6G • RAN
Learned Constellation
Deep Learning Autoencoder Constellation Shaping
Non-uniform signal constellations shaped by deep autoencoders to match physical transceiver hardware imperfections and non-linear power amplifiers.
Technical Explanation
Standard cellular networks transmit square grid constellations (like 64-QAM or 256-QAM). Real-world radio hardware suffers from non-linear power amplifier compression and phase noise that distorts the outer constellation points. Learned Constellation Shaping uses autoencoders to design non-uniform, circularly symmetric constellations tailored to the specific amplifier profile of the transmitter, maximizing data throughput without causing non-linear distortion.
Key Functions
- Autoencoder-designed non-uniform signal constellations outperforming standard square QAM
- Optimal resilience against power amplifier non-linearities and saturation compression
- Lower peak-to-average power ratio (PAPR) reducing battery drain in mobile transmitters
- Tailored geometric and probabilistic shaping approaching theoretical Shannon capacity limits
- Dynamic constellation adaptation based on battery level and instantaneous RF power
Specifications
IEEE Transactions on Information Theory, 3GPP Rel-20 Study on New Waveforms
Interfaces
Learned-Constellation-PHYAutoencoder-Mod
Related 6G Concepts
Want to memorize 6G concepts like this one?
Study it with SuperMemo SM-2 spaced repetition flashcards.