A Communication Channel Density Estimating Generative ...
Transcript of A Communication Channel Density Estimating Generative ...
![Page 1: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/1.jpg)
A Communication Channel Density Estimating Generative Adversarial NetworkAaron Smith and Joseph Downey
NASA Glenn Research Center
Cleveland, Ohio
![Page 2: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/2.jpg)
2
Autoencoder-Based Communications Systems (ABCs)
Background:
• Recent advances in communication system design have applied deep learning to optimize the physical layer for arbitrary channels
• Promising approach to optimizing performance over channels with difficult analytic solutions
Motivation:
• Simplify the traditional communication system, reduce dependency on channel models
• Adapt to a changing environment, and optimize over the end-to-end system
• Generalize over hardware, medium, and waveform
![Page 3: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/3.jpg)
3
Approximating Channel Gradients
𝑠, Ƹ𝑠 − messages (bit sequences)𝑥, 𝑦 − complex baseband symbols
Transmitter Receiver
Channel𝑠𝑥 𝑦
Ƹ𝑠
𝛻𝑙(𝑠, Ƹ𝑠)
Channel Model
𝛻𝑙(𝑠, Ƹ𝑠)
• Backpropagation calculates loss gradients to update weights and biases
• Calculating transmitter updates requires a known channel function
• Ideally, an ABCs would optimize any channel
• NN approximation of a channel provides missing channel gradients
![Page 4: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/4.jpg)
4
Training and Evaluating a Model
Qualities of a Good Training Algorithm• Accurate
> Statistical distance
> Qualitative analysis
• Robust
> Input distribution 𝑝(𝑥)
> ‘Difficult’ channels
• Stable
> Converges to a solution
> Repeatable results
Black-box training
Untrainedmodel
ChannelMeasurements
Trainedmodel
Training Algorithm
(GAN)
Approximates a distribution
𝑝(𝑥)
𝑝(𝑥)
𝑝 𝑦
𝑝 𝑦
?
![Page 5: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/5.jpg)
5
Training Demonstration5M symbols / Random 64-QAM Input
Channel Response Neural Network Response Binned Difference
![Page 6: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/6.jpg)
6
Channel Simulations
• “Channels” include all processes between modulation/demodulation
• Non-linear amplification and pulse shaping cause ISI
• Dispersive channels cause ISI
• Equalization more difficult in non-linear channels
• Non-Gaussian noise
Power Amplifier
MultipathBand-limiting
FilterPhase Noise
![Page 7: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/7.jpg)
7
GAN Architecture
1. Discriminator learns to classify real/fake channel responses
2. Stabilize training with VAE cycle (y-z-y)3. Discourage mode collapse with latent regression (z-y-z)
1
2
3
![Page 8: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/8.jpg)
8
Results
The GAN learned• Arbitrary Inputs
> Amplifier amplitude/phase characteristics
> ISI distortions for high energy symbols
> Distortions specific to constellation
• Channel Parameters
> Distortions specific to the RRC roll-off factor
> Both high/low AWGN powers
Input Power
Input Constellation
Roll-off Factors
AWGN Power
Trained Model
![Page 9: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/9.jpg)
9
Results Cont.Dispersion
Multipath Phase Noise
The GAN learned• ISI due to group delay variation
• ISI from multipath model
• Non-Gaussian noise process
![Page 10: A Communication Channel Density Estimating Generative ...](https://reader031.fdocuments.in/reader031/viewer/2022012300/61e0c20fb09daa435a49bbeb/html5/thumbnails/10.jpg)
10
Conclusions
ABCs provide an end-to-end optimization, but gradients cannot be calculated by the transmitter. This can be circumvented by approximating the channel with a neural network. We provide a new GAN architecture for this purpose. We demonstrate the utility of our architecture by evaluating the GAN on channels that contain non-linearities, intersymbol interference, and non-Gaussian statistics.
Next steps include
• ABCs utilization of GAN architecture
• Lab demonstration