Learnable Model-driven Performance Prediction for MIMO Systems

‍ ‍Practical wireless communication systems are often complex and experience unknown cochannel interference and uncertainties in the channel state information （CSI）. Existing performance analysis and optimization schemes for multiple-input multiple-output （MIMO） systems frequently degrade or even fail. To adapt to these challenges and perform better network auto-tuning， we propose a generic data- and model-driven， i.e.， dual-driven， framework， which is a digital twin of an actual MIMO communication system. We theoretically prove that the proposed dual-driven scheme has fewer training errors than the data-driven schemes based on the Lipschitz continuity condition. To explain the mechanism of the proposed framework works， we use regularized zero-forcing （RZF） precoding in an imperfect MIMO system as an example. The system performance is determined by system parameters. In this example， first， on the model-driven side， deterministic equivalence theory is used to obtain an approximation result of the system performance indicators. However， the approximation result holds only under ideal conditions， e.g.， known CSI uncertainty of the system， infinite number of antennas， and absence of any unknown interference. Therefore， on the data-driven side， we propose the use of a neural network （NN） to further elicit a more accurate system performance. Specifically， the input of the NN includes the approximation results and system parameters， and the output is a refined estimation of the system performance indicators. Thus， the learned NN can inherently solve imperfect system problems. The proposed dual-driven performance estimator is also a digital twin of the actual system. Based on this digital twin， a channel uncertainty estimation flow and algorithm are designed to quickly sense the channel uncertainty to support adaptive optimization of this system. In this flow， the system first performs signaling with an initial channel uncertainty and obtains the detection results of the performance metrics. Subsequently， the channel uncertainty is inverted with this detection result based on the performance estimator using the gradient projection method. For an imperfect MIMO system with unknown inference， simulation results show that the proposed dual-driven scheme significantly outperforms both the data-driven and model-driven baselines in terms of the prediction error of system performance and CSI uncertainty. This work provides insights into dual-driven performance prediction and optimization design in imperfect MIMO systems.