Reinforcement Learning-Aided Design of Efficient Polarization Kernels

2025-05-07

AI Summary

This work addresses the challenge of designing large polar code kernels that jointly optimize error exponent performance and decoding complexity—particularly under recursive maximum-likelihood decoding (RMLD), where theoretical optimal bit-error rate is targeted. To this end, we propose PolarZero: the first reinforcement learning framework for polar kernel search that integrates Gumbel-perturbed AlphaZero, enabling fully automated, scalable, and prior-free large-kernel construction. Our method jointly incorporates RMLD-based performance evaluation, kernel complexity modeling, and Gumbel-enhanced policy optimization. Experiments demonstrate that PolarZero exactly recovers exhaustive-search-optimal kernels for small dimensions and discovers a 16×16 kernel achieving near-optimal error exponents—comparable to state-of-the-art handcrafted designs—while matching their decoding complexity. These results validate PolarZero’s effectiveness, scalability, and practical utility for large-kernel polar code design.

Technology Category

• Machine Learning: Kernel Methods
• Search and Optimization: Learning to Search
• Natural Language Processing: Learning & Optimization for NLP

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Incentives in network design for Web infrastructures and ecosystems
• Security and Privacy: Large-scale security measurements

Abstract

Polar codes with large kernels achieve optimal error exponents but are difficult to construct when low decoding complexity is also required. We address this challenge under recursive maximum likelihood decoding (RMLD) using a rein-forcement learning approach based on the Gumbel AlphaZero algorithm. The resulting method, PolarZero, consistently matches exhaustive search in identifying low-complexity kernels, and discovers a size-16 kernel with complexity comparable to handcrafted designs. Our results suggest that PolarZero is a scalable tool for large-kernel design, where brute-force search is no longer feasible.

Problem

• Design efficient polarization kernels for large polar codes
• Achieve low decoding complexity with reinforcement learning
• Discover scalable kernels where brute-force search fails

Innovation

• Reinforcement learning aids polarization kernel design
• Gumbel AlphaZero algorithm enables efficient RMLD
• PolarZero matches exhaustive search for low-complexity kernels