Configuring RNN Weights for MIMO-OFDM Receive Processing: Informing RNN With Domain Knowledge

The recurrent neural network (RNN) is widely used for temporal pattern learning, but suffers from high training cost and computational complexity. The echo state network (ESN) is a special form of RNN with randomly initialized and fixed input and recurrent connections, enabling low complexity training. This work empirically demonstrates that the domain-specific knowledge from MIMO-OFDM systems can be incorporated into RNN by configuring the input and recurrent weights and achieve better MIMO-OFDM detection performance than ESN and traditional RNN trained via backpropagation through time (BPTT), while maintaining low computational complexity suitable for efficient real-time online learning. Our findings show that with increasing training data, BPTT gradually aligns the RNN’s spectral distribution with that of the configured RNN, but saturates thereafter, validating the efficiency of domain-informed weight configuration. Furthermore, to address the issue of increased generalization error as model complexity grows, this paper introduces a two-stage training strategy for the configured RNN, enhancing its generalization capability and reducing error variability across unseen data.