ADAPTIVE REINFORCEMENT LEARNING FOR DYNAMIC RESOURCE ALLOCATION: MINIMISING COST AND MAXIMISING SLA COMPLIANCE

Abstract

In modern cloud and edge computing infrastructures, service providers face the dual challenge of minimising operational cost while ensuring that service‐level agreements (SLAs) are met. Static resource allocation strategies often lead either to over‐provisioning (and hence wasted cost) or under‐provisioning (and thus SLA violations). This paper proposes an adaptive reinforcement‐learning (RL) based resource‐allocation model that dynamically adjusts resource provisioning and workload distribution in response to real‐time system state and workload dynamics. The RL agent observes system metrics (e.g., resource utilisation, workload queue length, SLA violation severity) and selects allocation actions (scaling resources up/down, migrating workloads, adjusting task placement) so as to optimise a composite reward reflecting cost savings and SLA adherence. Experimental results on a simulated cloud‐environment benchmark show that our model reduces cost by X % while improving SLA-compliance by Y % relative to baseline heuristics, thus validating the approach. We discuss the design of the state‐action space, reward shaping, training methodology, and potential deployment issues.