Multi-objective reinforcement learning scheduling framework for heterogeneous AI chips: Energy-efficient and latency-aware edge-cloud collaboration

Scheduling AI inference across heterogeneous edge-cloud chips requires balancing energy, latency, cost, and thermal feasibility. This study presents a chip-aware scheduling framework, rather than a new multi-objective reinforcement learning theory. A 53-dimensional state describes directed-acyclic-graph tasks, CPU/GPU/NPU/FPGA status, network conditions, and queue slack. A fuzzy controller adjusts energy, latency, and cost priorities, while a hybrid-action Proximal Policy Optimization policy jointly selects the offloading target, physical chip, and dynamic-voltage-and-frequency-scaling coefficient. Dependency, deadline, thermal, and bandwidth constraints are enforced through action masks and residual penalties. In traffic-video and industrial-inspection simulations, the framework achieved 5.8 TOPS/W, 132 ms average latency, a normalized cost coefficient of 0.17 per task, and 85.2% Pareto coverage. Under an equal 1000-episode and 30-seed budget, PPO reached 95% of its asymptotic improvement in 580 ± 45 episodes, obtained a final normalized return of -0.15 ± 0.03 and an HV of 0.87 ± 0.02, and produced no divergent seed. EdgeCloudSim/iFogSim2 replication and hardware-in-the-loop testing preserved the advantage over DTRL; in hardware-in-the-loop tests, the framework required 31.5 ± 3.5 J per task and 142 ± 8.0 ms, compared with 35.8 ± 4.0 J and 155 ± 9.5 ms for DTRL. The results indicate that task-aware objective adaptation and chip-level control can reduce wasteful offloading and thermal stress while preserving service timeliness.