Reinforcement Learning Boosts Network Slicing in 5G-Advanced O-RAN

Open Radio Access Network (O-RAN) architecture makes 5G-Advanced networks more flexible by replacing proprietary hardware with software-based components known as xApps and rApps. Yet this flexibility brings a challenge: rapidly changing channel conditions make beam selection and radio resource allocation difficult. A new study proposes a reinforcement learning framework that tackles both problems at once, running across the Near-Real-Time and Non-Real-Time RIC control platforms that manage O-RAN networks.

The framework addresses two tasks. The second allocates bandwidth among three service slices: enhanced Mobile Broadband (eMBB) for high-throughput traffic, Ultra-Reliable Low Latency Communications (URLLC) for time-critical applications, and massive Machine Type Communications (mMTC) for huge numbers of connected devices. Minimum rate requirements for each slice are factored in, and a multi-objective reward function guides the learning process.

The researchers developed two tabular Q-learning xApps, one for each task. The beam-selection xApp operates over 864 states, while the slicing xApp uses only 9 states. This difference in state-space size turned out to be decisive for performance.

The method was evaluated through 30 independent Monte Carlo trials, with results reported as 95% confidence intervals. Comparisons were made against greedy and random baseline approaches, and an ablation study isolated the contribution of each xApp.

The slicing xApp performed well, achieving an average throughput of 1709.2 Mbps compared with 1639.4 Mbps for the greedy method. The beam-selection xApp, in contrast, showed only limited improvement, reaching 700.4 Mbps against 680.4 Mbps for random selection.

The study also examined a handover hysteresis mechanism designed to reduce unnecessary handovers between beams. This mechanism cut handover events by 23.8%, but aggregate throughput decreased by 10.5%, illustrating a clear trade-off between connection stability and overall network performance.

The overall conclusion is nuanced rather than uniformly positive. Tabular reinforcement learning appears suitable for network slicing, where the state space is small and manageable. For beam selection, channel-aware greedy methods may be more practical in current deployments. The authors suggest that deep reinforcement learning could be explored in future work to handle more complex beam-selection problems that exceed the capacity of tabular methods.

The article, an open-access research paper published in Discover Applied Sciences on 28 October 2026 by Amar Choudhary, Geetika Srivastava, Bandana Mallick and Pavithra P, is shared as a peer-reviewed accepted manuscript that remains subject to final editorial edits before the Version of Record. Its practical takeaway for network operators is that lightweight, table-based learning agents can already deliver measurable gains in slice-level resource allocation, while the harder problem of adaptive beam management in fast-changing channels likely awaits more expressive learning architectures.

Subject of Research: Reinforcement learning for joint beam selection and network slicing in 5G-advanced O-RAN

Article Title: Reinforcement learning-based joint beam selection and network slicing for 5G-advanced O-RAN

Article References: Choudhary, A., Srivastava, G., Mallick, B., & P, P. (2026). Reinforcement learning-based joint beam selection and network slicing for 5G-advanced O-RAN. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09813-5

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09813-5

Keywords: O-RAN, 5G-Advanced, reinforcement learning, network slicing, beam selection, xApps, Q-learning, URLLC, eMBB, mMTC, throughput, RIC