Adaptive Receiver Design Using Predictive Meta-Learning for Dynamic
A new trend is the use of deep neural networks (DNNs) in designing the receiver which can be perhaps used in an environment where knowledge of channel model is not paramount. Communicative channels, however, because they are dynamic, tend to experience very fast distribution shifts which sometimes call for relatively retraining after some time. In this paper, we present a two stage training scheme aimed at reducing workload on the user and achieving rapid online adaptation. Under out training scheme, we employ a meta-learning approach whereby the receiver can learn quickly from its prior deployment as well from its current context. Our method is intended for any DNN receiver and does not impose the obligation of sending additional pilot signals for further training. To this end, we apply primarily model-based high-velocity deep neural networks (DNNs) to develop data-aided receivers, where we propose a modular training on predictive meta-learning. We validate our methods with simulations on a synthetic linear channel, a synthetic non-linear channel, and a COST 2100 channel. Based on this fact the proposed online training allows to achieve quantitative performance improvement over completely different in design self-learning and joint learning strategies of up to 2.5 dB in terms of coded bit error rate at high dynamics of changes.