Online Optimization for Learning to Communicate over Time-Correlated Channels

Zheshun Wu, Junfan Li, Zenglin Xu, Sumei Sun, Jie Liu

TL;DR

This paper drops the unrealistic assumption of independently and identically distributed (I.I.D.) channels typically used to evaluate learning-based communication systems. By framing communication over time-correlated channels as an online optimization problem, the authors introduce two algorithms based on optimistic online mirror descent that leverage temporal dependencies to minimize symbol error rates with solid theoretical guarantees.

What Problem Does This Paper Address?

Machine learning has shown immense promise in optimizing physical-layer communication tasks like channel estimation, decoding, and modulation. However, proving that these systems will perform well on unseen data typically relies on statistical learning theory, which assumes channels are I.I.D.

In reality, wireless channels are highly dynamic and time-correlated due to factors like user mobility (causing Markov fading) and hardware non-linearities (introducing colored noise). When the I.I.D. assumption breaks down, standard theoretical generalization guarantees collapse because classic concentration inequalities no longer apply. This paper addresses this gap by providing a framework to learn communication parameters on-the-fly over time-varying, correlated channels without relying on I.I.D. assumptions.

Methods

The authors frame the problem using Online Optimization, where a system continuously updates its decisions based on real-time feedback. Specifically, they adopt the Optimistic Online Mirror Descent (OMD) framework. The "optimistic" flavor is crucial here: it uses a "hint" (such as the gradient or loss from the previous round) to predict the next state, making it perfectly suited for exploiting temporal correlation.

The authors apply this framework to two distinct tasks:

  1. Decoder Learning (Online Convex Optimization): For time-correlated fading channels, they optimize a linear channel decoder. They design a novel hinge-type convex surrogate loss to upper-bound the expected error probability and solve it using optimistic OMD with a Euclidean regularizer.
  2. Codebook Selection (Multi-Armed Bandit): For time-correlated additive noise channels, the receiver must dynamically select the best modulation scheme (codebook) from a predefined set. This is modeled as an adversarial multi-armed bandit problem and solved using optimistic OMD with a Log-Barrier regularizer, where the hint is the most recently observed loss for each codebook.

The overall procedure for both tasks is illustrated below:

Fig. 1: The procedure of learning decoder and codebook via online optimization.

The transmission process assumes that within a single communication round (coherence time), the channel remains constant while a deterministic training sequence (similar to a pilot signal) is sent alongside random data symbols:

Fig. 2: Transmission process of uplink communication from mobile user to base station.

Datasets

No real-world datasets were used in this study. Instead, the authors evaluated their methods using extensive synthetic simulations modeling realistic channel dynamics:

  • First-order Markov fading channels.
  • Noise and channel variations generated via Gaussian Mixture Distributions (GMD) and Laplace Mixture Distributions (LMD).
  • Standard Rayleigh fading and Additive White Gaussian Noise (AWGN) baselines adapted for time-correlation.

Key Insights and Ideas

  • Harnessing the Past as a Hint: In standard online learning, temporal variation is often treated as an adversarial nuisance. Here, by using the previous round's gradient or loss as an optimistic hint, the algorithm turns channel correlation into an asset rather than a liability.
  • Sub-linear Regret Guarantees: The authors prove that both algorithms achieve sub-linear regret bounds. This means that as time goes on, the average performance of the online learner converges to the performance of the best fixed decision chosen in hindsight.
  • Quantifying Channel Dynamics: The derived theoretical bounds explicitly show that smaller cumulative variation ($\Sigma_{1:T}^H$) in the channel leads to tighter bounds, mathematically proving why and how the algorithm successfully exploits distribution dependency.

Key Results

  • Lower Symbol Error Rate (SER): In both decoder learning and codebook selection tasks, the optimistic OMD approach consistently outperformed standard baselines like Least Squares (LS) estimation, standard Online Gradient Descent (OGD), and the EXP3 bandit algorithm across both GMD and LMD noise profiles.
  • Faster Convergence: The optimistic updates allowed the system to adapt to channel conditions much faster than traditional online gradient descent methods.
  • Robustness to Correlation Levels: Simulations verified that as the channel became more temporally correlated (more predictable), the average symbol error rate dropped significantly, validating the authors' theoretical corollaries.

Limitations and Future Work

  • Simplified Channel Models: The channel models considered are relatively simplified (e.g., linear fading transformations or identity gains for the noise channel). Real-world channels can exhibit more complex, non-linear behaviors.
  • Codebook Constraints: The case study for explicit derivation of codebook selection focused heavily on a simplified 2-ary codebook scenario.
  • Future Directions: The authors note that future work will focus on deriving fundamental lower bounds for learning over time-correlated channels and developing algorithms capable of leveraging finer-grained channel state information.

Summary updated

Abstract

Machine learning techniques have garnered great interest in designing communication systems owing to their capacity in tackling with channel uncertainty. To provide theoretical guarantees for learning-based communication systems, some recent works analyze generalization bounds for devised methods based on the assumption of Independently and Identically Distributed (I.I.D.) channels, a condition rarely met in practical scenarios. In this paper, we drop the I.I.D. channel assumption and study an online optimization problem of learning to communicate over time-correlated channels. To address this issue, we further focus on two specific tasks: optimizing channel decoders for time-correlated fading channels and selecting optimal codebooks for time-correlated additive noise channels. For utilizing temporal dependence of considered channels to better learn communication systems, we develop two online optimization algorithms based on the optimistic online mirror descent framework. Furthermore, we provide theoretical guarantees for proposed algorithms via deriving sub-linear regret bound on the expected error probability of learned systems. Extensive simulation experiments have been conducted to validate that our presented approaches can leverage the channel correlation to achieve a lower average symbol error rate compared to baseline methods, consistent with our theoretical findings.
BibTeX
@article{wu2026online,
  author = {Zheshun Wu and Junfan Li and Zenglin Xu and Sumei Sun and Jie Liu},
  title = {{Online Optimization for Learning to Communicate over Time-Correlated Channels}},
  journal = {IEEE Trans. Cogn. Commun. Netw.},
  year = {2026},
  doi = {10.1109/tccn.2025.3595551},
}
Figure