Multi-Agent Reinforcement Learning for Explainable Spacecraft Configuration Design

This work presents an attention-based multi-agent reinforcement learning (MARL) framework for spacecraft configuration design, formulated as a sequential decision process with hard geometric constraints and five competing physical objectives. Each objective is assigned to a dedicated expert trained with proximal policy optimization, while a differentiable gating network allocates control among experts at each placement step. This architecture enables experts to develop reusable spatial strategies and allows the gate to synthesize them dynamically to explore the multi-objective trade space. Across 8000 training episodes, the system expands the achievable Pareto frontier, producing diverse non-dominated configurations and steadily increasing hypervolume coverage. Interpretability analyses show that Pareto diversity emerges not solely from objective-linked specialization but from coordination regimes in which the gate recombines a small set of geometric “skills” that generalize across objectives. One expert associated with wire spatial entropy minimization consistently acts as a robust generalist whose coherent placements benefit multiple criteria. Spatial density maps and surrogate gating models reveal that routing decisions follow smooth, low-dimensional geometric partitions governed primarily by distances to the bus center and power hub. These findings indicate that the modular multi-agent architecture naturally discovers interpretable spatial skills whose coordinated use expands the attainable Pareto frontier.