Trainable Communication Systems: Concepts and Prototype

Abstract

We consider a trainable point-to-point communication system, where both transmitter and receiver are implemented as neural networks (NNs), and demonstrate that training on the bit-wise mutual information (BMI) allows seamless integration with practical bit-metric decoding (BMD) receivers, as well as joint optimization of constellation shaping and labeling. Moreover, we present a fully differentiable neural iterative demapping and decoding (IDD) structure which achieves significant gains on additive white Gaussian noise (AWGN) channels using a standard 802.11n low-density parity-check (LDPC) code. The strength of this approach is that it can be applied to arbitrary channels without any modifications. Going one step further, we show that careful code design can lead to further performance improvements. Lastly, we show the viability of the proposed system through implementation on software-defined radios (SDRs) and training of the end-to-end system on the actual wireless channel. Experimental results reveal that the proposed method enables significant gains compared to conventional techniques.

WGAN-based Autoencoder Training Over-the-air

“…As previously mentioned, the idea of this autoencoder setup is to maximize the BMI at the receiver's output, which is shown in [12] to be closely related to minimizing the total binary CE, and leads to the following loss definition 1…”

“…Contrary to [9], [10], the wireless channel can become an attractive subject of investigation once multipath and, thus, frequency-selectivity becomes part of the transmission. We utilize the orthogonal frequency division multiplex (OFDM)-autoencoder structure from [11] and optimize the autoencoder for bit-wise information transmission as introduced in [12]. Further, we utilize Wasserstein generative adversarial networks (WGANs) [13] for improved convergence and training stability.…”

“…[14]). Most of the previously proposed fully end-to-end trained autoencoder-based communication systems rely on optimizing the mutual information between channel input and channel output by minimizing the symbol-wise categorical crossentropy (CE) (see [12] for a detailed derivation). The big drawback of this symbol-wise architecture is that it cannot be scaled to practical (bit) sequence lengths, as it suffers from the curse of dimensionality [15].…”

“…To reduce this complexity, practical systems usually rely on bitinterleaved coded modulation (BICM) and bit-metric decoding (BMD). We follow the approach of [12] and combine the autoencoder NN in the BICM framework with an outer channel code which can be decoded by a fully differentiable belief propagation (BP) decoder. Such an autoencoder system can then be trained in an end-to-end manner to maximize the bit-wise mutual information (BMI) at its output, which is also the decoder's input, and, thereby, inherently learns the optimal constellation shaping and bit labeling.…”

“…Such an autoencoder system can then be trained in an end-to-end manner to maximize the bit-wise mutual information (BMI) at its output, which is also the decoder's input, and, thereby, inherently learns the optimal constellation shaping and bit labeling. Throughout this work, we use the bit-wise iterative autoencoder architecture as described in [12] with an outer IEEE 802.11n WLANirregular low-density parity-check (LDPC) code of rate r = 1 /2, length n = 1296 bit and 40 iterations of iterative demapping and decoding (IDD) between the autoencoder receiver and the differentiable BP decoder.…”