Scholarlink

Coefficient optimization is critical for enhancing the performance of Sigma-Delta Modulator(SDMs), particularly in high-resolution analog-to-digital conversion. This paper proposes a reinforcement learning-based design framework using the Soft Actor-Critic (SAC) algorithm to optimize the key coefficients in discrete-time SDMs. The method autonomously adjusts feedforward, feedback, and integrator parameters to maximize signal-to-noise ratio (SNR) while ensuring loop stability. To validate the approach, a third-order single-bit CIFF-structured SDM was implemented and co-simulated using the Ngspice environment. Compared with conventional linearized model-based optimization, the proposed method achieves higher SNR and better stability margins with significantly reduced manual intervention. This approach enhances design efficiency, minimizes reliance on empirical parameter tuning, and demonstrates strong potential for integration into intelligent analog-digital co-design workflows.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Discrete-Time Sigma-Delta Modulator Coefficient Optimization via Soft Actor-Critic Reinforcement Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wanli Zhao,
  • Junpeng Zhang,
  • Zhenyu Rao,
  • Chen Yang,
  • Minglei Tong,
  • Yongqing Sun

摘要

Coefficient optimization is critical for enhancing the performance of Sigma-Delta Modulator(SDMs), particularly in high-resolution analog-to-digital conversion. This paper proposes a reinforcement learning-based design framework using the Soft Actor-Critic (SAC) algorithm to optimize the key coefficients in discrete-time SDMs. The method autonomously adjusts feedforward, feedback, and integrator parameters to maximize signal-to-noise ratio (SNR) while ensuring loop stability. To validate the approach, a third-order single-bit CIFF-structured SDM was implemented and co-simulated using the Ngspice environment. Compared with conventional linearized model-based optimization, the proposed method achieves higher SNR and better stability margins with significantly reduced manual intervention. This approach enhances design efficiency, minimizes reliance on empirical parameter tuning, and demonstrates strong potential for integration into intelligent analog-digital co-design workflows.