellites to be maintained — requires thousands of times the total cost of a tour. It is this calculation, costly if it is precisely and imprecisely optimized if it is estimated segment by segment, which hinders the search for the best solutions.

Ignore the order of targets

In an article filed on 6 October on arXiv, An-yi Huang proposes a network that directly predicts the cost of an unordered set of targets. The transfer costs between each pair of targets are summarized into vectors, then an encoder transforms insensitive to permutations by deducting the overall cost, taking into account the time window. Driven on about 850,000 labelled sequences, it reaches a relative error close to 5% and responds in less than one millisecond, about five orders of magnitude faster than accurate optimization.

Integrated into evolutionary research and applied to the GTOC9 problem (removal of debris in low orbit), it provides the best known solution for fleets of seven to ten vehicles, and serves as a starting point for refining that further improves published records. The method depends only on a cost function between pairs of targets: it also applies to satellite maintenance or visits of several bodies in remote space.