Source-linked AI summary
Physical Layer Communications System Design Over-the-Air Using Adversarial Networks
Timothy J. O'Shea, Tamoghna Roy, Nathan West, Benjamin C. Hilburn
TL;DR
Real-world channel impairments can be difficult to capture with closed-form analytic models, limiting physical-layer learning over such channels. This paper jointly learns a channel approximation and communication encoder-decoder, yielding an over-the-air non-standard 16-QAM mode with a symbol error rate around 0.00714.
Problem
Closed-form models can be difficult to construct for real-world channel impairments and may fail to express complexities such as device-to-device hardware variation.
Source: §I · p. 1Method
An adversarial system jointly learns a channel-function approximation and an encoder-decoder scheme using physical channel measurements and a learned channel network.
Source: §II · p. 2Results
The learned non-standard 16-QAM mode achieved a symbol error rate of around 0.00714 over the air.
Source: §III · p. 3Takeaways & Limitations
The learned channel-function approximation is sufficient to back-propagate and adapt both encoder and decoder networks on an unknown, uncharacterized physical channel.
Source: §IV · pp. 3–4Further work is needed to analyze performance, train on information streams, evaluate adaptation under changing conditions, and scale to larger symbol blocks.
Source: §IV · p. 4
Abstract
from arXiv · showhide
This paper presents a novel method for synthesizing new physical layer modulation and coding schemes for communications systems using a learning-based approach which does not require an analytic model of the impairments in the channel. It extends prior work published on the channel autoencoder to consider the case where the channel response is not known or can not be easily modeled in a closed form analytic expression. By adopting an adversarial approach for channel response approximation and information encoding, we can jointly learn a good solution to both tasks over a wide range of channel environments. We describe the operation of the proposed adversarial system, share results for its training and validation over-the-air, and discuss implications and future work in the area.
I. INTRODUCTION
Channel autoencoders learn physical-layer encoding schemes by jointly optimizing encoder and decoder networks, but real-world impairments can be difficult to model accurately in a differentiable form. The paper motivates an adversarial approach to learning over physical channels whose responses lack a closed-form model.
I. INTRODUCTION
- Jointly optimized deep encoder and decoder networks have learned new modulation schemes with inherent error-correction capability for common channel models,.Figure 1 depicts the channel autoencoder optimized for a differentiable analytic channel model.Source: §I · p. 1, Figure 1 · p. 1
- Directly learning complex modulation and demodulation functions can optimize global metrics such as symbol error rate, but real-world optimization requires an accurate differentiable channel expression.The modeled effects include digital conversion, analog RF hardware, and other distortions and impairments.Source: §I · p. 1
- Many systems lack closed-form models that capture real-world complexity, and receiver-only updates after synthetic pretraining cannot update the transmitter without a channel gradient.The paper instead frames the problem as adversarial function approximation; the next section describes a hardware-in-the-loop system for adapting transmitter and receiver over the air.Source: §I · p. 1
II. TECHNICAL APPROACH
CommGAN jointly approximates an arbitrary communications channel response and learns an encoding and decoding scheme. It replaces an analytic channel model with physical measurements and a learned channel approximation during training.
II. TECHNICAL APPROACH
- Figure 2’s CommGAN uses neural encoder and decoder networks, while representing a black-box channel with physical and learned response functions rather than an analytic model.The encoder maps bits or codeword indices to transmit samples; the decoder maps received samples to likelihoods.Source: Figure 2 · p. 2, §II · p. 2
- CommGAN jointly optimizes channel-response approximation and information encoding and decoding for a performance metric such as low symbol error rate.Its adversarial objectives are trained together or iteratively, with each informing and improving the other.Source: §II · p. 2
- The physical response h0 measures the communications system, while neural h1 approximates it; training uses channel-approximation MSE and autoencoder cross-entropy losses.The losses are identified as L0 and L1, respectively.Source: §II · p. 2
- Training cycles between objectives with manually tuned learning rates and small fully connected ReLU networks; Figure 3 plots the channel-approximation and autoencoder loss curves.The channel-approximation loss and symbol cross-entropy are the two training quantities described alongside the curves.Source: §II · p. 2, Figure 3 · p. 3
- The physical channel uses a USRP B210 in a 900 MHz indoor setup, with shared sample clocks but uncalibrated hardware offsets and impairments.The system maps information to 16 symbols, transmits three samples per symbol, and considers one receive sample within the unaligned three-sample window.Source: §II · p. 3
III. RESULTS
The channel-approximation and autoencoder losses reinforce each other during training, culminating in an over-the-air non-standard 16-QAM result with a symbol error rate around 0.00714.
III. RESULTS
- Normalization and noise insertion aid training, while an average-power constraint mitigates gradients that would continually increase transmit power.Source: §III · p. 3
- The channel-approximation and autoencoder losses reinforce one another as training epochs increase, and training finishes in under five minutes.Training used an NVIDIA Titan V GPU.Source: §III · p. 3
- Around 0.00714 symbol error rate was achieved over the air by encoding 4 bits into one receive symbol from 3 unaligned transmit symbols.Figure 4 shows the learned scheme’s I/Q samples.Source: Figure 4 · p. 3, §III · p. 3
IV. CONCLUSION
The adversarial approach jointly learns a channel function approximation and encoder/decoder for effective communication without a closed-form channel model. The authors identify substantial work needed to mature the approach.
IV. CONCLUSION
- Jointly learning a channel approximation and encoder/decoder can produce effective systems for arbitrary communications channels without a closed-form channel model or implementation.Source: §IV · p. 3
- The learned channel function approximation is sufficient to back-propagate and adapt both encoder and decoder networks.Source: §IV · p. 4
- Future work includes stream training, online adaptation under changing environmental conditions, and scaling to modern LTE-sized symbol blocks.Source: §IV · p. 4
- The paper frames holistic optimization across many effects and transforms as a fundamental requirement for the next major communications-performance leap.Source: §IV · p. 4