A Deep Reinforcement Learning-Based Multi-Resource Pricing Game in Computing Power Network

Abstract

Abstract Computing Power Network (CPN) integrates network, computing and storage resources to achieve efficient collaboration of cloud, edge and end, and meet the industry's demand for highly differentiated computing services. In contrast to fixed pricing, dynamic pricing allows mobile users (MUs) to request and pay for resources based on their demands. Resource providers (RPs) meet the needs of MUs through personalized network services and benefit from differentiated pricing. Traditional edge computing is limited to study the pricing of single resource, but the integration of multiple resources in CPN makes the scene more complex. To address this problem, we propose a multi-leader multi-follower Stackelberg game model for the dynamic pricing of multiple resources problem between RPs acting as leaders and MUs acting as followers. Specifically, we first give the initial unit price announced by RPs, MUs use the trust region method to solve the optimal purchase amount of computing resources and communication resources. After that, each leader observes the request and iteratively adjusted the pricing strategy based on the Multi-Agent Deep Deterministic Policy Gradient (MADDPG). Simulation results show that stakeholders' benefits in multi-resource pricing are better than that in single-resource pricing.