PaperlayerView exact revision on arXiv

Source-linked AI summary

Wireless Networks Design in the Era of Deep Learning: Model-Based, AI-Based, or Both?

Alessio Zappone, Marco Di Renzo, Mérouane Debbah

arXiv:1902.02647v2eess.SPcs.IT

TL;DR

Future wireless networks must adapt to heterogeneous users, changing scenarios, and complex smart radio environments. The paper surveys and develops deep-learning methods that complement mathematical models, with case studies showing practical performance and reduced data requirements. Its conclusion is that combining model-based and data-driven approaches is the most promising direction.

  • Problem

    Future wireless networks require flexible adaptation to heterogeneous service needs and increasingly complex environments that are difficult to model and optimize in real time.

  • Method

    The paper surveys deep learning for wireless networks and develops model-assisted approaches using artificial neural networks, transfer learning, and deep reinforcement learning.

  • Results

    The case studies show that ANN-based power allocation improves the complexity-performance trade-off, transfer learning combines model-based and data-driven methods, and deep Q-Network performance approaches an offline upper bound while exceeding an MDP-based solution.

  • Takeaways & Limitations

    Model-based and data-driven approaches should be used jointly because approximate mathematical models can provide useful prior information and reduce the data needed for deep learning.

  • Takeaways & Limitations

    Acquiring the large datasets needed by deep learning remains challenging because measurement overhead, time, energy, storage, and processing costs can be substantial.

Abstract

from arXiv · show

This work deals with the use of emerging deep learning techniques in future wireless communication networks. It will be shown that data-driven approaches should not replace, but rather complement traditional design techniques based on mathematical models. Extensive motivation is given for why deep learning based on artificial neural networks will be an indispensable tool for the design and operation of future wireless communications networks, and our vision of how artificial neural networks should be integrated into the architecture of future wireless communication networks is presented. A thorough description of deep learning methodologies is provided, starting with the general machine learning paradigm, followed by a more in-depth discussion about deep learning and artificial neural networks, covering the most widely-used artificial neural network architectures and their training methods. Deep learning will also be connected to other major learning frameworks such as reinforcement learning and transfer learning. A thorough survey of the literature on deep learning for wireless communication networks is provided, followed by a detailed description of several novel case-studies wherein the use of deep learning proves extremely useful for network design. For each case-study, it will be shown how the use of (even approximate) mathematical models can significantly reduce the amount of live data that needs to be acquired/measured to implement data-driven approaches. For each application, the merits of the proposed approaches will be demonstrated by a numerical analysis in which the implementation and training of the artificial neural network used to solve the problem is discussed. Finally, concluding remarks describe those that in our opinion are the major directions for future research in this field.

I. INTRODUCTION AND VISION

Future wireless networks will combine unprecedented complexity, heterogeneity, and adaptability requirements with AI-enabled, model-based design. The paper argues for synergy between mathematical models and deep learning, develops the supporting foundations, and evaluates this approach through wireless-network case studies.

  • I. INTRODUCTION AND VISION: Flexible architectures such as network slicing and UAV-based access points are proposed to support diverse, on-demand connectivity requirements.Network slicing creates customizable virtual networks over shared infrastructure, while UAVs can be redeployed according to heterogeneous traffic conditions.
  • I. INTRODUCTION AND VISION: Future networks combine dense, heterogeneous technologies and services, creating complexity that traditional deployment, design, and operation approaches may not adequately handle.Projected requirements include 1 ms latency with reliability above 99.999% for URLLC, one million terminals per square kilometer for mIoT, and user rates above 50 Mb/s for mBB.
  • I. INTRODUCTION AND VISION: AI capabilities distributed across network segments can support self-aware, self-organizing, and self-healing nodes that sense and process their environments.The paper links this cognitive behavior to intelligent wireless-network infrastructure rather than faster transmission technology alone.
  • I. INTRODUCTION AND VISION: The paper advocates using model-based and AI-based techniques in synergy rather than dismissing mathematical models in favor of data-driven methods.Its stated goal is cross-fertilization between the approaches, including using approximate models to reduce the amount of field data required.
  • I. INTRODUCTION AND VISION: The work surveys deep learning for wireless communications, explains its theoretical foundations and related learning frameworks, and quantitatively analyzes practical case studies.The case studies discuss ANN design, implementation, training, and resulting performance, while the paper also covers reinforcement learning, transfer learning, federated learning, and deep unfolding.

