Accuracy, latency and overhead test in Federated Deep Reinforcement Learning, L2FPPO in Focus

With the advent of 6G Open-RAN architecture, multiple operational services can be simultaneously executed in RAN, leveraging the near-Real-Time Radio Intelligent Controller (near-RT-RIC) and real-time (RT) nodes. The architecture provides an ideal platform for Federated Learning (FL): The xAPP is hosted in the near-RT-RIC to perform global aggregation, whereas the Open Radio Unit (ORU) allocates power to users to participate in FL in a RT manner. This paper identifies power and latency optimization as critical factors for enhancing FL in a stochastic environment. Due to the complexity of the problem involving multi-objective constraints, we formulate the problem as a mixed-integer nonlinear programming problem, then decompose the problem into modular units to find the perfect solutions. We employ dual Lagrange decomposition to derive optimal power allocation solutions, thereby improving user association with the ORU during FL. We introduce the Lyapunov Drift Plus Penalty framework to address latency issues during FL. Lastly, the Proximal Policy Optimization is utilized to solve for users\u2019 weights to upload to ORU. Empirical results demonstrate that our approach achieves a 6\\% improvement in accuracy, a 0.5\\% reduction in overhead, a 0.6\\% reduction in energy consumption, and a 0.4\\% reduction in latency compared to state-of-the-art methods.