Learn to Detect and Detect to Learn: Structure Learning and Decision Feedback for MIMO-OFDM Receive Processing

One of the major open challenges in MIMO-OFDM receive processing is how to efficiently and effectively utilize the extremely limited over-the-air pilot symbols to detect the transmitted data symbols. Recent advances have been devoted to investigating effective ways to utilize the limited pilots. However, we notice that besides exploiting the pilots, one can take advantage of the data symbols to improve the detection performance. Thus, this paper introduces an online subframe-based approach, namely RC-StructNet, that can efficiently learn from the precious pilot symbols and be dynamically updated with the detected payload data using the decision feedback (DF) approach. The network consists of a reservoir computing (RC) module in the time domain and a neural network StructNet in the frequency domain. The unique design of the network allows it to be dynamically updated with the changes of the channel by learning from the detected data symbols. Experiments demonstrate the effectiveness of RC-StructNet in detection under dynamic transmission modes and in reducing the training overhead requirement when taking the DF approach.