C. Deep Learning for Network Deployment and Planning

Future wireless networks are envisioned as intelligent, reconfigurable environments whose complexity exceeds conventional model-based optimization. The paper argues that deep learning should complement mathematical models to make such networks practically implementable.

  • Future wireless networks must integrate communication, sensing, and computing while supporting context-aware and locally processed services.
  • Smart radio environments turn the wireless environment into a software-reconfigurable entity that helps provide uninterrupted connectivity.
  • Reconfigurable meta-surfaces can modify radio-wave propagation, enabling customized reflection and refraction around cellular-network obstacles.
  • Mobile users and static base-station deployments create uneven loads, while densification increases power consumption and interference.
  • Smart radio environments introduce many environmental degrees of freedom, making accurate modeling and optimization computationally difficult.
  • Deep learning can simplify resource management, especially when AI-based methods are jointly used with model-based approaches.

E. Deep Learning for Network Operation and Maintenance

Network operation and maintenance involve diverse inference tasks for which tractable statistical models are often unavailable. Machine learning, including deep learning, offers a data-driven alternative, but its deployment must address data and distribution constraints.

  • Operation and maintenance covers localization, channel estimation, quality-of-service monitoring, anomaly detection, handovers, and intrusion detection.
  • These tasks can be formulated as estimating a random vector x from correlated observations y through an unknown transformation.
  • Classical decision and estimation methods require prior and conditional distributions that are often unavailable for wireless operation and maintenance tasks.
  • Machine learning addresses missing model information by learning from data rather than requiring explicit statistical distributions.
  • Supervised learning uses labeled input-output examples, whereas unsupervised learning uses input features alone to infer statistical structure.
  • Model capacity creates a trade-off: insufficient capacity causes underfitting, while excessive capacity can memorize training-specific structure and cause overfitting.

C. Beyond classical machine learning

Classical machine learning becomes increasingly limited as task dimensionality and complexity grow, while deep learning can extract relevant features and improve performance more rapidly with larger training sets. Deep learning uses hierarchical artificial neural networks, whose layered architectures support general-purpose input-output modeling.

  • C. Beyond classical machine learning: Deep learning is motivated by the limitations of classical machine learning as task complexity increases.Classical methods face constraints from training-data requirements, dimensionality, and task-specific assumptions.
  • C. Beyond classical machine learning: Deep learning performance improves faster than other machine learning techniques as training-data dimensionality increases.For small-to-medium training sets, classical algorithms can sometimes slightly outperform deep learning.
  • C. Beyond classical machine learning: Deep learning automatically extracts important features by modeling data as a hierarchy of underlying factors and correlations.This supports non-local generalization without requiring task-specific assumptions.
  • Artificial neural networks: Artificial neural networks consist of input, hidden, and output layers connected by weights and biases.A network with one hidden layer is shallow; one with multiple hidden layers is deep.
  • Artificial neural networks: Feed-forward networks propagate information forward, whereas recurrent networks permit feedback loops; convolutional and pooling layers provide additional architectures and operations.The section focuses primarily on fully connected feed-forward networks.
  • Artificial neural networks: A fully connected feed-forward network transforms its input through successive affine operations and activation functions across L hidden layers.The output of each neuron depends on the preceding layer’s activations, weights, bias, and activation function.

1) Convolutional neural networks:

Convolutional neural networks process multi-channel matrix data with shared filters, while pooling reduces local regions through maxima or averages. Convolution, padding, stride, and pooling determine the spatial structure and size of intermediate representations.

