Conceptos Básicos
Egret, a deep reinforcement learning-based approach, achieves near-optimal revenue for the edge computing service provider by determining the optimal price and visiting order for clients in a sequential computation offloading mechanism, without requiring clients' private preferences.
Resumen
The paper proposes a novel sequential computation offloading mechanism (SCOM) in edge computing, where the edge computing service provider (ECSP) posts prices of computing resources with different configurations to clients in turn. Clients independently choose which computing resources to purchase and how to offload based on the posted prices, without disclosing their preferences.
To maximize the ECSP's revenue in this SCOM setting, the authors design Egret, a deep reinforcement learning-based approach. Egret determines the optimal price and visiting order online, without considering clients' preferences.
The key highlights are:
- Egret achieves near-optimal revenue, only 1.29% lower than the theoretical optimum in the SCOM setting, and 23.43% better than the state-of-the-art in a dynamic SCOM setting where clients arrive dynamically.
- Egret employs several techniques to enhance the DRL training, including price ranking, invalid action masking, and state space optimization, which significantly improve the training efficiency and performance.
- Extensive experiments demonstrate Egret's superiority over other baselines in both static and dynamic SCOM settings, in terms of revenue, client offloading decisions, and resource utilization.
Estadísticas
The average revenue margin of Egret is 31.22% lower than the Oracle solution across different trace lengths.
The average revenue margin per step of Egret is around 0.48 across different trace lengths.
Citas
"Egret determines the optimal price and visiting order online without considering clients' preferences."
"Experimental results show that the revenue of ECSP in Egret is only 1.29% lower than Oracle and 23.43% better than the state-of-the-art when the client arrives dynamically."