Attention Network for the Fast Generation of a Multitarget Rendezvous Sequence

Global trajectory optimization of multitarget successive rendezvous missions is studied. A learning-based method is proposed to train a neural network offline, which can directly generate the near-optimal rendezvous sequence online. A neural network structure based on the attention mechanism is developed to extract combinatorial features among multiple on-orbit bodies. Physical equations and neural networks are integrated to generate solutions in a recurrent manner. The sequence generator can be trained and applied in cases with diverse values and sizes. A local refinement approach is applied to the generator to enhance the final solution quality. A reinforcement parallel learning algorithm is specifically designed for efficient model training. Numerical simulations are conducted on general random cases and global trajectory optimization competition (GTOC) problems. The well-trained solver can efficiently produce a sequence that is equivalent to the solution obtained via traditional advanced heuristics. When the trained sequence generator is applied to multispacecraft rendezvous sequence planning, the target assignment optimization process clearly accelerates due to fast sequence generation. The sequence and time of the GTOC winner's solution is reoptimized and better results were obtained, indicating that the proposed model performs significantly well in terms of solution quality.