Improving the Effectiveness of Bitwidth-Aware Federated Learning in Wireless Networks

This research explores the potential of model quantization in enhancing the efficiency of federated learning (FL) in the domains of wireless communication and computing. The process involves collecting the quantized local FL model parameters obtained from edge devices and combining them to create a global-quantized server. Afterwards, the devices are synchronised using the suggested bitwidth FL method. We need to collaboratively decide on the devices that will be included in a FL training iteration, as well as the specific quantization bitwidth set to be utilised for compressing the local model. In this optimisation issue, we are provided with a budget for device samples and a limit on the end-to-end latency each iteration for quantized federated learning (FL). The objective is to minimise the training loss. To address the said problem, it is essential to possess a comprehensive comprehension of how quantization impacts the general efficiency of machine learning. Additionally, the capability to conduct server-side inference that can accurately predict the functioning of this procedure is crucial. In order to achieve this objective, we carry out an extensive examination of the effectiveness of the suggested Federated Learning (FL) method under the conditions of limited communication and quantization errors occurring in wireless connections. This work validates our idea by presenting quantitative findings that demonstrate the degree to which the training loss in FL is enhanced in each cycle, depending on the choice of device, quantization technique, and many characteristics particular to the model. Next, we presented a proposal base of reinforcement learning (RL that is utilised to choose actions in a sequential manner. We demonstrate that the FL process of training may be seen as a specific case of a Markov decision process. This is the manner in which we confront the second obstacle. This is a training approach for FL that differs from unconstrained by a specific model RL. Unlike framework RL, this model-based learning method seeks to accurately imitate the behaviour of agents utilising Belief MDP, which is a mathematical characterization.