  • 1) Convolutional neural networks:: A CNN applies an F × F × Nc filter to an N × N × Nc input, producing an (N − F + 1) × (N − F + 1) output.The operation computes cross-correlations over sliding input regions, followed by bias addition and activation.
  • 1) Convolutional neural networks:: CNN convolution can be represented as a sparse fully connected operation with many permanently zero weights.This reduces parameter counts and helps process large data while avoiding overfitting.
  • 1) Convolutional neural networks:: CNNs assume strong spatial correlations so that the same filter can be applied across different input locations.This assumption enables shared filtering and avoids unnecessary neuron connections.
  • 1) Convolutional neural networks:: Padding adds border zeros, and stride controls how many positions the filter moves between successive operations.Both choices affect the dimensions of the output matrix.
  • 1) Convolutional neural networks:: Pooling divides each channel into local blocks and computes either the maximum or average within each block.Max pooling selects the largest element, whereas average pooling computes the arithmetic mean; stride may also be used.
  • 1) Convolutional neural networks:: Practical feed-forward networks commonly combine convolutional and pooling layers early with fully connected layers later.The early operations decrease data size before final fully connected processing.

B. Recurrent neural networks

The training section describes recurrent processing, supervised neural-network training, loss optimization, and the non-convexity that complicates global optimization. It also explains why gradient-based methods and suitable initialization remain important in practice.

  • B. Recurrent neural networks: Recurrent layers combine current inputs with previous intermediate activations, enabling neural networks to exploit temporal correlations.RNNs allow feedback loops, unlike purely feed-forward networks.
  • Training neural networks: Supervised training tunes network weights and biases to reduce the loss between actual and desired outputs over training samples.The mean squared error is a broadly applicable loss, while cross-entropy can converge faster in suitable probability-output settings.
  • Training neural networks: Training neural networks is a non-convex optimization problem because multiple layers combine nonlinear activation functions.The objective can contain local minima, maxima, and saddle points, and global optimization is not guaranteed to have polynomial complexity.
  • Training neural networks: Experimental evidence indicates that sufficiently wide networks often obtain satisfactory training error at local minima, while gradient descent can escape saddle points.The text presents saddle points, rather than local minima, as the main practical non-convexity concern.

PNT R

This section develops stochastic and adaptive optimization procedures for neural-network training. It emphasizes computational efficiency, learning-rate control, momentum, and the practical difficulty of initialization and non-convex optimization.

  • PNT R: Backpropagation computes derivatives efficiently through forward and backward passes, but processing the full training set makes each update scale with its size.This motivates stochastic gradient descent for large training sets.
  • PNT R: Stochastic gradient descent estimates the gradient from a randomly selected mini-batch and updates weights and biases using that estimate.Mini-batch size determines whether the method is batch, online, or intermediate SGD.
  • PNT R: The learning rate controls update speed and is commonly adjusted during early iterations using validation-error monitoring.The initial, final, and transition values are typically selected empirically.
  • PNT R: SGD has faster individual iterations and lower total computation to convergence than deterministic gradient descent, despite requiring more iterations.Its per-update complexity does not scale with the full training-set size.
  • PNT R: Momentum accelerates SGD by updating with an exponentially decaying moving average of previous gradients.The velocity term controls the relative speed of updates.
  • PNT R: Nesterov momentum evaluates the gradient after accounting for the current velocity term.Its convex-case convergence advantages are not guaranteed for the non-convex objectives typical of FNN training.
  • PNT R: AdaGrad adapts learning rates using accumulated gradients across mini-batches, while RMSProp and related methods address adaptive-rate behavior in training.The supplied passage introduces AdaGrad’s scaling principle and the RMSProp transition.
  • PNT R: Parameter initialization remains poorly understood because non-convex training can reach suboptimal points and low training loss may not imply low generalization error.Initialization must therefore be treated as a central training design issue.

2) Hyperparameter tuning - Fitting the data:

Hyperparameter tuning balances fitting the training data against generalization, using manual or automated approaches to adjust network capacity. The paper emphasizes monitoring training and validation errors and responding differently to underfitting and overfitting.

