This study provides a comprehensive methodological contribution through rigorous comparative analysis between deep learning approaches and traditional optimization algorithms for adaptive beamforming in MIMO systems. Traditional optimization methods face significant challenges in dynamic environments due to computational complexity and convergence issues. Through systematic experimentation with standardized datasets (DeepMIMO, 3GPP TR 38.901, and IEEE MIMO Data Challenge), we evaluate performance using statistically validated metrics including signal-to-interference-plus-noise ratio, bit error rate, computational efficiency, and adaptation speed. Our findings reveal that deep learning approaches achieve significantly faster convergence (23.7%, p < 0.01) and higher SINR (18.5%, p < 0.01) in dynamic channel conditions, while traditional algorithms maintain superior performance in steady-state scenarios. Traditional methods outperform deep learning by 12.3% (p < 0.01) in terms of BER in low-SNR environments. Computational complexity analysis shows traditional methods scale as O(N?) with MIMO size N, while deep learning maintains O(N) inference complexity. The main contribution of this work is a novel adaptive selection framework with mathematically proven optimality bounds that dynamically switches between methodologies based on current channel conditions, achieving 15.3% higher average SINR (p < 0.01) in mixed scenarios compared to fixed algorithms.
Mohamed et al. (Thu,) studied this question.