PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning based End-to-End Wireless Communication Systems with Conditional GAN as Unknown Channel

Hao Ye, Le Liang, Geoffrey Ye Li, Biing-Hwang Fred Juang

arXiv:1903.02551v1cs.IT

TL;DR

The paper addresses end-to-end wireless learning when instantaneous CSI is difficult to obtain and channel uncertainty blocks transmitter training. It models channel effects with a conditional GAN that uses pilot information and bridges transmitter–receiver gradients, while CNNs address long sequences. Simulations report effectiveness on AWGN, Rayleigh fading, and frequency-selective channels.

  • Problem

    Accurate instantaneous CSI is difficult to obtain because wireless channels vary with time and location, hindering transmitter-DNN training.

  • Method

    A conditional GAN models channel effects using received pilot information and bridges transmitter and receiver DNNs for gradient back-propagation; convolutional layers address long sequences.

  • Results

    Simulations report effectiveness on AWGN, Rayleigh fading, and frequency-selective channels.

  • Takeaways & Limitations

    The framework supports data-driven end-to-end communication across a wide range of wireless channels without requiring an explicitly known channel transfer function.

Abstract

from arXiv · show

In this article, we develop an end-to-end wireless communication system using deep neural networks (DNNs), in which DNNs are employed to perform several key functions, including encoding, decoding, modulation, and demodulation. However, an accurate estimation of instantaneous channel transfer function, i.e., channel state information (CSI), is needed in order for the transmitter DNN to learn to optimize the receiver gain in decoding. This is very much a challenge since CSI varies with time and location in wireless communications and is hard to obtain when designing transceivers. We propose to use a conditional generative adversarial net (GAN) to represent channel effects and to bridge the transmitter DNN and the receiver DNN so that the gradient of the transmitter DNN can be back-propagated from the receiver DNN. In particular, a conditional GAN is employed to model the channel effects in a data-driven way, where the received signal corresponding to the pilot symbols is added as a part of the conditioning information of the GAN. To address the curse of dimensionality when the transmit symbol sequence is long, convolutional layers are utilized. From the simulation results, the proposed method is effective on additive white Gaussian noise (AWGN) channels, Rayleigh fading channels, and frequency-selective channels, which opens a new door for building data-driven DNNs for end-to-end communication systems.

I. INTRODUCTION

End-to-end wireless communication replaces conventional transmitter and receiver blocks with jointly trained DNNs, but unknown, time-varying channels obstruct gradient propagation and long blocks create dimensionality challenges. The proposed conditional-GAN channel model uses pilot information and CNNs to support channel-agnostic learning across practical wireless channels.

  • End-to-End Paradigm: End-to-end communication replaces encoding, modulation, decoding, and demodulation modules with transmitter and receiver DNNs trained for recovery accuracy.The transmitter maps symbols to encoded signals, while the receiver recovers symbols from the received signal.
  • Challenges: Unknown channel transfer functions block gradients from the receiver DNN to the transmitter, while assumed models can bias learned weights relative to actual channels.Instantaneous CSI is difficult to obtain because wireless channels are noisy, time-varying, and analytically uncertain.
  • Challenges: Nearly 90% of codewords in training can still leave decoding performance poor on unseen codewords, motivating methods that handle longer practical blocks.Prior end-to-end work therefore concentrated on small block sizes such as Hamming codes (7,4).
  • Proposed Approach: The proposed channel-agnostic system learns channel-output distributions with a conditional GAN conditioned on encoded signals and received pilot information.The GAN bridges transmitter and receiver training so end-to-end loss gradients can reach the transmitter.
  • Results and Scope: CNNs mitigate dimensionality challenges, extending the block length from several bits to a couple of hundred bits and supporting frequency-selective channels with inter-symbol interference.The reported system is applicable to AWGN, Rayleigh fading, and frequency-selective channels.

II. RELATED WORKS

The proposed method is related to GANs, end-to-end learning-based communication systems, and learning-based decoders.

  • The work connects GAN-based modeling with end-to-end learning-based communication systems.
  • It is also related to approaches that use learning-based decoders.

A. GANs and Conditional GANs

GANs use competing generator and discriminator networks to produce samples resembling real data, while conditional GANs add context to control generated samples. In wireless communications, GANs can model channel effects, and the proposed conditional approach uses pilot information to address time-varying channels.

  • GAN training uses a generator and discriminator that compete, with discriminator feedback improving generated samples.
  • Conditional GANs add context information to the generator and discriminator to generate samples with a specific property.
  • Originally, conditional GANs used label information so generators could produce samples from particular categories.
  • In computer vision, GANs have been used for tasks including generating high-resolution images from low-resolution images.
  • GANs have been applied beyond computer vision to model additive white Gaussian noise channel effects.
  • The proposed conditional GAN extends channel modeling to more realistic time-varying channels by conditioning generation on received pilot information.

B. DNN based End-to-End Communications

Prior end-to-end communication methods learn transmitter and receiver DNNs but face challenges from unknown channels and long block lengths. The proposed approach uses a conditional GAN as a differentiable channel bridge and CNN-based joint encoding and decoding, addressing channel modeling and dimensionality challenges.

