In multiple-input multiple-output (MIMO) visible light communication (VLC) systems, collaborative constellation design (CC) is effective in improving performance while significantly reducing the total optical power. This letter presents the first attempt to exploit deep learning (DL) in the signal design for MIMO VLC systems using CC, namely CCNet. More specifically, we present a deep neural network with fully connected layers to design an optimal constellation which can be derived conventionally at a high computational cost using the convex optimization solving procedure. Moreover, to enhance the performance of the proposed CCNet, instead of employing only the ordinary channel state information (CSI), the input is efficiently preprocessed before entering the network. By employing training samples collected in simulations, CCNet is first trained offline with the objective of minimizing the mean square error (MSE), and the online CC designing process can then be effectively deployed using the trained model. Simulation results show that CCNet can achieve a near-optimal MSE with a remarkably lower running time than the previous CC designing scheme.

中文翻译：

---

基于深度学习的可见光通信协同星座设计

在多输入多输出 (MIMO) 可见光通信 (VLC) 系统中，协同星座设计 (CC) 可有效提高性能，同时显着降低总光功率。这封信首次尝试在使用 CC（即 CCNet）的 MIMO VLC 系统的信号设计中利用深度学习 (DL)。更具体地说，我们提出了一个具有完全连接层的深度神经网络来设计一个最佳星座，该星座可以使用凸优化求解程序以高计算成本进行常规推导。此外，为了提高所提出的 CCNet 的性能，而不是仅使用普通的信道状态信息（CSI），输入在进入网络之前被有效地预处理。通过使用在模拟中收集的训练样本，CCNet 首先以最小化均方误差 (MSE) 为目标进行离线训练，然后可以使用训练模型有效部署在线 CC 设计过程。仿真结果表明，与之前的 CC 设计方案相比，CCNet 可以以显着缩短的运行时间实现接近最优的 MSE。