Local and Remote Digital Pre-Distortion for 5G Power Amplifiers with Safe Deep Reinforcement Learning

Christian Spano, Damiano Badini, Lorenzo Cazzella, Matteo MatteucciSensors 2025 · DOI

TL;DR

This paper introduces DRL-DPD, a novel approach using Deep Reinforcement Learning (DRL) to tackle the complexity and resource limitations of Digital Pre-Distortion (DPD) for 5G power amplifiers. By employing a safe DRL algorithm (CRE-DDPG), it ensures regulatory compliance and device safety while achieving efficient linearization in both local and remote configurations.

What Problem Does This Paper Address?

Modern wireless communication systems, especially 5G, demand higher data rates and energy efficiency. This pushes power amplifiers (PAs) to operate closer to their limits, leading to nonlinear distortions that degrade signal quality and reduce efficiency. Digital Pre-Distortion (DPD) is a crucial technique to counteract these nonlinearities. However, traditional DPD methods often require complex, power-hungry hardware, especially in the feedback path for coefficient updates. This is particularly challenging in 5G's high-bandwidth scenarios (like FR2) and in remote DPD setups where the PA and DPD controller are physically separated, relying on over-the-air feedback. Existing DPD solutions struggle with complexity, adaptability to changing conditions, and resource constraints.

Methods

The core of this work is DRL-DPD, a system that leverages Deep Reinforcement Learning (DRL) to adaptively control DPD coefficients. Instead of relying on complex wideband feedback, DRL-DPD uses a simplified, scalar feedback signal: Adjacent Channel Leakage Ratio (ACLR) for local DPD and Signal-to-Noise Ratio (SNR) for remote DPD. These metrics are used by a DRL agent to learn optimal coefficient updates.

A key innovation is the CRE-DDPG (Cautious-Recoverable-Exploration Deep Deterministic Policy Gradient) algorithm. This "Safe RL" approach enhances the standard DDPG by incorporating a mechanism to prevent the agent from making exploration decisions that could lead to unacceptable ACLR violations, thus ensuring regulatory compliance and protecting the PA. CRE-DDPG modulates the agent's actions based on the current reward and the magnitude of the action, employing an attenuation factor and a recovery strategy if the safety threshold is approached.

The DRL agent is implemented using an actor-critic architecture, with the actor determining actions (coefficient updates) and the critic evaluating them. The paper details the state (DPD coefficients), action (coefficient updates), and reward definitions for both local and remote DPD scenarios.

Datasets

Not discussed in the paper. The paper focuses on simulated environments and a real hardware setup, rather than pre-existing datasets.

Key Insights and Ideas

  • Scalar Feedback is Sufficient: The paper demonstrates that simplified scalar feedback (ACLR or SNR) is enough for DRL to effectively learn and optimize DPD coefficients, significantly reducing hardware complexity and power consumption compared to traditional wideband feedback.
  • DRL for Adaptive DPD: DRL's ability to learn in dynamic environments makes it a natural fit for DPD, allowing adaptation to changing PA behavior (e.g., due to temperature drift or aging) without requiring explicit system models.
  • Safety First with CRE-DDPG: The introduction of CRE-DDPG is crucial for practical deployment. It ensures that the DRL agent's exploration doesn't violate safety and regulatory standards (like ACLR limits), a critical concern for real-world hardware.
  • Local vs. Remote DPD Synergy: The DRL-DPD framework is shown to be effective in both local (co-located PA and DPD) and remote (distributed) configurations, highlighting its versatility. Remote DPD leverages the routinely available SNR signal, making it a practical solution.

Key Results

  • Simulation Results: DRL-DPD, particularly with the "Noisy-Nets" variant, showed significant improvements in ACLR over time in simulated environments, converging to optimal coefficients.
  • Hardware Experiments: On a real 1W GaAs PA, CRE-DDPG demonstrated its ability to linearize the PA effectively. The agent consistently operated within the safe region, improving ACLR after approximately 850 coefficient updates.
  • Reliability: CRE-DDPG showed resilience to environmental variations, consistently maintaining performance and staying within the safe region across multiple simulations. It achieved ACLR improvements within 950 updates in 25 tested simulations.
  • Complexity Reduction: The proposed DRL-DPD architecture significantly reduces hardware complexity and power consumption by eliminating the need for expensive, high-bandwidth ADCs in the feedback path.

Limitations and Future Work

  • Hardware Experiments: While hardware experiments were conducted for local DPD, further extensive experiments are needed for diverse hardware implementations and challenging conditions (e.g., memory effects, high temperatures).
  • Model Compression: Future work could explore model compression techniques (pruning, knowledge distillation) to further optimize the DRL agent's architectures for resource-constrained platforms.
  • Remote DPD in Real-World Scenarios: Evaluating DRL-DPD in practical remote DPD scenarios is highlighted as an important area for future research, especially considering time constraints limited this study to simulations.
  • Alternative Feedback Mechanisms: The paper suggests exploring alternative feedback approaches, such as direct comparison of input and output signals to estimate an error signal, which could offer different implementation trade-offs.
  • Channel Conditions: Analyzing the impact of radio channel propagation conditions, particularly in mobility scenarios, on remote DPD estimation is a key future direction.

Summary updated

Abstract

The demand for higher data rates and energy efficiency in wireless communication systems drives power amplifiers (PAs) into nonlinear operation, causing signal distortions that hinder performance. Digital Pre-Distortion (DPD) addresses these distortions, but existing systems face challenges with complexity, adaptability, and resource limitations. This paper introduces DRL-DPD, a Deep Reinforcement Learning-based solution for DPD that aims to reduce computational burden, improve adaptation to dynamic environments, and minimize resource consumption. To ensure safety and regulatory compliance, we integrate an ad-hoc Safe Reinforcement Learning algorithm, CRE-DDPG (Cautious-Recoverable-Exploration Deep Deterministic Policy Gradient), which prevents ACLR measurements from falling below safety thresholds. Simulations and hardware experiments demonstrate the potential of DRL-DPD with CRE-DDPG to surpass current DPD limitations in both local and remote configurations, paving the way for more efficient communication systems, especially in the context of 5G and beyond.
BibTeX
@inproceedings{spano2025local,
  author = {Christian Spano and Damiano Badini and Lorenzo Cazzella and Matteo Matteucci},
  title = {{Local and Remote Digital Pre-Distortion for 5G Power Amplifiers with Safe Deep Reinforcement Learning}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25196102},
}
Figure