A decentralized game theoretic approach for energy-aware resource management in federated learning

The resource management in Federated Learning (FL) system has been a challenging issue since mobile users can save the energy consumption by limiting their computing resources and dataset in training their local models in which users have the energy limitation. We analyze the performance of the global model on the size of dataset and computing resources used for the local training. The performance of the final model is significantly influenced by the resource management of users. Moreover, the decisions of the users on the communication, computing resources and size of dataset can affect the time taken for one computing round. Since a large number of mobile users participate in the FL, a centralized resource management is not practical. Thus, we formulate an energy-aware resource management problem for FL in which users are interested in minimizing the time taken for one computing round with the constraints of energy consumption, communication resources and performance of the training model. Due to the coupling in the communication resource allocation, we formulate the resource management problem as a Generalized Nash Equilibrium Problem (GNEP) and propose a decentralized algorithm. In addition, we analyze the performance of the proposed algorithm on the resource management, energy and time consumption.