Supplementary information files for "Effective UAV-aided asynchronous decentralized federated learning with distributed, adaptive and energy-aware gradient sparsification"

Supplementary files for article "Effective UAV-aided asynchronous decentralized federated learning with distributed, adaptive and energy-aware gradient sparsification"<br><br>We consider decentralized federated learning (DFL) in unmanned aerial vehicle (UAV) networks where UAVs collaboratively train their machine learning (ML) models in a serverless peer-to-peer manner without sharing local data. We focus on three challenges affecting the performance and feasibility of UAV-aided DFL: i) communication inefficiency, ii) dynamics, heterogeneity and energy constraints of UAV networks, and iii) high synchronization overheads. To address these challenges, we propose an asynchronous DFL (A-DFL) model for UAV networks and design a novel distributed, adaptive and energy-aware model compression method based on the gradient sparsification. In this method, UAVs communicate asynchronously and apply the time-varying and non-identical compression parameters to adjust to a dynamic, heterogeneous environment. This reduces synchronization overheads and improves the communication efficiency given the strict battery constraints of UAVs. We show that our method can be formulated as a Markov potential game where the UAVs act as the players which decide on their compression parameters and the number of training data samples used for model updates. We prove that our game admits a dominant pure-strategy Nash equilibrium (NE) that maximizes its potential function and develop a new sparsified A-DFL algorithm enabling every UAV to reach its dominant strategy independently, in polynomial time. We then prove that the proposed algorithm converges to the Pareto-optimal NE representing the most efficient solution of our game. Using extensive simulations, we verify that our algorithm outperforms the state-of-the-art methods in terms of the key evaluation metrics of DFL.<br><br>© IEEE, All rights reserved.