  • DNN based End-to-End Communications: End-to-end communication systems jointly learn transmitter and receiver DNNs, with receiver recovery performance guiding transmitter training.
  • DNN based End-to-End Communications: Earlier model-free methods used reinforcement learning or stochastic perturbations, but focused on small block lengths.
  • DNN based End-to-End Communications: Extending end-to-end learning to large block sizes and modeling unknown channels with data-driven methods remained open problems addressed by this approach.
  • DNN based End-to-End Communications: Related work includes neural decoding of convolutional, turbo, linear, and short Polar codes, with short Polar-code decoding reported as similar to maximum likelihood decoding.
  • DNN based End-to-End Communications: CNN-based simultaneous learning of encoding and decoding can mitigate the curse of dimensionality compared with learning only to decode human-designed codewords.
  • DNN based End-to-End Communications: Unknown CSI blocks back-propagation needed to train the overall end-to-end system.
  • DNN based End-to-End Communications: A conditional GAN learns channel effects and bridges transmitter and receiver so gradients can pass through while modeling channel outputs data-driven.

A. Conditional GAN

The paper uses a conditional GAN to learn channel-output distributions from conditioning information, providing a differentiable surrogate that connects transmitter and receiver DNNs.

  • End-to-end integration: The learned GAN serves as a surrogate channel during transmitter training, allowing gradients to pass from the receiver to the transmitter.The transmitter, generator, and receiver are trained as connected DNN components while optimizing their respective objectives.
  • GAN fundamentals: A GAN learns a target distribution by having a generator produce samples and a discriminator distinguish generated samples from real data.Training alternates between generator and discriminator updates until the discriminator can no longer improve beyond random guessing.
  • Conditional GAN: A conditional GAN feeds extra information into both generator and discriminator, producing outputs conditioned on that information.The paper denotes the conditional outputs as G(x|m) and D(x|m).
  • Channel modeling: In the communication system, the conditional GAN models channel-output distributions conditioned on encoded signals and received pilot data.The pilot signal is included because channel outputs depend on the channel condition associated with the observation.

B. Modeling Channels

The channel model uses conditional generation with received pilot data and convolutional architectures to represent channel effects and manage long symbol sequences.

  • Conditional channel model: A conditional GAN learns the channel output distribution p(y|x) by using the transmitted signal x as conditioning information.The generator produces channel-like samples while the discriminator distinguishes them from real-channel outputs.
  • Conditional channel model: Received pilot data y_p is added to the conditioning information so generated outputs follow the distribution of y given x and y_p.The pilots provide information associated with the instantaneous channel condition.
  • Convolutional architecture: Convolutional layers reduce connections and trainable weights relative to fully connected layers, reducing complexity and improving training convergence speed.Their shared weights connect local receptive fields rather than every neuron in adjacent layers.
  • Convolutional architecture: Convolutional layers alleviate the curse of dimensionality for long sequences and are suitable for ISI channels because channel effects use convolutional operations.The paper uses hierarchical one-dimensional convolutional layers in the channel GAN, transmitter, and receiver.
  • Convolutional architecture: CNN-learned codes are reported to be more easily recovered at the receiver than conventional hand-designed codes.This claim is presented as an additional merit of using CNNs in the end-to-end communication system.

IV. END-TO-END COMMUNICATION SYSTEM

The end-to-end system treats transmitter, channel generator, and receiver as jointly trained modules that map information bits through an embedding and recover them from channel observations.

  • System architecture: The system is framed as a deep auto-encoder in which transmitter and receiver DNNs learn the communication mapping.The transmitter maps information bits to a fixed-length embedding, while the receiver reconstructs the original information.
  • System architecture: The transmitter sends an embedding x ∈ R^K through the channel, and the receiver recovers the original information from the received signal y.The reconstruction is compared with the original bit sequence s.
  • Training objective: Binary cross-entropy measures the distance between original information bits s and reconstructed outputs ŝ.The loss is computed from corresponding elements of the original and recovered sequences.
  • Training procedure: Training data combine randomly generated information bits with instantaneous CSI sampled randomly from a channel set.The training and testing procedures are described as using different objectives for the system modules.
  • Testing procedure: During testing, end-to-end reconstruction performance is evaluated using the learned transmitter and receiver with real channels.The conditional GAN is used during training as part of the learned communication system.

B. Training Receiver

The receiver learns to recover transmitted bits from received data and pilots, while the channel GAN enables transmitter optimization through back-propagated end-to-end gradients.

