AutoML and Deep Unfolding Optimize Wireless Beamforming
A recent research initiative has successfully integrated automated machine learning (AutoML) techniques with advanced deep unfolding strategies to enhance wireless communication processes, specifically beamforming and waveform optimization. Utilizing a modified proximal gradient descent (PGD) algorithm, the study reformulated it into a deep neural network to facilitate parameter learning at varying layers. This approach was enriched by a hybrid layer that enables a trainable linear gradient transformation prior to the proximal projection phase. The novel auto-unrolled PGD model achieves an impressive 98.8% of the spectral efficiency of traditional 200-iteration PGD methods with just five unrolled layers. For more details, refer to arXiv identifier 2603.17478.
Key facts
- Combines AutoML with model-based deep unfolding
- Converts iterative PGD into a deep neural network
- Hybrid layer with learnable linear gradient transformation
- Uses AutoGluon and TPE for hyperparameter optimization
- Search space includes network depth, step-size, optimizer, scheduler, layer type, activation
- Auto-PGD achieves 98.8% spectral efficiency of 200-iteration PGD
- Only five unrolled layers used
- Published on arXiv:2603.17478
Entities
Institutions
- arXiv