  • Automated hyperparameter selection: Automated hyperparameter selection includes grid search, random search, and hyperparameter optimization, but some methods are rarely practical for deep learning.Grid search is practical mainly for three or fewer hyperparameters, while optimization is hindered by unavailable loss derivatives and its own hyperparameters.
  • Manual hyperparameter setting: Manual tuning uses training and validation errors to identify underfitting or overfitting and adjust network capacity through trial and error.It can achieve desired performance at affordable complexity but requires substantial experience.
  • Fitting the training data: Underfitting requires improving optimization or increasing capacity, beginning with the learning rate and potentially changing optimizers, normalization, mini-batch size, or training data.Common optimizer choices include SGD with momentum, RMSProp, and Adam.
  • Fitting the validation data: Overfitting is addressed by gathering more data, applying regularization such as early stopping, or reducing the number of neurons and layers.If careful tuning fails, collecting more data remains the only stated way to avoid overfitting.
  • Validation reliability: Validation error estimates test error only when the validation procedure is reliable; otherwise, enlarging the validation set may be necessary.If that does not help, the validation procedure or a more fundamental issue may be responsible.

D. Deep Reinforcement Learning

Deep reinforcement learning combines reinforcement learning, which learns through environment interaction and feedback, with deep learning. The framework uses neural networks to estimate value or policy functions when the environment is complex or incompletely known.

  • Framework: Reinforcement learning learns adaptively by interacting with an environment, taking actions, and receiving feedback.Deep reinforcement learning combines this framework with deep learning.
  • Markov Decision Processes: An MDP describes reinforcement-learning scenarios using states, actions, transition probabilities, rewards, a discount factor, and a policy.The discount factor adjusts the weight assigned to more recent actions.
  • Value functions: The action-value function measures expected reward for taking an action in a state under a policy, including immediate and long-term rewards.It is the basis for evaluating action sequences.
  • Approach categories: Reinforcement-learning methods are classified as value-based, policy-based, or actor-critic approaches.Actor-critic methods estimate both the action-value and policy functions.
  • Role of deep learning: ANNs provide function approximators for action-value and policy functions when environment statistics and parameters are not fully known.This supports practical reinforcement learning in complex settings.

1) Deep Q-Network. Estimating the action-value function:

Deep Q-Networks extend Q-learning by using an ANN to approximate the action-value function. This avoids prescribing its functional form while learning from action-value samples or environmental experience.

  • Objective: Q-learning seeks to estimate the optimal action-value function, whose values characterize the MDP and determine the best policy.The Bellman optimality equation can solve the problem when the complete MDP model is known.
  • Motivation: Complete MDP knowledge is often unrealistic because practical agents typically observe environmental responses without transition statistics.The action-value function can then be learned from experience.
  • Q-learning: Traditional Q-learning assumes a parameterized functional form and fits its parameters to sampled action-value values by minimizing mean squared error.The samples may come from trying actions and observing the environment.
  • Deep Q-Network: A Deep Q-Network uses an ANN that takes a state-action pair as input and outputs the corresponding action-value estimate.Its weights are trained using action-value samples, making it a Q-learning method with a learned approximation function.
  • Neural approximation: ANN-based approximation leaves the functional form of the action-value function unspecified and can approximate the true function within any desired tolerance with proper training.This follows from the universal function-approximation ability attributed to ANNs.

2) Deep Policy Iteration. Estimating the policy function:

Deep policy iteration uses neural networks to represent policies, while actor-critic methods jointly estimate policies and action values. Deep unfolding instead maps iterative algorithm steps onto ANN layers, fixing architecture dimensions from the unfolded algorithm.

  • Deep Policy Iteration: Policy iteration parametrizes the policy function and adjusts its parameters to maximize the average reward using gradient ascent.The policy-gradient theorem supplies the gradient expression, while REINFORCE uses sampled returns from the environment.
  • Neural policy representation: An ANN policy maps a state to action probabilities, samples an action, and uses the resulting reward to weight training and update the policy adaptively.The process repeats after the action moves the agent to a new state.
  • Deep Actor-Critic: Actor-critic methods combine a critic ANN estimating action values with an actor ANN estimating policies.They improve purely policy-based methods by supplying an action-value estimate, and advantage functions can reduce estimation variance.
  • Deep unfolding: Deep unfolding maps each iteration of a recursive algorithm onto one ANN layer, with the final layer producing the algorithm’s final output.The unfolded network takes the algorithm inputs and initial state, computes successive iterates, and outputs y.
  • Architecture determination: Deep unfolding determines ANN depth from the algorithm’s iteration count and layer widths from the sizes of its input, intermediate, and output vectors.This provides a systematic alternative to manually choosing the number of neurons and layers.
  • Model-based and AI-based methods: Combining deep unfolding with transfer learning initializes a model-based ANN from an unfolded model and refines it with empirical data.The approach avoids tuning the number of layers and neurons because the model determines the architecture.

