Statistical Evidence for a DQN-Family Advantage in a Resource-Coupled Satellite-Ground Scheduling Benchmark

Abstract

Resource-coupled satellite-ground scheduling requires decisions whose computational, energy, thermal, and communication effects persist across task arrivals. We tested whether the deep Q-network (DQN) family provides a reproducible advantage over short-horizon planning in a synthetic benchmark with clipped, soft-constrained resources. Six DQN objectives used the same state representation, network architecture, replay procedure, optimizer, interaction budget, 20 model seeds, and held-out traces. Across five fully coupled regimes, every one of the 30 objective–regime comparisons had lower mean episode cost for DQN than for four-step model-predictive control (MPC-4). Every paired 95% bootstrap interval excluded zero, and all Holm-adjusted sign tests supported this ordering. The least-favorable mean difference was −5.92, with a family-wide simultaneous 95% upper bound of −5.03. Differences among DQN objectives were smaller and scenario-dependent. Additional nominal tests assessed discount-factor, training-length, greedy-rollout, and CPU decision-time sensitivity. These findings apply only to this synthetic, soft-constraint benchmark. Hard energy and thermal limits can change the feasible action set and therefore the algorithm ordering.