Abstract
Two of the main principles underlying the life cycle of an artificial intelligence (AI) module in communication networks are adaptation and monitoring. Adaptation refers to the need to adjust the operation of an AI module depending on the current conditions; while monitoring requires measures of the reliability of an AI module’s decisions. Furthermore, the dynamic nature of a communication system imposes short coherent resources due to rapid conditions changes. This makes the available data for module training, e.g., known pilot transmission, to be of small size, requiring the AI module to be sample-efficient.In the first part of this thesis, we integrate both meta-learning and Bayesian learning for these challenges. Meta-learning addresses sample-efficiency by extracting useful shared properties across different channel conditions by learning how to learn when facing a new condition with few pilots. Bayesian learning increases reliability by producing better-calibrated, e.g., less overconfident, decisions. A well-calibrated AI model is one that can reliably quantify the uncertainty of its decisions, assigning high confidence levels to decisions that are likely to be correct and low confidence levels to decisions that are likely to be erroneous. As an application, we apply the integration of meta-learning and Bayesian learning to symbol demodulation and validate its improvements. The capacity to quantify uncertainty in the model parameter space is further leveraged by extending Bayesian meta-learning to an active setting. In it, the designer can select in a sequential fashion channel conditions under which to generate data for meta-learning from a channel simulator. Bayesian active meta-learning is seen in experiments to significantly reduce the number of frames required to obtain efficient adaptation procedure for new frames of an equalization problem.
While Bayesian meta-learning is better than frequentist learning, it is a best-effort approach with no formal guarantees. To obtain mathematical reliability guarantees, we in-corporate in the second part of the thesis the framework of conformal prediction. Conformal prediction post-processes in an ad-hoc manner a probabilistic predictor into set predictor, producing set of labels that is guaranteed to contain the correct label with a probability chosen by the designer. This is done irrespectively of the true, unknown, distribution underlying the generation of the variables of interest, and can be defined in terms of ensemble or time-averaged probabilities. We apply recent conformal prediction advances, including cross-validation-based schemes that reduce average set sizes, to communication problems such as symbol demodulation and modulation classification. As a communication system may have its data distribution change over time, online conformal prediction is used and investigated for received signal strength prediction, as well for 5G dynamic scheduling of ultra-reliable and low-latency communication traffic. Experiments compare empirical coverage rate and averaged set sizes of different schemes.
Conformal prediction is a special case of conformal risk control for which the risk is the miscoverage indicator. We develop a novel cross-validation-based conformal risk control and show its ability to meet a target risk while keeping the average predictive set sizes smaller than the validation-based counterparts.
| Date of Award | 1 Jun 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Osvaldo Simeone (Supervisor) |