Logic-informed reinforcement learning for cross-domain optimization of large-scale cyber-physical systems