FedRIC: Heterogeneity-Aware Federated Learning with Hard Deadline Constraints in Open RAN

Abstract

Open Radio Access Networks (O-RANs) bring intelligence to disaggregated RANs through virtualized platforms and open interfaces, but the stringent latency budgets of the RAN Intelligent Controller (RIC) and heterogeneous traffic conditions make real-time deployment costly. Centralized learning is ill-suited to the RIC due to its computational burden, timing constraints, and communication overhead, motivating Federated Learning (FL) as a distributed alternative. We formulate FL in O-RAN as a Mixed-Integer Nonlinear Program (MINLP) that minimizes communication cost and learning time under heterogeneity in User Equipment (UE) compute, traffic, and channel conditions, and decompose it into a UE Assignment (UA) subproblem and a Computation and Bandwidth Allocation (CBA) subproblem. We then propose FedRIC, a heterogeneity-aware FL framework combining compressed gradient descent, slice-aware resource allocation, and multilevel hierarchical aggregation. Experiments show that FedRIC outperforms FedCOM, FedAvg, and QFedAvg in convergence rate, communication cost, and latency.