Dense network for joint Transmitter and Receiver noncoherent codebook design

Unsupervised Deep Learning for MU-SIMO Joint Transmitter and Noncoherent Receiver Design. Songyan Xue et.al. IEEE Wireless Communications Letters, February 2019 (pdf) (Citations 24)

Quick Overview

• Joint transmitter and ==noncoherent== receiver optimization fo MU-SIMO through unsupervised learning. Find the optimal waveform set \(\mathbf{A}\) for multiusers.
• Communication chain can be represented by a DNN with three essential layers.
  • The first layer is a partially-connected linear layer responsible for multiuser waveform joint optimization.
  • The others (the second and the third layers) are nonlinear dense layers for noncoherent multiuser detection at the received side

其是很简单，就是在收发机联合训练，找到最适合的波形，使得其能够在发射端分离出来。所谓的创新点还有：

• 减少hidden layers，从而减轻梯度爆炸或消失，并且性能不会收到损失。
• 采用部分链接的Dense层来模拟多发射机并行传输的行为。
• 非对称DNN结构。

Transmitter

Codebook for transmitter:

Choose a transmitter codebook \(\mathbf{A}\) to minimize the Channel Ambiguity.

Each codebook has \(K\) codewords. So the available codebook for \(m\)-th user is \(\mathbf{W}_m\in\mathbb{R}^{L\times K}\). The one-hot vector \(\mathbf{a}_m\) (uniformly distributed) has the size of \(K\times 1\), which denotes the chosen index of codebooks.