Strategic Cache Allocation via Game-Aware Multi-Agent Reinforcement Learning

Abstract

This paper addresses the problems of low cache allocation efficiency and unstable strategies caused by multi-tenant resource competition in edge computing environments. A game-driven resource allocation mechanism based on multi-agent reinforcement learning is proposed. The mechanism consists of two core modules: the Game-aware Adaptive Policy Optimization (GAPO) framework and the State-aware Decentralized Agent Network (SADAN). GAPO introduces a local incentive adjustment function that guides agents to make more reasonable resource allocation decisions in dynamic competitive environments. It helps avoid convergence to suboptimal game equilibria. SADAN combines neighborhood state interaction with structured state encoding. It enables agents to capture system dynamics under partial observability and enhances policy coordination and learning efficiency. The cache resource allocation problem is modeled as a multi-agent game process. The proposed learning framework is applied to an edge caching system and evaluated using real-world datasets and a constructed simulation environment. Experimental results show that the proposed method outperforms existing approaches in key metrics such as cache hit rate, response delay, and policy convergence speed. Moreover, the method demonstrates strong robustness and stable system performance under varying conditions. These include multi-tenant scaling, reduced observation completeness, and changing resource constraints. The results effectively validate the adaptability and superiority of the proposed mechanism in edge cache resource allocation tasks.