Abstract
Open Radio Access Networks (Open RANs) offer a flexible and interoperable wireless network infrastructure to address growing demands for scalability and resource optimization. Current traffic steering and resource allocation approaches use either static policies or machine learning (ML) workflows, each with significant drawbacks-static policies lack adaptability, while ML-based methods can be computationally intensive and inefficient under certain load conditions. To overcome these issues, we propose a hybrid traffic steering mechanism within the Open RAN framework, integrating an enhanced rule-based policy with a global-local reinforcement learning (RL) algorithm. The enhanced rule-based policy is triggered for efficient resource allocation under low network load, encompassing a broader range of User Equipment (UE) slices for finer-grained control. Under high-load variability, the system transitions to the RL-based strategy that learns optimal allocation policies in real-time, factoring in slice types and base station capacities. Evaluation results in a city-scale scenario demonstrate that the proposed adaptive approach significantly improves user satisfaction, reduces unmanaged UEs, and balances cell utilization across varying traffic conditions compared to existing schemes. These findings underscore the potential of combining rule-based policies with reinforcement learning workflows to advance the efficiency and adaptability of Open RANs.