Federated Convolutional Auto-Encoder for Optimal Deployment of UAVs with Visible Light Communications

Abstract

In this paper, the problem of unmanned aerial vehicles (UAV) deployment is investigated for visible light communication (VLC)-enabled UAV networks. Here, UAVs can simul-taneously provide communications and illumination services to ground users. In this model, ambient illumination distribution of the service area must be considered since it can cause interference over the VLC link and affects the illumination requirements of users. This problem is formulated as an optimization problem, which jointly considers UAV deployment, user association, power efficiency, and predictions of the illumination distribution. To solve this problem, we first need to predict illumination distribution to proactively determine the UAV deployment and user association so as to minimize total transmission power of UAVs. To predict the illumination distribution of the entire service area, a federated learning framework based on the machine learning algorithm of convolutional auto-encoder (CAE) is proposed. Compared to the centralized machine learning algorithms that requires complete illumination data for centralized training, the proposed algorithm enables the UAVs to train their local CAE with partial illumination data and cooperatively build a global CAE model that can predict the entire illumination distribution. Using these predictions, the optimal UAV deployment and user association policy that minimizes the total transmission power of UAVs is determined. Simulation results demonstrate that the proposed approach reduces the transmission power of UAVs up to 14.8% and 25.1%, respectively, compared to the local CAE prediction models and the conventional optimal algorithm without illumination distribution predictions.