Federated Learning: Sum Power Constraints Optimization Design

The recent rise of high-risk scenarios such as unmanned aerial vehicles and intelligent driving, which require low latency and high privacy, has given rise to a new field called Federated Learning (FL). Limited communication bandwidth becomes a major bottleneck for aggregating local models in FL. Therefore, we use fast global model aggregation based on over-the-air computation (Aircomp). Wireless data transmission consumes most of the power of the edge nodes compared with the energy consumption required for sensing and processing data. Optimizing the transmission power of nodes can significantly save system power. To the best of our knowledge, in most of the research works on FL, the optimum design of system parameters usually takes the peak power of each sensor as the constraint condition. In this paper, the sum power constraint of edge devices is innovatively proposed to minimize the aggregation error of FL. We also study the optimization problem to minimize the total power of all edge sensors in the FL system under certain aggregation error constraints. Unfortunately, these problems presented are nonconvex biquadratic programming problems that are very difficult to solve. Therefore, we propose an alternative optimization method based on the difference-of-convex algorithm. The numerical results show that our proposal significantly reduces the aggregation error under the same total power. The simulation results also show that Intelligent Reflector Surface-assisted communication can greatly reduce the aggregation error and save system power.