VP-PPO: Real-Time Learning-Based Vector Perturbation Precoding for O-RAN Cell-Free Massive MIMO Networks

Abstract

Cell-Free (CF) massive Multiple-Input Multiple Output (MIMO) networks are a key enabler for future sixth generation wireless systems, as they allow distributed Radio Units (RUs) to cooperatively serve users and mitigate inter-user interference. However, under dense deployments and high spatial loading, conventional linear precoding schemes such as Maximum Ratio Transmission (MRT), Zero Forcing (ZF), and Minimum Mean-Square Error (MMSE) suffer from channel ill-conditioning, spatial correlation, and noise amplification, which limit their achievable spectral efficiency (SE). Non-linear Vector Perturbation (VP) precoding can overcome these limitations by adding an integer perturbation vector before transmission, thereby reducing transmit energy while preserving the original information symbols through a modulo operation at the receiver. Despite its strong performance, optimal VP precoding requires an exhaustive integer search whose complexity grows exponentially with the number of users, making it unsuitable for real-time Open Radio Access Network (O-RAN) operation. This work proposes VP-Proximal Policy Optimization (PPO) , a real-time learning-based non-linear precoding framework for O-RAN-compliant CF massive MIMO networks. The proposed method is implemented as an intelligent xApp within the near-real time RAN Intelligent Controller (near-RT RIC), where it replaces the exhaustive VP perturbation search with a single neural network inference. A hybrid PPO agent is designed to jointly determine the discrete perturbation vector and the continuous transmit power level. The state representation includes the MMSE precoding matrix, transmitted symbols, effective channel informa tion, signal and interference features, channel condition number, and spatial loading ratio. The hybrid action space combines categorical distributions for user-wise perturbation selection with a continuous power-control head. The reward function jointly maximizes spectral efficiency, reduces transmit power, and min imizes the number of occupied Resource Blocks (RBs), enabling an efficient trade-off between throughput, energy consumption, and radio-resource utilization. Simulation results using 3GPP CDL-D channel conditions demonstrate that VP-PPO provides a practical balance between non-linear precoding performance and real-time feasibility. For nine users, VP-PPO achieves 28.69 bits/s/Hz, corresponding to a 53% spectral-efficiency improvement over MMSE precoding, while reducing transmit power by approximately 8% compared with the fixed 1 W power budget used by the baseline methods. In addition, VP-PPO reduces RB consumption compared with linear precoders and closely approaches the exhaustive VP oracle. Most importantly, the proposed method achieves millisecond scale inference latency, satisfying the 10–100 ms near-RT RIC control-loop requirement, whereas exhaustive VP search becomes computationally infeasible. These results show that reinforcement learning-assisted xApps can enable practical deployment of ad vanced physical-layer techniques in future O-RAN CF networks.