Accelerating Resource Allocation in Open RAN Slicing via Deep Reinforcement Learning

The transition to beyond-fifth-generation (B5G) wireless systems has revolutionized cellular networks, driving unprecedented demand for high-bandwidth, ultra low-latency, and massive connectivity services. The open radio access network (Open RAN) and network slicing provide B5G with greater flexibility and efficiency by enabling tailored virtual networks on shared infrastructure. However, managing resource allocation in these frameworks has become increasingly complex. This paper addresses the challenge of optimizing resource allocation across virtual network functions (VNFs) and network slices, aiming to maximize the total reward for admitted slices while minimizing associated costs. By adhering to the Open RAN architecture, we decompose the formulated problem into two subproblems solved at different timescales. Initially, the successive convex approximation (SCA) method is employed to achieve at least a locally optimal solution. To handle the high complexity of binary variables and adapt to time-varying network conditions, traffic patterns, and service demands, we propose a deep reinforcement learning (DRL) approach for real-time and autonomous optimization of resource allocation. Extensive simulations demonstrate that the DRL framework quickly adapts to evolving network environments, significantly improving slicing performance. The results highlight DRL’s potential to enhance resource allocation in future wireless networks, paving the way for smarter, self-optimizing systems capable of meeting the diverse requirements of modern communication services.