Configuration Learning in Underwater Optical Links

A new research problem named configuration learning is described in this work. A novel algorithm is proposed to address the configuration learning problem. The configuration learning problem is defined to be the optimization of the Machine Learning (ML) classifier to maximize the ML performance metric optimizing the transmitter configuration in the signal processing/communication systems. Specifically, this configuration learning problem is investigated in an underwater optical communication system with signal processing performance metric of the physical-layer communication throughput. A novel algorithm is proposed to perform the configuration learning by alternating optimization of key design parameters and switching between several Recurrent Neural Network (RNN) classifiers dependant on the learning objective. The proposed ML algorithm is validated with the datasets of an underwater optical communication system and is compared with competing ML algorithms. Performance results indicate that the proposal outperforms the competing algorithms for binary and multi-class configuration learning in underwater optical communication datasets. The proposed configuration learning framework can be further investigated and applied to a broad range of topics in signal processing and communications.