Model-Based Q-Learning for Dynamic Cache Allocation in Multi-Tenant Edge Computing

In multi-tenant edge computing, several Service Providers (SPs) share a limited cache. Because the traffic demand of each SP changes over time, a fixed cache partition can leave space unused for one SP while another SP suffers cache misses and longer response delay. This paper proposes a dynamic cache allocation framework that monitors the traffic demand and cache usage of every SP and moves unused cache space to the SPs that need it. Allocation decisions are supported by model-based reinforcement learning (MB-RL) built on Q-learning. A regression model estimates cost or performance from system conditions, experience replay reuses past experiences during learning, and an epsilon-greedy policy balances trying new actions with choosing learned ones. Every candidate action is checked against two hard rules before it is applied: the total allocation must not exceed the cache capacity, and every SP must keep a minimum cache share. Fair sharing is therefore enforced by construction rather than left to the learning signal. The paper presents the system architecture, the allocation procedure, a worked example with a 100 GB cache, and a planned evaluation based on cache utilization, response delay, cache hit ratio, and fairness. The framework is expected to improve cache usage, reduce unnecessary cache misses and response delay, and support fair sharing among tenants compared with fixed allocation.