Detection over Rapidly Changing Communication Channels Using Deep Learning

In this work, we demonstrate that deep learning algorithms can be trained for detection in rapidly changing channels, without any knowledge about the underlying channel models. This is achieved by training a deep learning algorithm called the sliding bidirectional recurrent neural network (SBRNN). The algorithm is trained using a dataset that is generated by transmitting a modulated data sequence (symbol sequence) over a parameterized Poisson channel, where the channel parameters are constant over each training sequence and then change for the next sequence. We show that although the SBRNN algorithm is trained using data where the channel state information (CSI) is constant over each sequence of symbols, it still performs well in rapidly changing channels, where the CSI changes on the order of the symbol duration.