Joint User Association and Resource Allocation for Hierarchical Federated Learning Based on Games in Satisfaction Form

Abstract

Hierarchical Federated Learning (HFL) has emerged to overcome the shortcomings of conventional Federated Learning (FL) due to communication obstacles between the end users and the cloud server and the congestion at the backhaul of wireless network implementations. In this paper, we consider a wireless user-edge-cloud HFL network where the transmissions of the users' local model parameters to the edge are multiplexed via the Non-Orthogonal Multiple Access (NOMA) technique. The joint problem of association and uplink transmission power allocation of the users to the edge is formulated and solved as a non-cooperative game in satisfaction form. Diverging from the prevailing research that proposes centralized solution concepts, each user makes autonomous decisions regarding its association and power level so as to attain a minimum acceptable tradeoff of three vital network factors. The latter includes the global model's training accuracy and the users' consumed energy and time during transmission. Different types of equilibria are explored, i.e., the Satisfaction Equilibrium (SE) and Minimum Efficient Satisfaction Equilibrium (MESE) which not only fulfills users' minimum tradeoff but also minimizes the overall network's cost. Algorithms based on Reinforcement Learning (RL) and Best Response Dynamics (BRD) are, then, devised to conclude the SE and MESE points. The proposed framework is evaluated via modeling and simulation, verifying its efficiency in achieving an equitable balance in the network.