Collaborative Safe Bayesian Optimization

TL;DR

A new safe collaborative optimization algorithm called CoSBO is developed, capable of safely tuning the network parameter online with very few iterations, and is capable of safely tuning the network parameter online with very few iterations.

Abstract

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.

Figures (6)

• Figure 1: Illustration of two mobile network scenarios where the antenna parameter being adjusted on the left is the horizontal beamwidth, and on the right is the electrical tilt. Each base station includes three cells. Both scenarios show two negative outcomes (non connected users) resulting from beamwidth/tilt configuration and insufficient coordination.
• Figure 2: Optimization process of the horizontal beamwidth of an antenna using the CoSBO algorithm to maximize the performance function (gray curve on the top) while satisfying the safety constraint (gray curve on the bottom). The visualization shows the progress of the estimated functions (light blue curves) at the first, fifth, and tenth iterations. The algorithm starts with collaborator data (orange crosses) that is added as observed data (black crosses). Then, based on the GP posterior and its confidence intervals (light blue areas), it iteratively selects and evaluates new data (red crosses) that are above the safety threshold (gray dashed line) and within the current safe set (green set). The input parameter region is truncated on the plot.
• Figure 3: Two of the five simulator-generated topological maps for traffic volumes 15e7s. The mobile network is comprised by base stations (macro, red stars), the users (green dots), the obstacles or indoor areas (white squares), and the outdoor areas (black squares). The indoor-outdoor probability corresponds to a 0.5 quantity.
• Figure 4: Optimization method comparison for tilt and beamwidth using high-quality collaborators (on the left) and low-quality ones (on the right). Curves correspond to the optimum obtained at each iteration of the process for a budget of 60 iterations. The shaded areas represent the 25-75 inter-quartile range of values obtained over the 15 simulated networks times the 10 different starting points. CoSBO method is displayed in blue, SafeOpt-MC in orange, and an unsafe random exploration method in violet red.
• Figure 5: Optimization method comparison for tilt and beamwidth using high-quality collaborators and a GP model with lengthscale value 4. Curves correspond to the optimum obtained at each iteration of the process for a budget of 60 iterations. The shaded areas represent the 25-75 inter-quartile range of values obtained over the 15 simulated networks times the 10 different starting points. CoSBO method is displayed in blue, SafeOpt-MC in orange, and an unsafe random exploration method in violet red.