MCORANFed: Communication Efficient Federated Learning in Open RAN
Amardip Kumar Singh, Kim Khoa Nguyen · 2022
To bring network intelligence closer to the end devices, Open Radio Access Networks (O-RAN) specifies a disaggregated and vendor agnostic framework of hierarchical processing units. Although this framework can be useful for certain use cases of 5G smart services, no standardised method to train Machine Learning (ML) models has been defined. Recently, Federated Learning (FL) has emerged as a promising solution for training in disaggregated systems. Unfortunately, the stringent deadline of O-RAN control loops and fluctuating network band-width poses challenges for FL implementation. In this paper, we tackle this problem by proposing an accelerated gradient descent method to expedite the FL convergence, and a compression operator to reduce the communication cost. We formulate a joint optimization problem to select the participating local trainers in each global round of FL and allocate the resources to these trainers while minimizing the overall learning time and resource costs. We design an FL algorithm (MCORANFed) which adheres to the deadline of O-RAN control loops. Extensive experimental results show that MCORANFed outperforms state-of-the-art FL methods such as MFL, FedAvg, and FedProx in terms of its convergence and objective costs.