Deep Learning-Based MAP Selection Techniques for Media-Based Modulation Scheme

This article develops a deep neural network (DNN)-assisted mirror activation pattern (MAP) subset-selection framework for media-based modulation (MBM). MBM conveys information through an <tex-math notation="LaTeX">M </tex-math>-ary quadrature amplitude modulation (QAM) symbol and radio-frequency (RF) mirror-state indices. Classical Euclidean-distance, MAP-correlation, and capacity-optimized selection (EMS, MCMS, and CMS) improve reliability by selecting favorable MAP subsets from instantaneous CSI. To avoid repeating these rule-based searches after each CSI update, the proposed deep learning (DL) schemes—DL-EMS-MBM, DL-MCMS-MBM, and DL-CMS-MBM—learn the corresponding subset labels offline and perform online selection with one feedforward inference pass. Under the article’s bit-mapping model, <tex-math notation="LaTeX">mrf=2 </tex-math> and <tex-math notation="LaTeX">M∈ \2,4,8\ </tex-math> yield nominal spectral efficiencies of 3, 4, and 5 bits/s/Hz, respectively. At <tex-math notation="LaTeX">BER=10⁻⁵ </tex-math> in Rayleigh fading, MAP selection lowers the required signal-to-noise ratio (SNR) by 0.93–12.31 dB relative to conventional MBM, corresponding to 19.3%–94.1% less required radiated symbol energy for fixed noise power. The DNN predictors achieve 98.72%–99.04% validation accuracy, and their BER curves remain within 0.059 dB of the corresponding classical selectors at <tex-math notation="LaTeX">BER=10⁻³ </tex-math>. Under Nakagami- <tex-math notation="LaTeX">m </tex-math> fading, CMS/DL-CMS provides 3.209–5.00 dB SNR savings. Thus, learned MAP selection preserves the reliability gains of classical selection with a predictable online inference stage. The reported energy values are link-level radiated-energy savings, not a complete hardware energy-efficiency assessment.