Transferable Multi-Fidelity Bayesian Optimization for Radio Resource Management

Yunchuan Zhang, Sangwoo Park, Osvaldo Simeone

Abstract

Radio resource allocation often calls for the optimization of black-box objective functions whose evaluation is expensive in real-world deployments. Conventional optimization methods apply separately to each new system configuration, causing the number of evaluations to be impractical under constraints on computational resources or timeliness. Toward a remedy for this issue, this paper introduces a multi-fidelity continual optimization framework that hinges on a novel information-theoretic acquisition function. The new strategy probes candidate solutions so as to balance the need to retrieve information about the current optimization task with the goal of acquiring information transferable to future resource allocation tasks, while satisfying a query budget constraint. Experiments on uplink power control in a multi-cell multi-antenna system demonstrate that the proposed method substantially improves the optimization efficiency after processing a sufficiently large number of tasks.

Introduction

In modern 5G and 6G networks, operators must constantly tune complex radio settings, such as uplink power control, to maximize spectral efficiency. However, finding optimal configurations is a "black-box" challenge where objective functions lack simple analytical forms and evaluations via high-fidelity simulations or real-world tests are time-consuming and costly.

Standard Bayesian Optimization (BO) is typically employed for such tasks but suffers from two primary limitations in telecommunications: short-sightedness and cost-blindness. Conventional BO treats each network change (e.g., user mobility) as an independent problem, discarding previously learned insights. Furthermore, it often assumes uniform evaluation costs, failing to exploit the availability of cheap, low-fidelity simulations. This paper proposes the Multi-Fidelity Transferable Max-Value Entropy Search (MFT-MES) framework to address these inefficiencies by learning from past tasks and strategically utilizing varying fidelity levels.

The MFT-MES Framework

The MFT-MES framework introduces a novel acquisition function designed to optimize resource allocation over a sequence of tasks. The system operates by balancing immediate performance gains with the acquisition of "transferable knowledge" that benefits future optimizations.

MFT-MES Framework Overview

Key Components:

  • Multi-Fidelity Modeling: The optimizer can choose between $M$ levels of evaluation. It utilizes a Multi-Fidelity Gaussian Process (MFGP) where kernel parameters are shared across tasks, allowing the system to "scout" the landscape with cheap, noisy simulations before committing to expensive, high-fidelity tests.
  • Look-Ahead Acquisition Function: Unlike greedy methods, MFT-MES evaluates candidate points based on their potential to resolve the current task and their ability to improve the global model for future tasks. This is controlled by a weight parameter $\beta$, which scales the value of transferable knowledge.
  • Stein Variational Gradient Descent (SVGD): To manage the computational complexity of learning environment characteristics, the authors employ SVGD to maintain a set of "particles" representing the most likely system states.

Experimental Evaluation

The framework was evaluated using a simulated Uplink Open Loop Power Control (OLPC) scenario based on 3GPP TR 38.901 standards. The setup involved an Urban Microcellular (UMi) street canyon with 3 cells and 10 users per cell utilizing MIMO configurations. The experiment spanned a sequence of 40 tasks, each defined by a new random distribution of users.

Optimality Ratio vs. Number of Tasks

Key Results:

  • Efficiency Gains: MFT-MES treats early tasks as an "investment phase." While it may initially perform similarly to greedy methods, it begins to significantly outperform baselines after approximately 12 tasks as the accumulated knowledge pays off.
  • Superior Optimality: By the 40th task, MFT-MES achieved an optimality ratio of approximately 98%. This represents a 7% improvement over Continual GIBBON (which transfers knowledge but lacks active look-ahead) and a 23% improvement over standard, non-transferable GIBBON.
  • Horizon Sensitivity: The study found that a higher $\beta$ is beneficial for long task horizons, whereas a greedy approach ($\beta=0$) is preferable for very short sequences.

Conclusion

MFT-MES demonstrates that treating information gathering as a long-term investment significantly enhances the efficiency of radio resource management. While the framework introduces some computational overhead due to its complex acquisition function, the trade-off is highly favorable when objective function evaluations are expensive. Future work may explore adaptive $\beta$ parameters that adjust based on the remaining optimization horizon and methods to mitigate "negative transfer" when environment dynamics shift drastically.

Summary updated

Abstract

Radio resource allocation often calls for the optimization of black-box objective functions whose evaluation is expensive in real-world deployments. Conventional optimization methods apply separately to each new system configuration, causing the number of evaluations to be impractical under constraints on computational resources or timeliness. Toward a remedy for this issue, this paper introduces a multi-fidelity continual optimization framework that hinges on a novel information-theoretic acquisition function. The new strategy probes candidate solutions so as to balance the need to retrieve information about the current optimization task with the goal of acquiring information transferable to future resource allocation tasks, while satisfying a query budget constraint. Experiments on uplink power control in a multi-cell multi-antenna system demonstrate that the proposed method substantially improves the optimization efficiency after processing a sufficiently large number of tasks.
BibTeX
@inproceedings{zhang2024transferable,
  author = {Yunchuan Zhang and Sangwoo Park and Osvaldo Simeone},
  title = {{Transferable Multi-Fidelity Bayesian Optimization for Radio Resource Management}},
  booktitle = {SPAWC},
  year = {2024},
  doi = {10.1109/spawc60668.2024.10694025},
}
Figure