Service Function Chain (SFC) placement under service-level agreements (SLAs) is inherently constrained and multi-objective, requiring both feasibility and flexibility in placement decisions. While greedy heuristics can achieve high success rates in resource-rich settings, they often collapse to narrow solution patterns and provide limited alternatives under constraints. This paper proposes a topology-aware reinforcement learning framework based on Graph Attention Networks and Proximal Policy Optimization (GAT-PPO) for SLA-aware SFC placement. The proposed method integrates graph-based state encoding with penalty-based reward shaping and feasibility masking, enabling the agent to learn constraint-sensitive placement policies without relying on complex constrained optimization schemes. Simulation results on randomly generated substrate networks show that the proposed method maintains high feasibility in bandwidth-constrained regimes and, more importantly, learns a thicker Pareto set in the success-diversity plane compared with greedy and random baselines. Hypervolume analysis confirms improved Pareto robustness, while training dynamics demonstrate stable learning and reduced SLA violations. These results indicate that combining constraint-aware reinforcement learning with topology-sensitive representations enables more expressive and robust SFC placement under practical SLA constraints.
SHAO et al. (Thu,) studied this question.