一种基于深度强化学习的多对多在轨服务优化调度方法

An intelligent approach based on deep reinforcement learning is proposed to address the optimization scheduling problem for multi-to-multi on-orbit spacecraft services. Initially，this problem is modeled as an orbit-related vehicle routing problem. Subsequently，an encoder-decoder neural network，equipped with an attention mechanism，is introduced to construct a stochastic policy that generates solutions for given problem instances. Within this framework，the encoder is responsible for producing graph embeddings and node embeddings，while the decoder generates solutions in a step-by-step manner based on these embeddings. The neural network is then trained utilizing the REINFORCE algorithm，augmented with a greedy rollout baseline. Extensive experimental results ultimately demonstrate the effectiveness and superiority of the proposed method. The advantages of this intelligent approach are manifold. It provides near real-time solutions to scheduling problems，offers superior solution quality for large-scale scheduling problems compared to meta-heuristic algorithms，and exhibits good generalization ability，as models trained on instances with a specific number of targets can be applied to instances with differing numbers of targets.