Model-Driven IEP-GNN Framework for MIMO Detection with Bayesian Optimization

In this letter, a model-driven detector called IEP-GNN is proposed for massive multiple-input and multiple-output (MIMO) systems. Graph neural network (GNN) and improved moment matching (IMM) are integrated into the expectation propagation (EP) algorithm to improve the accuracy of posterior distribution approximation and leverage the self-correction ability of EP algorithm. Moreover, to acquire the training experiences and optimize initial parameters, hotbooting and Bayesian parameter optimization (BPO) are employed respectively, which can further improve the performance of the proposed IEP-GNN. Simulation results show that our proposed IEP-GNN with BPO outperforms other state-of-the-art EP-based detectors while maintaining an acceptable convergence and computational complexity.