Link adaptation in Underwater Wireless Optical Communications based on deep learning

Abstract

A deep learning framework is proposed to address the research problem of link adaptation in Underwater Wireless Optical Communications (UWOCs). In this framework, the wireless receiver is assumed to be a black box due to hardware interface constraints; only the received time-domain signal waveform after analog-to-digital conversion is available at the receiver to perform link adaptation inference. The novelty of this framework is that this is the first investigation of this research problem and the first solution based on the deep learning approach. A solution based on a deep Recurrent Neural Network (RNN) named SwitchOpt RNN is proposed, with alternating optimization to tune hyperparameters. Based on the evaluation in a UWOC system with datasets generated from the link-level simulator, the proposed SwitchOpt RNN can effectively perform the link adaptation classifications. This work has significance in broader applications besides underwater networks including terrestrial wireless cellular communication systems.