Tree Meets Transformer: A Hybrid Architecture for Scalable Power Allocation in Cell-Free Networks

TL;DR

This paper tackles scalable per-user power allocation in cell-free mMIMO by introducing a hybrid Tree-Transformer that compresses user features into a binary-tree root, applies a Transformer at the root, and decodes per-user UL/DL powers with a shared decoder. The architecture achieves logarithmic depth and linear total complexity, with inference costs independent of the number of users $K$ and APs $L$, enabling scalable deployment. Empirically, it delivers near-optimal max-min fairness performance with significantly reduced latency compared to full-attention baselines, and generalizes across unseen network sizes due to its global root representation and modular design. The work highlights practical relevance for large-scale deployments and opens avenues for distributed learning among edge APs to further improve scalability and adaptability.

Abstract

Power allocation remains a fundamental challenge in wireless communication networks, particularly under dynamic user loads and large-scale deployments. While Transformerbased models have demonstrated strong performance, their computational cost scales poorly with the number of users. In this work, we propose a novel hybrid Tree-Transformer architecture that achieves scalable per-user power allocation. Our model compresses user features via a binary tree into a global root representation, applies a Transformer encoder solely to this root, and decodes per-user uplink and downlink powers through a shared decoder. This design achieves logarithmic depth and linear total complexity, enabling efficient inference across large and variable user sets without retraining or architectural changes. We evaluate our model on the max-min fairness problem in cellfree massive MIMO systems and demonstrate that it achieves near-optimal performance while significantly reducing inference time compared to full-attention baselines.

Figures (5)

• Figure 1: Illustration of a cell-free mMIMO network. Users are served by AP without cell boundaries.
• Figure 2: Block diagram of the proposed model architecture.
• Figure 4: Evolution of training and validation loss. The model converges rapidly and generalizes well across network configurations.
• Figure 5: CDF of optimal and predicted power on the test set. The predicted curves closely follow the optimal ones, confirming the model’s robustness and generalization capability.
• Figure 6: Comparison of SE trends: (\ref{['fig:se_vs_K']}) varying UE count $K$, (\ref{['fig:se_vs_L']}) varying AP count $L$. Predicted SE closely tracks the optimal SE across both scenarios.