Mobile networks require safe optimization to adapt to changing conditions in traffic demand and signal transmission quality, in addition to improving service performance metrics. With the increasing complexity of emerging mobile networks, traditional parameter tuning methods become too conservative or complex to evaluate. For the first time, we apply safe Bayesian optimization to mobile networks. Moreover, we develop a new safe collaborative optimization algorithm called CoSBO, leveraging information from multiple optimization tasks in the network and considering multiple safety constraints. The resulting algorithm is capable of safely tuning the network parameter online with very few iterations. We demonstrate that the proposed method improves sample efficiency in the early stages of the optimization process by comparing it against the SafeOpt-MC algorithm in a mobile network scenario.
Modern mobile networks consist of complex systems where key parameters, such as antenna tilt or beamwidth, directly influence the coverage quality experienced by users. Optimizing these parameters is essential for maintaining efficient network performance and adapting to changing environmental conditions. However, evaluating the impact of new configurations is complex. It requires simulations or testing with real-world systems which might take days and can be challenging due to safety constraints such as maintaining minimum coverage levels or minimum signal quality.
This paper addresses the problem of safely and efficiently optimizing performance-related parameters in mobile network systems, since the evaluation of each new parameter might take up to a day to complete. The goal is to develop a method that maximizes performance subject to safety constraints through online optimization, that is, by safely exploring parameter values directly in the network. In addition, our objective is to use data from collaborative agents to improve the sample efficiency of current methods.
With the increasing complexity of modern mobile networks, such as 5G and future 6G networks, traditional methods like rule-based strategies have become too conservative. Promising approaches like reinforcement learning (RL) require extensive offline training data or high-quality simulators, and safe evaluation in the network is challenging. Bayesian Optimization (BO), and in particular safe Bayesian Optimization (safe BO), addresses some of these limitations. Unlike RLbased methods that require training, safe BO is an online This work has been performed with support of Sweden’s Innovation Agency. optimization approach. It uses probabilistic models, typically Gaussian processes (GPs), to explore the parameter space and iteratively update the knowledge about performance while ensuring that each new evaluation is safe. Safe BO uses fewer evaluations than RL methods. However, it requires an initial safe set which might be small or poorly located, hence increasing the number of evaluations needed to reach an optimal configuration, or causing the algorithm to be stuck local optima. We propose a new algorithm called Collaborative Safe Bayesian Optimization (CoSBO) that builds on the standard safe BO algorithm for multiple constraints SAFEOPT-MC [1], introducing a collaborative initialization strategy. The proposed method differs from previous methods in several ways. Unlike rule-based systems, it adapts dynamically to the data. In contrast to RL-based methods, it optimizes online with no pre-training required. Although BO already improves efficiency by balancing exploration and exploitation, we further improve sample efficiency by integrating external data into the optimization process and show that it helps discovering new optimal regions.
The first contribution of this paper is CoSBO, the new algorithm that takes advantage of the similarity between agents, such as antennas in similar geographical or network traffic conditions, to enrich the initial estimation model of the network performance. By identifying the most correlated collaborator among a set of network optimization problems and transferring a subset of its data, the optimization process starts with a more informed model, without requiring additional evaluations at the initial stage. The quality and availability of the collaborator data are key factors to consider, since the benefit may be reduced if the selected collaborator is poorly correlated. The second contribution consists of applying a safe Bayesian optimization method to the telecommunications domain for the first time. This is done through a careful modeling of safety and performance functions for mobile networks and is evaluated through series of simulation-based experiments. The method is shown to reach high-performing configurations with very few parameter evaluations in the network while maintaining formal safety guarantees. In general, this work contributes to the development of adaptive, data-efficient, and safety-constrained optimization techniques for modern mobile networks.
The outline of this paper continues with related work and background, followed by a description of the problem and the CoSBO algorithm, and finishes with details on the experiments performed, results obtained, and concluding remarks. The code for the algorithm is available at: https://github. com/EricssonResearch/collaborative-safe-bo, including hyper-parameters and simulated datasets used in our experiments.
Recent studies provide RL-based approaches that optimize parameters in mobile networks in a multi-agent setting, such as antenna tilt optimization [2]. These methods require an offline training phase, using a large number of simulated data, since it is too risky and expensive to retrieve samples from real systems, and lack safety guarantees. In contrast, our work focuses on online optimization, aims to reduce the number of evaluations t
This content is AI-processed based on open access ArXiv data.