Reinforcement Learning-Enabled Agent for Transmitter Optimization in Digital-Analog Radio-over-Fiber Fronthaul

Technical Summary: Reinforcement Learning-Enabled Agent for Transmitter Optimization in Digital-Analog Radio-over-Fiber Fronthaul

Problem Statement
Digital-analog radio-over-fiber (DA-RoF) has emerged as a viable fronthaul solution for 6G and advanced Cloud-RAN architectures, offering a balance between the spectral efficiency of analog transmission and the robustness of digital transmission. However, the performance of DA-RoF systems is critically dependent on the joint optimization of several tightly coupled transmitter parameters: the rounding factor (RF), scaling factor (SF), geometric shaping (GS) factor, and pre-equalization tap coefficients. These parameters collectively influence quantization noise, peak-to-average power ratio (PAPR), nonlinear distortion, and bandwidth-induced inter-symbol interference (ISI).

Conventional optimization methods, such as fixed configuration, empirical tuning, or exhaustive grid search, are computationally prohibitive and often yield suboptimal results under practical hardware impairments and varying link conditions. Furthermore, existing end-to-end deep learning approaches are ill-suited for DA-RoF because the system relies on non-differentiable rounding operations, and deploying complex neural networks at the remote radio units (RRUs) is cost-prohibitive. There is a need for a model-free, hardware-efficient optimization method that can autonomously learn optimal parameters without requiring a differentiable channel model or additional receiver-side complexity.

Methodology
The authors propose a Reinforcement Learning (RL)-enabled DA-RoF fronthaul agent architecture designed to operate entirely at the transmitter side (Centralized Unit/Distributed Unit). The core methodology involves:

1. MDP Formulation: The transmitter parameter optimization problem is formulated as a Markov Decision Process (MDP).

   • State: Defined by the current values of RFs, SFs, the GS factor, and pre-equalization coefficients.
   • Action: Discrete variations (increases or decreases) applied to these parameters.
   • Reward: Derived from the Signal-to-Noise Ratio (SNR) of the recovered wireless signal, calculated via Error Vector Magnitude (EVM) feedback from the receiver. The reward function encourages exceeding modulation-specific SNR thresholds and monotonic improvement.
   • Environment: The physical IMDD (Intensity-Modulation Direct-Detection) DA-RoF link, including impairments like amplifier nonlinearity and bandwidth limitations.

2. Algorithm Selection: The authors employ a Dueling Deep Q-Network (DQN). This architecture separates the estimation of state value and action advantage, which is particularly effective for the coupled, discrete parameter space of DA-RoF where many local actions may yield similar immediate rewards but differ in long-term impact. The agent is model-free, relying solely on end-to-end SNR feedback rather than an analytic channel model.

3. System Architecture: The agent resides at the CU/DU. It interacts with the fronthaul link by adjusting transmitter parameters and receiving SNR feedback from the Remote Units (RUs). Crucially, the RUs require no hardware modifications or additional DSP modules; they simply provide feedback, ensuring low complexity at the network edge.

Key Contributions

• MDP Formulation for DA-RoF: The paper formulates the joint optimization of RF, SF, GS factor, and pre-equalization coefficients as an MDP, addressing the non-separable nature of these parameters.
• Hardware-Friendly RL Agent: The proposed framework deploys the learning agent exclusively at the transmitter, maintaining low receiver complexity and compatibility with existing IMDD fronthaul deployments.
• Model-Free Optimization: The system achieves optimization without requiring a differentiable channel model or exhaustive grid search, utilizing a dueling-DQN to learn optimal policies through trial and error.

Experimental Results
The proposed framework was validated on an experimental IMDD DA-RoF testbed using a 1-km standard single-mode fiber (SSMF) link. Key findings include:

• SNR Improvement: The trained agent achieved an average SNR improvement of approximately 2.7 dB compared to a grid-search baseline.
• Performance Across Orders: The system successfully supported high-order modulation formats, achieving final SNRs of:
  • 35.8 dB for 1-order DA-RoF (1024-QAM)
  • 42.9 dB for 2-order DA-RoF (4096-QAM)
  • 53.8 dB for 3-order DA-RoF (16384-QAM)
  • 63.2 dB for 4-order DA-RoF (65536-QAM)
• Convergence: The agent demonstrated rapid convergence, reaching optimal states within 32 interaction steps, significantly faster than exhaustive search methods which would require tens of thousands of measurements.
• Parameter Adaptation: The agent autonomously learned to adjust RF and SF to balance quantization noise and PAPR, utilized GS to mitigate amplitude nonlinearity, and optimized pre-equalization taps to compensate for bandwidth-induced ISI.
• Robustness: The optimized parameters maintained high performance across varying electrical driving amplitudes and received optical powers, demonstrating strong generalization.

Significance and Claims
The paper claims that the proposed RL-enabled framework provides an online, scalable, and hardware-efficient solution for parameter optimization in DA-RoF fronthaul systems. By eliminating the need for complex receiver-side DSP or exhaustive search, the method addresses the practical constraints of cost-sensitive access networks. The results validate that intelligent, model-free control can effectively manage the complex coupling of transmitter parameters, paving the way for the deployment of high-order modulation formats and intelligent next-generation radio access networks. The authors position this work as a step toward hardware-adaptive fronthaul systems capable of supporting beyond-Terabit capacities and ultra-dense deployments.