Small-Scale-Fading-Aware Resource Allocation in Wireless Federated Learning

TL;DR

The paper tackles wireless federated learning under rapid small-scale fading by introducing a small-scale-fading-aware MARL framework. It models gradient uploads as a Dec-POMDP and solves it with QMIX under centralized training and decentralized execution, enabling per-slot spectrum and power decisions that directly impact FL convergence. A one-step convergence bound links upload success to FL performance, guiding reward design that emphasizes both gradient participation and interference management. Experimental results show the QMIX-based strategy outperforms baselines across statistical and system heterogeneity, with ablations confirming the importance of small-scale fading dynamics in optimizing FL outcomes.

Abstract

Judicious resource allocation can effectively enhance federated learning (FL) training performance in wireless networks by addressing both system and statistical heterogeneity. However, existing strategies typically rely on block fading assumptions, which overlooks rapid channel fluctuations within each round of FL gradient uploading, leading to a degradation in FL training performance. Therefore, this paper proposes a small-scale-fading-aware resource allocation strategy using a multi-agent reinforcement learning (MARL) framework. Specifically, we establish a one-step convergence bound of the FL algorithm and formulate the resource allocation problem as a decentralized partially observable Markov decision process (Dec-POMDP), which is subsequently solved using the QMIX algorithm. In our framework, each client serves as an agent that dynamically determines spectrum and power allocations within each coherence time slot, based on local observations and a reward derived from the convergence analysis. The MARL setting reduces the dimensionality of the action space and facilitates decentralized decision-making, enhancing the scalability and practicality of the solution. Experimental results demonstrate that our QMIX-based resource allocation strategy significantly outperforms baseline methods across various degrees of statistical heterogeneity. Additionally, ablation studies validate the critical importance of incorporating small-scale fading dynamics, highlighting its role in optimizing FL performance.

Figures (6)

• Figure 1: Workflow of the proposed resource allocation strategy: In the centralized training phase, each network is deployed at the BS, interacts with the FL environment, and learns from stored experiences. After training, the BS sends the trained agent network parameters to the corresponding clients. In the decentralized execution phase, each client independently decides its sub-band allocation and power level based on local observations.
• Figure 2: QMIX network: The blue block on the right represents the agent network, while the red block on the left depicts the mixing network. Within this network, the yellow block illustrates the hypernetwork, which generates the weights and biases for its layers.
• Figure 3: The distribution of local training data across different degrees of statistical heterogeneity. Each column represents the dataset of an individual client, and each color indicats a distinct category of training data.
• Figure 4: Test accuracy over the number of training rounds under different levels of system and statistical heterogeneity. The concentration parameter $\alpha$ controls the statistical heterogeneity among local datasets, and the cluster number $n_c$ controls the system heterogeneity among devices.
• Figure 5: Cumulative global reward for each training episode with increasing training episodes. The concentration parameter and cluster number are $\alpha=0.5$ and $n_c=21$.

Theorems & Definitions (3)

• Lemma 1: Bounded Local Drift
• Lemma 2: Bounded Average Squared Norm
• Theorem 1: One-Step Convergence of FedAvg Algorithm