Non-Cooperative Edge Server Selection Game for Federated Learning in IoT

Computational offloading is an efficient way to help constrained IoT devices by performing heavy tasks on Edge servers, especially tasks related to Machine Learning. Moreover, due to their limited learning capacity and memory size, such devices can only store a limited amount of data as a training set for their learning. Consequently, learning prediction is bound to be smeared with relatively high error. To mend that issue, IoT devices can federate the learning process with their pairs via an Edge server. However, offloading repeatedly the learning model through a wireless access network is time consuming. Hence, although learning collectively can reduce the learned model variance, it inflicts a communication cost depending on the selected Edge server. Therefore, in this paper, we model the Edge Selection problem as a non-cooperative game where devices autonomously and efficiently select an Edge server to reduce both their learning error and their communication cost. Depending on the characteristics of the dataset, we discern two different types of games. For each game type, we implemented and compared a semi-distributed algorithm based on Best Response dynamics. We compared the obtained results with the optimal centralized approach and with a less computationally intensive meta-heuristics, to assess the price of anarchy. Our numerical analysis shows that the Best Response algorithm strikes a good balance between efficiency and swift convergence.