  • Training Receiver: The receiver DNN takes received data y as input and outputs an estimate ŝ of the transmitted information s.Comparing s and ŝ provides the loss used for receiver training.
  • Training Receiver: For time-varying channels, combining received data y with pilot data y_p lets the receiver infer channel conditions and detect symbols simultaneously.The method avoids explicitly estimating the channel in the receiver input-processing procedure.
  • Training Transmitter: The channel generator acts as a surrogate channel so transmitter training can use the same end-to-end loss-based procedure as receiver training.The transmitter, generator, and receiver are treated as one whole DNN during training.
  • Training Transmitter: Gradients from the receiver loss are propagated through the conditional GAN to update transmitter weights using SGD.The conditional GAN and receiver remain fixed during this transmitter update.
  • Training Transmitter: The transmitter learns an embedding constellation that makes the received signal easier for the receiver to detect.This describes the intended effect of updating transmitter weights through the learned channel model.
  • Training Channel GAN: The channel GAN is trained by alternating generator and discriminator updates using real-channel data and generator-produced encoded data.Its optimization follows a min-max objective with one model fixed while the other is trained.
  • Simulation results: The channel GAN models channel effects in a data-driven way, and the resulting end-to-end system achieves similar or better results with unknown channel information during training.The reported scope includes several commonly used channel types described in the paper.

A. Experimental Settings

The experiments evaluate DNN-based communication across AWGN, Rayleigh, and frequency-selective channels, using FCNs for small blocks and CNNs for large blocks. Rayleigh channel modeling conditions the GAN with received pilot data to represent time-varying channel effects.

  • 1) Implementation Details: Two DNN architectures are used: FCNs for small block sizes and CNNs for large block sizes to mitigate dimensionality challenges.The model parameters are reported separately for the FCN and CNN architectures.
  • 2) Channel Types: Three channel types are evaluated: AWGN, Rayleigh fading, and frequency-selective multipath channels.The frequency-selective model uses a three-tap channel with equal average path power and delays of 0, T, and 2T.
  • 2) Channel Types: For Rayleigh fading, the channel coefficient is time-varying and unknown during transceiver design, making instantaneous CSI estimation necessary for receiver detection.The channel output is modeled as y_n = h_n·x_n + w_n, with h_n ∼ CN(0, 1).
  • 3) Baseline System: The conventional baseline uses modular signal processing with 4 QAM, Hamming or convolutional coding, Viterbi/MAP decoding, and OFDM for frequency-selective multipath.The proposed system is compared with this channel-knowledge-based baseline using BER and BLER.
  • B. Modeling the Channel Effects: The conditional GAN models Rayleigh channel effects using the encoded signal and received pilot data as conditioning information.The experiments use an FCN for Rayleigh fading and test the GAN with standard 16 QAM encoded symbols.
  • B. Modeling the Channel Effects: The conditional GAN generates Rayleigh-channel samples with different channel gains and phase rotations according to the conditioning information.This evaluates whether the learned conditional distribution reflects changes in the channel output.

C. End-to-end Communication System

A channel-agnostic end-to-end system is built around the channel GAN and evaluated on AWGN, Rayleigh fading, and frequency-selective multipath channels. Its BER and BLER are compared with conventional coded-modulation baselines across small and large block sizes.

  • C. End-to-end Communication System: The channel GAN enables a channel-agnostic end-to-end communication system spanning AWGN, Rayleigh fading, and frequency-selective multipath channels.The approach is compared with traditional methods designed using channel-transfer-function knowledge.
  • 1) AWGN Channel: For a small AWGN block transmitting four information bits with output length seven, the learning-based BER and BLER are similar to Hamming (7,4) with MLD.This comparison is reported for the small-block AWGN experiment.
  • 1) AWGN Channel: CNNs are used for large block sizes, with AWGN training performed at a fixed 3 dB SNR and evaluation across different SNRs.The CNN is trained with noise added directly to the hidden layer.
  • 1) AWGN Channel: For small blocks, FCN-based E2E-64 and E2E-100 perform similarly to RSC at low SNR and significantly outperform RSC at high SNR.The transmit information sequence lengths are 64 and 100 bits, respectively.
  • 2) Rayleigh Channel: For 64-bit large-block transmission over Rayleigh fading, the end-to-end approach shows BER and BLER performance similar to traditional QAM-plus-RSC methods.The baseline uses QAM modulation and rate-1/2 RSC coding.

3) Frequency-Selective Fading Channel:

The proposed end-to-end system is evaluated on frequency-selective multipath channels, including mismatched training and testing power-delay profiles. It outperforms OFDM at higher SNR and remains robust to the mismatch.

  • Experimental setup: The evaluation uses coded and uncoded CNN-based end-to-end systems, with OFDM as the baseline.The OFDM baseline uses 64 subcarriers, a cyclic prefix of 16, and 4-QAM; the end-to-end block size is 64 bits.
  • Results: The proposed end-to-end system significantly outperforms OFDM when the SNR exceeds 7 dB.
  • Results: Despite substantial differences between training and testing power-delay profiles, the mismatched model performs comparably to the model without mismatch and remains much better than OFDM.Training uses an exponential-decay power-delay profile, whereas testing uses an equal-strength profile.
  • Conclusion: Across AWGN, Rayleigh fading, and frequency-selective channels, simulations show similar or better performance than traditional expert-knowledge-based approaches.The end-to-end pipeline iteratively trains transmitter DNNs, a channel GAN, and receiver DNNs to optimize end-to-end loss.
Loading 1903.02551v1…