Transfer Learning for Sample-Efficient Training of Space Trajectory Cost Approximators

The preliminary stage of space mission design often requires evaluating a large number of orbital transfers. This is especially the case for multi-target missions, where a combinatorial problem involving millions of orbital transfers must be solved. Existing techniques for calculating the cost of each transfer are time-consuming, so can render the combinatorial problem intractable. Machine learning has been used in the last decade to accelerate the calculation of orbital transfer cost, at the expense of some accuracy. However, the machine learning approach incurs an upfront computational expense for generating a database of optimal transfers, and training an artificial neural network (ANN) to approximate them. This work addresses the scenario where a number of spacecraft models, each with its own thrust, specific impulse and mass, need to to analysed for a mission. Normally, this would involve creating a large database and training a different ANN for each spacecraft. In the proposed approach, an ANN is trained on one spacecraft (acting as a ‘baseline’), and fine-tuned on smaller databases for each other spacecraft. It is shown that, if a sufficiently good accuracy can be obtained on the baseline database, the fine-tuning approach requires fewer training samples to achieve a given accuracy than if an entirely fresh model were trained, and saves time on the hyperparameter tuning required to construct a fresh model. Therefore, we propose to train an ANN once on a baseline spacecraft, then transfer this knowledge to an arbitrary number of other spacecraft by fine-tuning the baseline ANN.