Scalable Base Station Configuration via Bayesian Optimization with Block Coordinate Descent
Kakeru Takamori, Koya Sato
TL;DR
This work tackles the scalability challenge of optimizing dense base-station configurations where the objective is the area-averaged throughput $f(\Phi)=\frac{1}{|\mathcal{A}|}\sum_i\int_{\mathcal{A}} C_i(\mathbf{x})\,d\mathbf{x}$ with $C_i(\mathbf{x})=B\log_2(1+\mathrm{SINR}_i(\mathbf{x}))$. It introduces a scalable framework that decomposes the joint optimization into per-BS blocks via block coordinate descent and solves each block with Bayesian optimization using a Gaussian process with an RBF kernel, within an outer loop that randomizes the BS update order. Each per-BS subproblem optimizes $g_{r_i}(\phi_{r_i})=f(\Phi\mid \Phi^{\mathrm{best}}_{-r_i})$ while keeping other BS parameters fixed, with initialization and budget settings $N_{\mathrm{init}}$, $t_{\mathrm{sub}}$, $T_{\mathrm{sub}}$, and $T_{\mathrm{total}}$ guiding the iterations until convergence. Empirical results in a realistic 3D urban setting show substantial throughput gains over naive BO and fixed-antenna baselines, validating the approach’s scalability and effectiveness for dense deployments.
Abstract
This paper proposes a scalable Bayesian optimization (BO) framework for dense base-station (BS) configuration design. BO can find an optimal BS configuration by iterating parameter search, channel simulation, and probabilistic modeling of the objective function. However, its performance is severely affected by the curse of dimensionality, thereby reducing its scalability. To overcome this limitation, the proposed method sequentially optimizes per-BS parameters based on block coordinate descent while fixing the remaining BS configurations, thereby reducing the effective dimensionality of each optimization step. Numerical results demonstrate that the proposed approach significantly outperforms naive optimization in dense deployment scenarios.