A hybrid transformer-zero-shot learning framework with Muon optimization for intelligent channel estimation in MIMO...

Area of Science:

• Wireless Communication
• Machine Learning
• Signal Processing

Background:

• Accurate channel estimation is crucial for MIMO systems.
• Current Deep Learning (DL) methods struggle with generalization and require retraining for new scenarios.
• Existing techniques lack adaptability to unseen Signal-to-Noise Ratio (SNR) levels and fading conditions.

Purpose of the Study:

• To propose a novel hybrid framework for intelligent MIMO channel estimation.
• To enhance generalization capabilities for unknown wireless environments.
• To overcome limitations of traditional and current DL-based estimation techniques.

Main Methods:

• Integration of Transformer architectures for spatial-temporal feature extraction.
• Application of Zero-Shot Learning (ZSL) for inference in unknown conditions without retraining.
• Utilization of the Muon optimizer for improved convergence and generalization over Adam.
• Assessment of three configurations: ZSL-Muon, Transformer-ZSL, and the complete Transformer-ZSL-Muon model.

Main Results:

• The hybrid Transformer-ZSL-Muon model consistently outperforms LS, MMSE, CNN, GRU, and Transformer-only baselines in Mean Squared Error (MSE).
• Significant MSE reduction (94.44%) observed with increased antennas (2 to 128) at 30 dB SNR, demonstrating massive MIMO potential.
• Achieved substantial MSE improvements: ~93.75% over MMSE and ~96.67% over LS at 30 dB SNR.
• Demonstrated superior performance across quasi-static and time-varying fading models and a wide SNR range.

Conclusions:

• The proposed hybrid framework offers a scalable, intelligent, and resilient solution for next-generation MIMO systems.
• Combining semantic-aware inference, attention mechanisms, and adaptive optimization enhances channel estimation robustness.
• The framework effectively handles complex and unseen channel conditions, reducing the need for constant retraining.