Data-Driven Solutions for Digital Communications

In this paper, we show that data-driven communication methods can overcome limitations of pure model-based designs. Classical communication theory works well with ‘ideal’ channels (AWGN, Rayleigh, etc.). However, an open area of research exists in how to transmit across non-ideal channels, such as high Doppler, nonlinear, or contested. Communications researchers have already started to make great progress using machine learning to address a variety of communications problems. We build on their findings and propose a variety of machine-learned communication techniques. We show a transmitter that uses reinforcement learning algorithms to learn efficient modulation schemes. The benefit of a reinforcement learned system is that it can change its transmission scheme based on the current characteristics of the channel. We show our reinforcement learned transmitter successfully able to learn an efficient modulation scheme given power constraints. We show our transmitter not only learning good modulation schemes for AWGN channels, but also adapting when suddenly presented with a non-linear amplifier in its path. In addition, while it takes a long time to initially train the transmitter, it relatively quickly adapts to the introduction and removal of non-linear amplification. This approach outperforms static modulation schemes that suffer when the non-linearities are introduced. Furthermore, we propose a receiver that uses a machine-learning classifier to decode either a fixed transmitter scheme or a learned transmitter scheme. We also implement and propose neural networks to simultaneously learn transmitter and receiving algorithms via supervised learning. Finally we implement a learned parameter selection model to select communication system parameters such as modulation order and other values dynamically based on current channel state. In this way we can utilize machine learning to improve the robustness and resiliency of critical SATCOM links.