Summary

Chapter Summary

Key Points

• 1.

  Supervised learning maps observations to estimates without explicit models. Neural-network channel estimators learn the mapping from pilot observations to full channel estimates, implicitly capturing the power delay profile and cross-subcarrier correlations that classical LS estimation ignores. A two-layer network with $\sim$5000 parameters, trained on $\sim$500 channel realisations, can approach MMSE performance without requiring explicit knowledge of the channel correlation matrix $\mathbf{R}_{HH}$. The universal approximation theorem guarantees that a sufficiently wide network can approximate the MMSE estimator $\hat{\mathbf{x}}_{\mathrm{MMSE}} = \mathbb{E}[\mathbf{x}|\mathbf{y}]$ to arbitrary accuracy.

• 2.

  End-to-end autoencoders jointly optimise transmitter and receiver. By treating the communication chain (encoder $\to$ channel $\to$ decoder) as a single neural network, the autoencoder framework (O'Shea and Hoydis, 2017) discovers optimal constellation geometries and detection rules simultaneously. For AWGN channels, the learned constellations converge to classical PSK/QAM arrangements, validating the approach; for non-standard channels (non-linear amplifiers, limited feedback), the autoencoder can discover novel solutions that outperform hand-designed schemes. Backpropagation through the AWGN channel is trivial ($\partial\mathbf{y}/\partial\mathbf{x} = \mathbf{I}$); for non-differentiable channels, surrogate models or policy-gradient methods are required.

• 3.

  Deep unfolding converts iterative algorithms into trainable networks with dramatic efficiency gains. LISTA unfolds $L$ ISTA iterations into $L$ layers with per-layer learnable thresholds, achieving linear convergence ($\rho^L$) compared to ISTA's sublinear $O(1/t)$ rate. Typically $L = 5$--$20$ learned layers match $T = 50$--$500$ classical iterations, yielding a $5$--$25\times$ reduction in computational cost. The principle extends beyond sparse recovery: any iterative algorithm (ADMM, gradient descent, belief propagation, WMMSE) can be unfolded, with the algorithm structure providing a strong inductive bias that dramatically reduces the number of trainable parameters compared to black-box alternatives.

• 4.

  Reinforcement learning enables adaptive resource allocation without explicit labels. Power control, scheduling, and beam management are sequential decision problems naturally modelled as MDPs. Tabular Q-learning discovers near-optimal power allocations for small systems ($K \leq 4$ users), while deep RL (DQN, DDPG, PPO) scales to larger networks. The key challenges are sample efficiency (RL requires many environment interactions), the curse of dimensionality in the joint action space ($J^K$ for $K$ users with $J$ power levels), and multi-agent coordination in decentralised settings.

• 5.

  Federated learning trains models across distributed base stations without sharing raw data. FedAvg aggregates local model updates from $C$ clients in communication rounds, reducing communication cost and preserving data privacy. Under IID data, FedAvg converges to the global optimum at a rate comparable to centralised SGD. Under non-IID data (the common case in wireless, where each BS sees different channel statistics), a persistent bias $O(E \cdot \Gamma / \mu)$ remains, proportional to the data heterogeneity $\Gamma$ and the number of local epochs $E$. Over-the-air aggregation, client selection, and gradient compression are active research areas for making FL practical over wireless links.

• 6.

  Model-based ML outperforms black-box networks in the data-scarce regime that characterises wireless. Deep unfolding requires $10$--$100\times$ fewer training samples than black-box NNs for comparable performance, because the algorithm structure encodes physical invariants. Model-based approaches also generalise better under distribution shift (different SNR, channel model, or number of users) and offer interpretability (each layer corresponds to one algorithm iteration). Black-box NNs are preferred when no suitable algorithm exists, abundant data is available, or the channel model is too complex for analytical treatment. In practice, hybrid approaches --- model-based backbone with NN refinement --- often achieve the best trade-off.