F. Deep Transfer learning

Deep transfer learning reuses knowledge from related source tasks or domains to support target tasks, especially when target-domain data are costly or scarce. The section surveys instance-, mapping-, network-, and adversarial-based approaches, with network-based transfer highlighted as particularly useful for wireless communications.

  • F. Deep Transfer learning: Transfer learning transfers knowledge from a source task and domain to a related target task and domain.The framework identifies source and target tasks and domains as its four fundamental components.
  • F. Deep Transfer learning: Transfer learning can reduce the data required for deep-learning training when target-domain measurements are expensive or impractical.Knowledge may come from related scenarios, including datasets generated from approximate mathematical models.
  • Instance-based transfer learning: Instance-based transfer learning augments target data with source-domain instances while weighting them according to their relevance to the target domain.Source and target datasets are both used for the target task, with source-specific instances receiving less or no importance.
  • Mapping-based transfer learning: Mapping-based transfer learning combines source and target losses with a weighting parameter and a regularizer accounting for domain differences.The formulation requires labelled data from both domains; standard training algorithms can then exploit the combined dataset.
  • 3) Network-based transfer learning: Network-based transfer learning first trains an ANN on the source task, then reuses or refines its configuration for the target task.A common two-step approach initializes target-domain training with source-trained weights and biases, which is useful when source data are abundant but target labels are scarce or costly.
  • 4) Adversarial-based transfer learning: Adversarial transfer learning uses a GAN to identify features shared by source and target tasks.The source-task ANN is divided into feature extraction and task-processing segments before adversarial alignment.

4) Adversarial-based transfer learning:

The paper surveys deep learning applications in wireless communications and develops expert-knowledge-aided approaches for practical resource allocation and receiver design. These methods aim to combine mathematical models with ANNs, reducing complexity or empirical-data requirements while retaining strong performance.

  • Physical-layer applications: Deep learning research at the physical layer follows two directions: operating the layer and managing radio resources.Operating tasks include detection, decoding, channel estimation, and localization; management focuses on radio resource allocation.
  • Physical-layer applications: ANNs have been used for physical-layer detection, estimation, localization, encoding, decoding, and joint equalization tasks.Examples include autoencoder-based end-to-end communication systems and neural implementations of channel estimation and signal detection.
  • Physical-layer applications: Resource-allocation studies increasingly use deep learning with minimum complexity or side-information requirements.The paper emphasizes ANN-based physical-layer resource allocation as a major application area.
  • B. Learning to optimize: Expert-knowledge-aided deep learning incorporates prior problem knowledge because purely data-driven approaches may require excessive data and processing for large-scale applications.Synthetic data generated from models can augment small real datasets and improve estimation accuracy in other application domains.
  • B. Learning to optimize: The proposed approaches reduce complexity relative to purely model-based methods, enabling real-time resource allocation with near-optimal performance.The paper presents this as a principal advantage of the methodologies.
  • B. Learning to optimize: They also reduce empirical-data requirements relative to purely data-driven methods, avoiding expensive and impractical measurement campaigns.Most subsequent case studies address resource allocation using the two proposed methodologies.
  • 1) Physical layer design: Optimizing the receiver of a molecular communication system: In molecular communication, the ANN-based demodulator learns a threshold that agrees well with the numerically optimized threshold across signal-to-noise ratios.The learned threshold is identified from the input producing an output probability of 0.5.

2) Optimizing a model: power control in wireless networks:

The paper formulates weighted sum energy-efficiency power control using mathematical models, then trains an ANN to approximate optimal allocations for low-complexity online operation.

