Chapter 25: AI/ML for Massive MIMO

Learning Objectives

• Distinguish model-based and data-driven deep learning for wireless, and argue when each is appropriate
• Formulate channel estimation as a learned denoising problem and compare against LS / MMSE baselines under distribution shift
• Describe CsiNet and Transformer-based CSI feedback compression, and place them in the rate-distortion framework against 5G NR Type II codebooks
• Set up a beam prediction task on mmWave mobility traces and analyze why sequence models (LSTM, Transformer) beat per-slot beam search
• Formulate scheduling and power control as a Markov decision process, describe PPO, and honestly catalogue its pitfalls in production (reward gaming, simulation-to-real gap, training instability)
• State the CommIT / 6G@TU Berlin-Huawei workshop position that model-based DL is the right path for physical-layer learning and justify it with examples from deep unfolding and physics-informed losses