Multi-Timescale Latent-Action DRL for Joint Optimization in Edge-Cloud Networks

Short summary

This paper proposes a two-timescale multi-layer deep RL framework with latent action space (2T-MDRL-LA) to jointly optimize service placement, computational delegation, and power control in edge-cloud networks. By decomposing the NP-hard problem into long-term and short-term subproblems and using a VAE-based latent action representation, the framework achieves up to 20.8% latency reduction and 13% better resource utilization. It converges approximately 50% faster than conventional PPO while approaching branch-and-bound optimality.

• •Two-timescale DRL framework with VAE-based latent action space for edge-cloud joint optimization
• •Achieves 20.8% latency reduction and 13% resource utilization improvement over baselines without computational delegation
• •Converges ~50% faster than conventional PPO while approaching branch-and-bound optimal performance

Generated with AI, which can make mistakes.