  • WSEE power control is difficult because multi-user interference makes rate numerators nonconcave and the objective has a sum-of-ratios form.
  • The ANN maps system parameters to the optimal K-dimensional power vector, avoiding repeated numerical optimization when parameters change.
  • ANN inference has negligible computational complexity compared with sophisticated numerical optimization once its parameters and hyperparameters are fixed.
  • Training-set generation can remain practical because optimization is performed offline and the dataset is updated less frequently than network parameters change.
  • The ANN-based method outperforms practical alternatives, while the strongest comparable SCA method is more complex and performs slightly worse.The results support online power allocation through a better complexity-performance trade-off.

3) Optimizing a model: user-cell association in massive MIMO networks:

This case study uses an ANN to approximate optimal user-cell association in massive MIMO networks, reducing the complexity of solving the underlying discrete allocation problem.

  • User-cell association is formulated with binary transmitter-receiver variables, per-user association constraints, base-station capacity limits, and minimum-rate requirements.
  • Traditional branch-and-cut or relaxation-based optimization can remain computationally demanding, motivating an ANN approximation.
  • A fully connected ANN maps channel and system inputs to association outputs, using the same general implementation approach as the power-control case.
  • The ANN was trained with 140,000 samples and validated with 15,000 samples generated from independent user-position realizations.
  • ANN-based association performs similarly to optimal association with both MMSE and MR reception while requiring much lower computational complexity.

4) Refining a model by deep transfer learning - Cellular networks beyond the Poisson point process:

For cellular networks whose true deployment differs from a tractable model, the paper combines model-based pretraining with a small amount of empirical data through deep transfer learning.

  • The case study addresses energy-efficiency optimization when an analytical model is available but inaccurate for non-Poisson cellular networks.
  • The proposed method first trains an ANN with a simpler Poisson point-process model, then refines it using a few empirical samples from the square-grid network.
  • With 300, 600, 1500, 2100, or 3000 empirical samples, transfer learning achieves much lower training and validation errors than using only empirical samples.
  • Combining model-based data with 3,000 empirical samples provides near-optimal performance, whereas 30,000 model-based samples alone are unsatisfactory.
  • The paper concludes that transfer learning can combine the advantages of model-based and data-driven optimization.

5) Refining a model by deep transfer learning - Cellular networks with inaccurate power consumption models:

The paper extends model-assisted learning to energy-harvesting power control, using deep reinforcement learning to handle unknown processes and large discretized state-action spaces.

  • The energy-harvesting problem seeks to maximize long-term achievable rate using only causal channel and energy-arrival information.
  • An MDP formulation requires statistical process information and fine discretization, which can create large state and action spaces and high complexity.
  • The deep Q-Network uses 150 discretized transmit-power actions with step 10^-3.
  • The deep Q-Network achieves performance very close to the non-causal offline policy and outperforms the MDP-based solution.
  • The deep reinforcement learning method requires a few thousand time-slots for throughput stabilization in the reported convergence experiment.

V. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS

Deep learning is presented as an indispensable but still developing tool for future wireless networks, whose data demands make model-based prior knowledge especially valuable. The paper identifies major open challenges involving data acquisition, distributed integration, security, robustness, and the continued role of communications theory.

  • Deep learning is considered indispensable for future wireless network design, but deployment still requires solving many open problems.Recent advances in processing and distributed storage make deep learning more practical, yet ANN-based wireless networks are not ready for deployment.
  • Large datasets needed for satisfactory ANN performance are expensive to acquire in wireless networks and difficult to collect within channel coherence times.The paper therefore emphasizes combining data-driven methods with approximate mathematical models to reduce required live data.
  • Distributed ANN architectures can produce unequal learning capabilities across nodes and may destabilize the network when local datasets differ in size or quality.The paper also identifies the risk that nodes optimize individual rather than system-wide utility, motivating security mechanisms.
  • Training must become robust to corrupted data because feedback and storage errors can produce undesirable results and amplify distributed-network inconsistencies.The paper calls for methods that tolerate such corruption, particularly in distributed ANN implementations.
  • Deep learning should be initialized with wireless communications expertise to reduce the data, computation, energy, and overhead needed for target gains.The conclusion preserves a fundamental role for communications theory alongside data-driven learning.
Loading 1902.02647v2…