Optimizing Space Communications Using Deep Learning
Brianna I. Robertson, Aaron Dennis Smith · 2021
The paradigm shift from standard communication models to artificially intelligent solutions continues to be an exciting avenue of research. Through applying neural networks to a communication-learning model, the autonomous development of robust signal constellations emerges; however, optimal training architectures for these algorithms have not been determined. This work focuses on improving the modulation and coding solutions produced by deep neural networks by evaluating various training methods and configurations. The communications system is modeled with an autoencoder, an auto-associative neural network that learns how to transmit and recover input data, representing transmitter and receiver. Effective training of an autoencoder-based system encourages the model to learn modulation and coding schemes through optimizing symbol placement, which is achieved in this work through proper network normalization and adding channel noise. A normalization layer bounds symbol energy in a learned constellation, but these layers can impair the reproducibility of the training processes. This work assesses the effect of various normalization layers on the minimum distance between symbols in a learned constellation. Furthermore, a nonlinear noise technique, which intentionally mislabels symbols during training, is implemented to encourage symbol separation. This work concludes with a comparison between bit-wise versus symbol-wise training to determine if the learned constellation can achieve gray coding-like results.