SLAQ: An SLA-Driven 6G O-RAN QoS Framework Using Deep Reinforcement Learning

As 6G scenarios grow in complexity, network operators need an automated and optimized approach to managing Quality-of-Service (QoS) in the radio access network (RAN). Transitioning from resource-focused management to one that considers evolving operator intents is essential. This can be achieved by translating service-level agreement (SLA) intents into operational rules to ensure compliance and save costs. O-RAN introduced the RAN intelligent controller (RIC) and intelligent applications (xApps) to enhance the RAN operation. Furthermore, RAN slicing has been shown as the most promising method to provide O-RAN-based QoS in 6G. However, while existing methods optimize resources allocated per slice, they often overlook SLAs. This leads to suboptimal resource usage and outages when the system fails to meet target key performance indicators (KPIs). This work introduces SLAQ, an xApp designed to translate service operator requirements to optimize spectrum resource sharing. SLAQ monitors key performance metrics (KPMs) and employs deep reinforcement learning (DRL) to make decisions within the RAN. SLA policies are modeled and integrated into the xApp to optimize resource distribution while preventing SLA conflicts and outages. Our highly-detailed system-level simulations show that SLAQ effectively learns optimal actions, achieving high communication reliability, i.e., close to $99 \%$ for ultra-reliable low-latency communications.