DeepEigen-Tabu: Deep Eigen Network Assisted Probabilistic Tabu Search for Massive MIMO Detection

Abstract

Massive multiple-input multiple-output (MIMO) is a promising technology for enhancing quality of service in communication systems, but deploying numerous antennas increases detection complexity. To address this challenge, this paper introduces a novel detection scheme called DeepEigen-Tabu, combining the deep learning-based eigen network (DeepEigNet) with probabilistic Tabu search (P-TS). In the proposed scheme, DeepEigNet, a deep neural network, is constructed to utilize the eigenvalues and eigenvectors of the channel matrix to provide approximate symbol estimates. Subsequently, these estimates serve as the initialization and are prioritized according to their probabilities of correction to support the P-TS. Furthermore, the P-TS integrates an early stopping mechanism based on correction probabilities to eliminate unnecessary iterations during the Tabu search process. Finally, computer simulations and complexity analysis demonstrate that the proposed DeepEigen-Tabu scheme outperforms existing methods while maintaining lower complexity. For instance, in communication scenarios with both transmit and receive antennas set to 16, the proposed DeepEigen-Tabu method demonstrates savings of approximately signal-to-noise ratio (SNR) 0.8 dB at bit error rate (BER) 10(-3), compared to existing approaches in 4-ary quadrature amplitude modulation (4-QAM) symbol modulation. When the number of antennas is increased to 24 and using 16-QAM, the proposed DeepEigen-Tabu provides an improvement of 0.5 dB in SNR performance. Specifically, the proposed DeepEigen-Tabu not only achieves superior performance, as mentioned earlier, but also incurs a lower computational cost. The performance enhancements can be attributed to the DeepEigNet's provision of effective initialization, along with the early stopping and efficient candidate movement mechanisms employed by the P-TS method.