User Terminal Power Control Based on Probabilistic Quantization in Federal Learning

In federal learning, the frequent interaction of model parameters between the central server and terminals leads to high communication overhead. Especially when the communication capabilities of terminals are weak and they are sensitive to energy consumption, optimizing communication overhead has become a key challenge. This paper comprehensively considers model parameter compression and communication algorithm optimization and proposes a power control method based on probabilistic quantization. Firstly, the MSE performance of parameter compression based on probabilistic quantization under non-perfect communication links is derived. Then, methods for minimizing the transmit power of terminals in the cases of equal and unequal transmit power are proposed respectively. Finally, the effectiveness of the proposed method is verified through numerical simulations. This method enables local terminals to reduce the local model parameters to be transmitted through quantization compression and dynamically adjust the transmit power according to network and channel conditions.